CN115639605A - Automatic high-resolution fault identification method and device based on deep learning - Google Patents

Abstract

The present invention provides a method and device for automatic identification of high-resolution faults based on deep learning, including: acquiring a seismic image to be identified; inputting the seismic image to be identified into a quality improvement model, and outputting an enhanced seismic image; wherein, the The quality improvement model is obtained after training a first training set and a first validation set constructed based on a plurality of sample low-quality seismic images and corresponding high-quality seismic image labels, and the first training set and the first validation set Generated by random parameter simulation; input the lifted seismic image into a fault identification model, and output a fault probability map; wherein, the fault identification model is a second training set constructed based on a plurality of sample seismic images and corresponding fault probability map labels obtained after training with the second validation set. Compared with the conventional method, the proposed method can obtain a cleaner and clearer fault probability map, and the predicted fault position is also more accurate.

Description

Method and device for automatic identification of high-resolution faults based on deep learning

technical field

The invention relates to the technical field of geophysical exploration, in particular to an automatic recognition method and device for high-resolution faults based on deep learning.

Background technique

Seismic image is an important geophysical data, which can provide continuous and large-scale subsurface geological information. Experienced geophysicists can infer the sedimentary age of the strata, analyze the sedimentary environment and tectonic movement, and assist oil and gas exploration through seismic images. The subsurface fault is an important geological structure, which controls the deposition of the basin, affects the migration of subsurface fluids and the distribution of minerals. Identifying subsurface faults from seismic images is a commonly used means, however, manual interpretation of faults is a laborious and highly subjective task due to the large scale and data complexity of seismic surveys. Therefore, with the development of computer technology, computer-aided automatic fault recognition is becoming more and more popular.

Patent application CN202010188216.0 discloses a three-attribute fusion fault identification method based on the dominant frequency of seismic data. Firstly, the seismic data of the whole frequency band is frequency-divided and processed, and the dominant frequency volume data is selected to make the coherence, dip and azimuth attribute volumes. Finally, the three Species attributes are displayed using HIS superimposition to obtain a fusion body for fault interpretation. This method improves the identification ability of faults through frequency division and image fusion technology, but it takes a lot of energy to select the dominant frequency volume, and the subsequent attribute volume calculation requires a lot of calculations, and the selection of parameters affects the subsequent interpretation accuracy.

Patent application CN201910854606.4 discloses a coherent enhanced fault identification method based on seismic analysis traces under horizon constraints, which improves the traditional coherent algorithm by introducing Hilbert transform and histogram equalization, which has high noise resistance However, due to the limitations of the coherent body itself, the result is prone to multiple solutions and cannot clearly describe small faults.

Patent application CN201710186136.X discloses an automatic identification method for seismic coherent volume image faults based on AdaBoost algorithm, which realizes automatic identification of seismic coherent volume seismic images. Although the accuracy of fault identification is improved, it needs to be adjusted when mining feature information. The size of the data block increases the space occupancy of the data.

Patent application CN202110946552.1 discloses a seismic fault identification method based on deep learning semantic segmentation, which uses the powerful fitting ability of deep convolutional neural network to realize fast and accurate fault identification, but its identification accuracy is affected by the signal-to-noise ratio of seismic data .

In 2016, Huang Cheng et al. used analysis techniques such as volume attributes, bedding-along structural attributes, and absorption-attenuation attributes to optimize seismic attribute bodies, and effectively predicted the fault system in the study area in stages and layers. The combined application of multiple methods reduced the Multisolution in fault identification process. In 2017, Zhong Weijun and others comprehensively used coherent volume, variance volume, structural curvature volume, multi-window dip scanning volume, structure-guided filtering, coherent energy gradient volume, ant body tracking, edge detection and other technical means to form a dip-guided technology. The technical process of fault identification represented by the company has identified faults developed at different levels and in different stages. In 2017, Zhang Rui and others proposed the frequency-division ant tracking technology with the help of generalized S-transform and wavelet transform analysis techniques. This method can better suppress noise interference, identify faults more clearly and accurately, and can identify conventional full-band data. Difficult to identify deep small faults. In 2017, Li Quan and Tong Liqing identified complex faults through the optimal combination of seismic attributes. First, they controlled the distribution trend of faults with coherent volume attributes based on guidance, and then optimized the fusion of dip volumes, azimuth volumes, and curvature volumes that highlight local details. The attribute body finely depicts the spatial distribution characteristics of third- and fourth-order fractures and micro-fractures. Also in 2017, Sun Zhenyu and others used seismic attributes as the input of support vector machine to build a fault identification model with an accuracy rate of 98%, which reduced the influence of human subjective factors and shortened the interpretation cycle. In 2020, Liu Wanjin and Xian Hailong fused images of coherent body and amplitude, coherent body and maximum positive curvature, made full use of their respective structural information, enhanced the ability to identify small faults, and improved fault plane combination, plane distribution characteristics and mutual relationship reliability. In 2022, Hou Juntao and others applied the automatic fault extraction technology to the frequency-division coherent body based on guided filtering, and obtained more precise and accurate fault description results by suppressing noise. In 2022, Ding Changwei and others proposed to use the information value to reduce the seismic attributes, combined with the improved Bayesian optimization algorithm, optimize the XGBoost parameters to further improve the accuracy of small fault seismic interpretation, this method effectively solves the uneven distribution of small fault samples It has a certain anti-interference ability. In 2022, Liu Naihao and others introduced the edge detection technology in deep learning and optimized the network structure according to the characteristics of seismic data and faults, and proposed an improved HED network using seismic fault interpretation, which has higher accuracy and continuity in fault prediction. better. Also in 2022, Zhang Zheng et al. combined the deep residual network with transfer learning and applied it to fault recognition. On the basis of the synthetic data training model, a small number of actual fault samples were used for transfer learning to enhance the generalization ability of the network and improve the Performance of fault identification.

The above methods have contributed to the automatic identification of seismic faults to a certain extent, but these methods lack the accuracy analysis of fault identification, and also lack the analysis of the identification effect when the quality of multi-seismic data is low, for example, seismic data contains Lots of random noise, low resolution seismic data. However, the seismic data collected in the field will inevitably be disturbed by random noise and low resolution caused by the absorption of high-frequency components of seismic waves by the underground medium, especially in deep and ultra-deep exploration. These reasons lead to the poor universality of the current fault identification method.

Therefore, how to avoid the low accuracy and low reliability of automatic fault identification when the quality of seismic data is low is still a problem to be solved urgently by those skilled in the art.

Contents of the invention

The present invention provides an automatic identification method of high-resolution faults based on deep learning, which is used to solve the problems of low accuracy and low reliability of automatic identification of faults in the prior art when the quality of seismic data is low.

The present invention provides an automatic recognition method for high-resolution faults based on deep learning, including:

Obtain the seismic image to be identified;

Input the seismic image to be identified into the quality improvement model, and output the enhanced seismic image;

Wherein, the quality improvement model is obtained after training on a first training set and a first verification set constructed based on a plurality of sample low-quality seismic images and corresponding high-quality seismic image labels, and the first training set and the The first verification set is generated by random parameter simulation;

Inputting the upgraded seismic image into a fault identification model, and outputting a fault probability map;

Wherein, the fault recognition model is obtained after training on a second training set and a second verification set constructed based on a plurality of sample seismic images and corresponding fault probability map labels.

According to the automatic identification method of a high-resolution fault based on deep learning provided by the present invention, it also includes:

input the seismic image to be identified into the fault identification model, and output the direct identification result of the seismic image;

The fault probability map is compared with the seismic image direct identification result.

According to a method for automatic recognition of high-resolution faults based on deep learning provided by the present invention, the first training set and the first verification set are generated through random parameter simulation, specifically including:

Generate formation reflectance models according to preset rules;

Generate different types of formation reflectance models through random parameter control;

Convolve any stratum reflectivity model with low-frequency Reich wavelets and add random noise to the seismic data for twice down-sampling to obtain sample low-quality seismic images;

The corresponding high-quality seismic image labels are obtained by convolving any stratum reflectivity model with the high-frequency Rake wavelet.

According to an automatic identification method of high-resolution faults based on deep learning provided by the present invention, the network structure of the quality improvement model in the training process is a generative confrontation network, and the generative confrontation network includes a generator and a discriminator , where the generator contains a skip connection and multiple residual connections, and the discriminator contains a convolutional layer and seven convolutional blocks.

According to an automatic identification method of high-resolution faults based on deep learning provided by the present invention, during the training process of the quality improvement model, the network parameters corresponding to the highest peak signal-to-noise ratio of the quality improvement model on the verification set are: Optimal model parameters, wherein the peak signal-to-noise ratio is the similarity between the high-quality seismic image reconstructed by the generative adversarial network and the real high-quality seismic image.

According to a method for automatic recognition of high-resolution faults based on deep learning provided by the present invention, the model network structure used in the training process of the fault recognition model is a U-net network, and the U-net network includes encoding branches and A decoding branch, the encoding branch consists of four convolution operations and downsampling, the decoding branch consists of four upsampling and convolution operations.

According to an automatic recognition method of high-resolution faults based on deep learning provided by the present invention, in the loss function of the fault recognition model training process, the weight assigned to pixels representing faults is greater than the weight assigned to pixels representing non-faults.

The present invention also provides an automatic recognition device for high-resolution faults based on deep learning, including:

an acquisition unit, configured to acquire a seismic image to be identified;

a quality improvement unit, configured to input the seismic image to be identified into a quality improvement model, and output the enhanced seismic image;

Wherein, the quality improvement model is obtained after training on a first training set and a first verification set constructed based on a plurality of sample low-quality seismic images and corresponding high-quality seismic image labels, and the first training set and the The first verification set is generated by random parameter simulation;

a fault identification unit, configured to input the lifted seismic image into a fault identification model, and output a fault probability map;

Wherein, the fault recognition model is obtained after training on a second training set and a second verification set constructed based on a plurality of sample seismic images and corresponding fault probability map labels.

The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and operable on the processor. When the processor executes the program, the depth-based Steps in learning the method for automatic identification of high-resolution tomograms.

The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the automatic identification of high-resolution faults based on deep learning as described in any one of the above is realized. method steps.

The method and device for automatic identification of high-resolution faults based on deep learning provided by the present invention obtain seismic images to be identified; input the seismic images to be identified into the quality improvement model, and output the enhanced seismic images; wherein, the quality improvement The model is obtained after training a first training set and a first validation set constructed based on a plurality of sample low-quality seismic images and corresponding high-quality seismic image labels, and the first training set and the first validation set are obtained by random Parameter simulation generation; input the upgraded seismic image into a fault identification model, and output a fault probability map; wherein, the fault identification model is based on a second training set and a second training set constructed based on a plurality of sample seismic images and corresponding fault probability map labels The second validation set is obtained after training. A cleaner and clearer fault probability map is achieved, and the predicted fault position is more accurate.

Description of drawings

In order to more clearly illustrate the present invention or the technical solutions in the prior art, the accompanying drawings that need to be used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings in the following description are the present invention. For some embodiments of the invention, those skilled in the art can also obtain other drawings based on these drawings without creative effort.

Fig. 1 is a schematic flow chart of the automatic identification method of high-resolution faults based on deep learning provided by the present invention;

Fig. 2 is a schematic structural diagram of an automatic identification device for high-resolution faults based on deep learning provided by the present invention;

Fig. 3 is the neural network structural diagram that the seismic data quality promotion that the present invention provides;

Fig. 4 is a schematic diagram of the training process of the seismic data quality improvement neural network provided by the present invention;

Fig. 5 is the neural network structural diagram of earthquake fault identification provided by the present invention;

Fig. 6 is a schematic diagram of the training process of the fault recognition neural network provided by the present invention;

Fig. 7 is a schematic diagram of the end-to-end workflow of high-resolution fault intelligent identification provided by the present invention;

Figure 8 is a schematic diagram of the application of the present invention on real seismic data in the field and the comparison with the traditional implementation process;

FIG. 9 is a schematic structural diagram of an electronic device provided by the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention , but not all examples. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

Because the problems of low precision and low reliability of automatic fault identification generally exist in the prior art when the quality of seismic data is low. The method for automatic recognition of high-resolution tomograms based on deep learning of the present invention will be described below with reference to FIGS. 1-9 . Fig. 1 is a schematic flow chart of the automatic identification method of high-resolution faults based on deep learning provided by the present invention. As shown in Fig. 1, the method includes:

Step 110, acquire the seismic image to be identified.

Specifically, the seismic image to be identified is obtained. Usually, the image quality of the seismic data image obtained is not high due to the limitation of the field acquisition environment. Here, the image quality refers to the resolution and image signal-to-noise ratio, that is, the seismic data image collected in the field The resolution is low and the signal-to-noise ratio of the image is also low, and the noise is relatively large.

Step 120, input the seismic image to be identified into the quality improvement model, and output the enhanced seismic image;

Wherein, the quality improvement model is obtained after training on a first training set and a first verification set constructed based on a plurality of sample low-quality seismic images and corresponding high-quality seismic image labels, and the first training set and the The first validation set is generated by simulation with random parameters.

Specifically, the low-quality seismic image to be identified is first input into the quality improvement model for quality improvement, and the output seismic image after upgrading has higher resolution and larger signal-to-noise ratio than the image before quality improvement, while the quality improvement The model is trained based on a large number of sample low-quality seismic images and corresponding high-quality seismic image labels. During the training process, the training data is divided into a training set and a verification set. After the training set is completed, the verification set is used for verification to select The model with the optimal network parameters is obtained. Here, low-quality seismic images and high-quality seismic images need to be explained. The low-quality and high-quality seismic images here mean that the image resolution is lower than the first threshold and the image resolution is higher than the second threshold. The images with the threshold, and the images with the image signal-to-noise ratio lower than the third threshold and higher than the fourth threshold, and the first threshold, the second threshold, the third threshold and the fourth threshold are set based on the application scenario. Moreover, both the first training set and the first verification set in the embodiment of the present invention are generated through random parameter simulation, which does not require manual labeling and saves labor costs.

Step 130, input the upgraded seismic image into the fault identification model, and output the fault probability map;

Wherein, the fault recognition model is obtained after training on a second training set and a second verification set constructed based on a plurality of sample seismic images and corresponding fault probability map labels.

Specifically, after the quality improvement of the seismic image to be recognized is completed, it is input to the fault recognition model, and the model outputs the corresponding fault probability map, and the fault recognition model is trained based on a large number of sample seismic images and the corresponding fault probability map labels. During the training process, the training data is also divided into a training set and a verification set like the quality improvement model. After the training set is trained, the verification set is used for verification to select the model with the optimal network parameters.

In the method provided by the present invention, by acquiring the seismic image to be identified; inputting the seismic image to be identified into a quality improvement model, and outputting the enhanced seismic image; wherein, the quality improvement model is based on a plurality of sample low-quality seismic images and corresponding The first training set and the first verification set constructed by high-quality seismic image labels are obtained after training, and the first training set and the first verification set are generated by random parameter simulation; the improved seismic images are input into the fault The identification model outputs a fault probability map; wherein, the fault identification model is obtained after training on a second training set and a second verification set constructed based on a plurality of sample seismic images and corresponding fault probability map labels. A cleaner and clearer fault probability map is achieved, and the predicted fault position is more accurate.

Based on the foregoing embodiments, the method also includes:

input the seismic image to be identified into the fault identification model, and output the direct identification result of the seismic image;

The fault probability map is compared with the seismic image direct identification result.

Specifically, after completing the input and output of the seismic image to be identified through the two layers of the quality improvement model and the fault identification model, the obtained fault probability map and the direct identification result of the seismic image output by directly inputting the seismic image to be identified into the fault identification model are carried out. Compared with the experimental results, it can be found that the fault probability map obtained by the quality improvement model is clearer at the fault display, and the fault display effect is better.

Based on the above embodiments, in this method, the first training set and the first verification set are generated through random parameter simulation, specifically including:

Generate formation reflectance models according to preset rules;

Generate different types of formation reflectance models through random parameter control;

Convolve any stratum reflectivity model with low-frequency Reich wavelets and add random noise to the seismic data for twice down-sampling to obtain sample low-quality seismic images;

The corresponding high-quality seismic image labels are obtained by convolving any stratum reflectivity model with the high-frequency Rake wavelet.

Specifically, training data is generated by random parameter simulation. First, a reflectance model with horizontal layers is randomly generated. Then, random vertical perturbations are added to the horizontal layer model to simulate formation dip. Next, Gaussian perturbations were randomly added to the model to simulate the fold structure. Next, the real fault distribution is simulated by adding volumetric vector field perturbations to the model. So far, different types of formation reflectance models can be generated through stochastic parameter control. Finally, the reflectivity model is convolved with the low-frequency Reich wavelet and the seismic data with random noise added is twice down-sampled to obtain a low-quality sample seismic image, and the reflectivity model is convolved with the high-frequency Reich wavelet to obtain a corresponding High-quality seismic image tags.

Based on the above embodiments, in this method, the network structure of the quality improvement model during training is a generative adversarial network, and the generative adversarial network includes a generator and a discriminator, wherein the generator includes a skip layer connection and multiple residual connections, the discriminator consists of one convolutional layer and seven convolutional blocks.

Specifically, the construction of the neural network is based on the current popular generative adversarial network to build a network for improving the quality of seismic data. The network structure consists of a generator and a discriminator. The network structure of the generator consists of a skip connection and multiple residual connections. First, shallow features of low-quality seismic images are extracted through a convolutional layer and a parameterized ReLU (PReLU) activation function. Then, five convolutional blocks with residual connections are used to extract deep features. Subsequently, the shallow and deep features are spliced and fused through layer-skip connections. Sub-pixel convolutional layers are used to upsample the feature maps by a factor of two to obtain an output image of the same size as a high-quality seismic image. Finally, a convolutional layer and Tanh activation function are used to predict high-quality seismic images. Compared with the generator, the discriminator has a simpler network structure because it only needs to determine the probability that the input image is a real high-quality seismic image. The discriminator network consists of one convolutional layer and seven convolutional blocks. Each convolutional block consists of a convolutional layer, a batch normalization layer (BN) and a Leaky ReLU activation function. Finally, two fully connected layers (Dense) and a Sigmoid activation function are performed to predict the probability that the input image is a real high-quality seismic image. Adversarial learning via generator and discriminant helps the generator generate more realistic high-quality seismic images.

Based on the above embodiments, in this method, during the training process of the quality improvement model, the network parameter corresponding to the highest peak signal-to-noise ratio of the quality improvement model on the verification set is the optimal model parameter, wherein the peak signal-to-noise ratio The noise ratio is the similarity between the reconstructed high-quality seismic image by the generative adversarial network and the real high-quality seismic image.

Specifically, by alternately optimizing the generator and the discriminator, the generator generates high-quality seismic images that are closer to reality in order to deceive the discriminator, while the discriminator is optimized to improve the ability to distinguish between the image generated by the generator and the real high-quality image. Use the Adam optimizer to update the generator and discriminator network parameters. The peak signal-to-noise ratio is used to characterize the similarity between the reconstructed image and the original real high-quality seismic image. The optimal model is determined by looking at the peak signal-to-noise ratio of the model on the training set and validation set as the number of iterations increases. The model The network parameters corresponding to the highest PSNR on the validation set are saved.

Based on the above-mentioned embodiment, in this method, the model network structure used in the training process of the fault recognition model is a U-net network, and the U-net network includes an encoding branch and a decoding branch, and the encoding branch is composed of a quadratic volume The decoding branch consists of four upsampling and convolution operations.

Specifically, fault recognition is regarded as an image semantic segmentation task, and each sampling point in the image to be recognized can be divided into two types: fault and non-fault. U-net, which is commonly used in semantic segmentation, is used as the fault recognition network. U-net is divided into an encoding branch and a decoding branch. The encoding branch consists of four convolution operations and downsampling, and the decoding branch consists of four upsampling and convolution operations. . Layer-skip connections fuse the upsampled feature maps of the decoding branch with the convolved feature maps of the encoding branch. Finally, a convolutional layer with a kernel size of 1×1 and a Sigmoid activation function are used to output the predicted tomographic probability map, which is the same size as the input image to be recognized.

Based on the above embodiment, in the method, in the loss function of the fault recognition model training process, the weight assigned to pixels representing faults is greater than the weight assigned to pixels representing non-faults.

Specifically, there is a strong sample imbalance in fault semantic segmentation since positive samples (fault pixels) are much less than negative samples (non-fault pixels). To solve this problem, a focal loss is used to give higher weights to pixels representing faults, as follows:

where N represents the number of pixels in the fault probability map, _yi represents the true binary label (0 for non-fault, 1 for fault), and _pi represents the predicted probability of the fault identification network. It is used to adjust the weight of loss value of positive and negative samples, and is used to control the weight of sample learning difficulty. According to our experimental experience, in this embodiment, α=0.9 and γ=2 are set.

The following describes the automatic identification device for high-resolution tomograms based on deep learning provided by the present invention, the automatic identification device for high-resolution tomograms based on deep learning described below is the same as the automatic identification device for high-resolution tomograms based on deep learning The identification methods can be referred to in correspondence with each other.

Fig. 2 is a schematic structural diagram of an automatic identification device for high-resolution faults based on deep learning provided by the present invention. As shown in Fig. 2, the device includes an acquisition unit 210, a quality improvement unit 220, and a fault identification unit 230, wherein,

The acquiring unit 210 is configured to acquire seismic images to be identified;

The quality improvement unit 220 is configured to input the seismic image to be identified into a quality improvement model, and output the enhanced seismic image;

Wherein, the quality improvement model is obtained after training on a first training set and a first verification set constructed based on a plurality of sample low-quality seismic images and corresponding high-quality seismic image labels, and the first training set and the The first verification set is generated by random parameter simulation;

The fault identification unit 230 is configured to input the lifted seismic image into a fault identification model and output a fault probability map;

Wherein, the fault recognition model is obtained after training on a second training set and a second verification set constructed based on a plurality of sample seismic images and corresponding fault probability map labels.

The device provided by the present invention obtains the seismic image to be identified; inputs the seismic image to be identified into a quality improvement model, and outputs the enhanced seismic image; wherein, the quality improvement model is based on a plurality of samples of low-quality seismic images and corresponding The first training set and the first verification set constructed by high-quality seismic image labels are obtained after training, and the first training set and the first verification set are generated by random parameter simulation; the improved seismic images are input into the fault The identification model outputs a fault probability map; wherein, the fault identification model is obtained after training on a second training set and a second verification set constructed based on a plurality of sample seismic images and corresponding fault probability map labels. A cleaner and clearer fault probability map is achieved, and the predicted fault position is more accurate.

Based on the foregoing embodiments, the device also includes a comparison unit for:

input the seismic image to be identified into the fault identification model, and output the direct identification result of the seismic image;

The fault probability map is compared with the seismic image direct identification result.

Based on the above embodiment, in this device, the first training set and the first verification set are generated through random parameter simulation, specifically including:

Generate formation reflectance models according to preset rules;

Generate different types of formation reflectance models through random parameter control;

Convolve any stratum reflectivity model with low-frequency Reich wavelets and add random noise to the seismic data for twice down-sampling to obtain sample low-quality seismic images;

The corresponding high-quality seismic image labels are obtained by convolving any stratum reflectivity model with the high-frequency Rake wavelet.

Based on the above embodiment, in this device, the network structure of the quality improvement model during training is a generative adversarial network, the generative adversarial network includes a generator and a discriminator, wherein the generator includes a skip layer connection and multiple residual connections, the discriminator consists of one convolutional layer and seven convolutional blocks.

Based on the above embodiment, in this device, during the training process of the quality improvement model, the network parameter corresponding to the highest peak signal-to-noise ratio of the quality improvement model on the verification set is the optimal model parameter, wherein the peak signal-to-noise ratio The noise ratio is the similarity between the reconstructed high-quality seismic image by the generative adversarial network and the real high-quality seismic image.

Based on the above-mentioned embodiment, in this device, the model network structure used in the training process of the fault recognition model is a U-net network, and the U-net network includes an encoding branch and a decoding branch, and the encoding branch consists of a quadratic volume The decoding branch consists of four upsampling and convolution operations.

Based on the above-mentioned embodiment, in the device, in the loss function of the fault recognition model training process, the weights assigned to pixels representing faults are greater than the weights assigned to pixels representing non-faults.

Based on the above embodiments, the present invention provides a method for improving the resolution and noise reduction of seismic images based on synthetic seismic data using a generative countermeasure network, including the following steps:

S1. Construct training data set, build seismic data quality improvement neural network, train and verify neural network;

S2. Construct a training data set, build a fault recognition neural network, train and verify the neural network;

S3. Establish an end-to-end workflow for automatic identification of high-resolution faults based on the seismic data quality improvement neural network and fault identification neural network;

S4. The real seismic data in the field are directly sent to the fault identification network for fault identification and the results of fault identification by the method proposed in this paper are compared to verify the effectiveness of the method proposed in this paper.

Specifically, in S1, it can be divided into three stages. The first stage is the establishment of the training data set, which uses the deep learning method of supervised learning. In order to establish the mapping relationship between low-quality seismic data and high-quality seismic data, a large amount of low-quality seismic data and their corresponding labels (high-quality seismic data) are required. Earthquake data) are used for training, however, it is very difficult to collect such data in reality. Therefore, the training data is generated by random parameter simulation. First, a reflectance model with horizontal layers is randomly generated. Then, random vertical perturbations are added to the horizontal layer model to simulate formation dip. Next, Gaussian perturbations were randomly added to the model to simulate the fold structure. Next, the real fault distribution is simulated by adding volumetric vector field perturbations to the model. So far, different types of formation reflectance models can be generated through stochastic parameter control. Finally, the reflectivity model is convolved with the low-frequency Reich wavelet and the seismic data with random noise added is twice down-sampled to obtain low-quality seismic data, and the reflectivity model is convolved with the high-frequency Reich wavelet to obtain the corresponding High-quality seismic data. Through different combinations of random parameters, a large number and variety of seismic data pairs can be generated. The low-quality seismic data is used as the input of the seismic data quality improvement neural network, and the high-quality seismic data is used as the label. The training set and the verification set are divided according to the ratio of 8:2, which completes the preparation of the training data. The second stage is the construction of the neural network. FIG. 3 is a structural diagram of the neural network for improving the quality of seismic data provided by the present invention. The network for improving the quality of seismic data is built based on the currently popular generative adversarial network (as shown in FIG. 3 ). The network structure consists of a generator and a discriminator. As shown in Fig. 3(a), the network structure of the generator consists of a skip connection and multiple residual connections. First, shallow features of low-quality seismic images are extracted through a convolutional layer and a parameterized ReLU (PReLU) activation function. Then, five convolutional blocks with residual connections (as shown in Fig. 3(c)) are used to extract deep features. Subsequently, the shallow and deep features are spliced and fused by layer-skip connections. Sub-pixel convolutional layers are used to upsample the feature maps by a factor of two to obtain an output image of the same size as a high-quality seismic image. Finally, a convolutional layer and Tanh activation function are used to predict high-quality seismic images. As shown in Fig. 3(b), compared with the generator, the network structure of the discriminator is simpler because it only needs to determine the probability that the input image is a real high-quality seismic image. The discriminator network consists of one convolutional layer and seven convolutional blocks (Fig. 3(d)). Each convolutional block consists of a convolutional layer, a batch normalization layer (BN) and a Leaky ReLU activation function. Finally, two fully connected layers (Dense) and a Sigmoid activation function are performed to predict the probability that the input image is a real high-quality seismic image. Adversarial learning via generator and discriminant helps the generator generate more realistic high-quality seismic images. The third stage is the training and verification of the neural network. The training of the generative confrontation network is divided into the training of the generator and the training of the discriminator. The loss function of the definition generator is the sum of the mean square error loss plus the perceptual loss and the adversarial loss.

The loss for the generator is defined as follows:

l _G ＝l _MSE +0.6×l _VGG +0.1×l _adv

where lMSE denotes the pixel-level MSE loss, lVGG denotes the VGG loss, and ladv denotes the adversarial loss.

lMSE calculates the pixel-by-pixel mean square error between the generated image and its corresponding real image, and the calculation formula is as follows:

In the formula, W and H represent the dimensions of low-resolution noisy seismic image ILR, and t represents the multiple of resolution improvement from ILR to IHR.

VGG networks have been proven to have strong feature extraction capabilities. The quality of the production image is improved by calculating the difference between the generated image and the real image in the feature extraction space of the VGG network. The VGG loss function is defined based on a pre-trained 16-layer VGG network (VGG16). Calculate the relationship between the high-resolution noise-free seismic image generated by the generator and the feature map obtained by the jth convolution (after the activation function) of the corresponding real image IHR before the i-th maximum pooling layer in the VGG16 network Euclidean distance. The Euclidean distance is used as the VGG loss to measure the similarity between the reconstructed image and the real image in the VGG network feature representation space, and the calculation formula is as follows:

In the formula, φ _i,j represents the operation of extracting the feature map of the j-th convolutional layer (after the activation function) before the i-th maximum pooling layer, W _i,j and H _i,j represent each The dimension of the feature map of the feature extraction layer.

To fool the discriminator, the generator learns the distribution of true high-resolution noise-free seismic images as much as possible. Therefore, ladv is defined in terms of the probability that the discriminator recognizes the reconstructed seismic image as a true high-resolution noise-free image:

where N represents the number of low-resolution noisy samples, and represents the probability that the discriminator regards the seismic image reconstructed by the generator as a true high-resolution noise-free seismic image.

The loss of the discriminator is only the adversarial loss. By alternately optimizing the generator and the discriminator, the generator is designed to deceive the discriminator to generate more realistic high-quality seismic data, while the discriminator is optimized to improve the ability to distinguish between the image generated by the generator and the real high-quality image. Use the Adam optimizer to update the generator and discriminator network parameters. The peak signal-to-noise ratio is used to characterize the similarity between the reconstructed image and the original image. Fig. 4 is a schematic diagram of the training process of the seismic data quality improvement neural network provided by the present invention. As the number of iterations increases, the peak signal-to-noise ratio changes to determine the optimal model, and the network parameters corresponding to the highest peak signal-to-noise ratio of the model on the verification set are saved for subsequent prediction tasks of the model.

It should be noted that, in the present invention, the randomly given high-quality seismic data has a peak frequency in the range of 30-55 Hz, and the low-quality seismic data has a peak frequency in the range of 10-30 Hz. Also given the signal-to-noise ratio of random noise in low-quality seismic data ranges from 3 to 13. For the division ratio of training data, it is divided according to the criterion of 80% for training and 20% for verification. In other embodiments, other parameters may be used for data simulation and training data set division.

Specifically, in S2, it can also be divided into three stages. The first stage is the establishment of the training data set, using the public data set (Wu et al., 2020) to prepare the training data set and the verification data set in a ratio of 8:2. Next is the construction of the fault recognition neural network in the second stage. Fault recognition is regarded as an image semantic segmentation task. Each sampling point in the seismic image can be divided into two types: fault and non-fault. Fig. 5 is the structural diagram of the neural network structure for earthquake fault recognition provided by the present invention, using U-net commonly used in semantic segmentation as the fault recognition network, its network structure as shown in Fig. 5, U-net is divided into encoding branch and decoding branch, encoding The branch consists of four convolution operations and downsampling, and the decoding branch consists of four upsampling and convolution operations. Layer-skip connections fuse the upsampled feature maps of the decoding branch with the convolved feature maps of the encoding branch. Finally, a convolutional layer with a kernel size of 1 × 1 and a sigmoid activation function are used to output the predicted fault probability map, which is the same size as the input seismic image. The third stage is the training and verification of the fault recognition network. Since the positive samples (fault pixels) are far less than the negative samples (non-fault pixels), there is a strong sample imbalance in fault semantic segmentation. In order to solve this problem, the focal loss is used to give higher weight to the pixels representing the fault, and the loss function L _fault of the fault recognition network is defined as follows:

where N represents the number of pixels in the fault probability map, _yi represents the true binary label (0 for non-fault, 1 for fault), and _pi represents the predicted probability of the fault identification network. It is used to adjust the weight of loss value of positive and negative samples, and is used to control the weight of sample learning difficulty. According to experimental experience, in the present invention, α=0.9 and γ=2 are set.

During training, the input seismic data were normalized to eliminate the influence of data distribution differences, and data augmentation operations of image flipping and rotation were performed to enhance the robustness of the model. The Adam optimizer is used to update the parameters of the network. Figure 6 is a schematic diagram of the training process of the fault recognition neural network provided by the present invention. The training process of the fault recognition network is shown in Figure 6, and the network parameters corresponding to the minimum loss of the model on the verification set are saved for subsequent fault prediction of the model Task. In other embodiments, other optimizers, training parameters, loss functions and data augmentation methods may be used.

Specifically, in S3, an end-to-end workflow for high-resolution fault identification of field seismic data is constructed. Fig. 7 is a schematic diagram of the end-to-end workflow of high-resolution fault intelligent identification provided by the present invention. The workflow is formed by the neural network trained in S1 and S2 in series. First, the original seismic image (as shown in Fig. 7(a) shown) is sent to the trained data quality improvement network in S1 for resolution improvement and random noise suppression to obtain high-quality seismic data (as shown in Figure 7(b)), and then the improved data quality It is sent to the fault recognition network in S2 for automatic fault recognition, and high-resolution fault recognition results are obtained (as shown in Fig. 7(c)). This workflow effectively integrates the two tasks of seismic data quality improvement and fault identification. Through the proposed method, high-precision fault identification results can be obtained directly from the original low-frequency seismic images full of random noise.

Specifically, in S4, compared with the conventional method, the result of directly sending the original seismic image into the fault identification network for fault identification is compared with the proposed result of first improving the quality of the seismic image and then performing fault identification to verify the proposed The effectiveness and advancement of the new method. First, the collected field seismic data are directly sent to the fault identification network trained in S2 for fault identification. Then, using the workflow proposed in S3, the data quality of the field seismic data is first improved and then the fault identification is performed. Finally, compare the fault identification results of the above two methods. In the present invention, Fig. 8 is a schematic diagram of the application on the real seismic data in the field provided by the present invention and the comparison with the traditional implementation process. First, the well-trained seismic data quality improvement network in S1 is used to improve the original seismic image (as shown in Fig. 8 ( a) Shown) for quality improvement. Figure 8(b) is the improved seismic image. Compared with the original seismic image, the random noise has been significantly suppressed, and some high-frequency geological features, such as faults and thin layers, are easier to observe with the naked eye. Then, the quality-improved seismic images are fed to the trained fault identification network to obtain high-resolution fault identification results (as shown in Fig. 8(d)). Comparing this result with the result of the conventional fault identification process (feeding the raw seismic image (as shown in Figure 8(a)) directly to the fault identification network) (as shown in Figure 8(b)), the proposed high-resolution The faults identified by the fault identification method are more refined and clean, and the identified noise interference is less. In addition, when some adjacent faults are directly recognized through the original image, the recognition results are mixed together, but the recognition results of the proposed method are separated and well-characterized . More importantly, some small-scale faults were not identified due to the low resolution of the original seismic image and the high random noise contained in it. However, after the quality of the seismic data was improved, these small-scale faults were detected by the neural network for fault identification. OK detected. Further comparing Fig. 8(d) and Fig. 8(b), the fault identified by the method is sharper than the traditional method, which will provide more accurate fault location information, which has great potential for subsequent accurate geological modeling. s help. These all show the superiority of the proposed new method for high-resolution fault identification compared to existing fault identification methods, especially for low-quality seismic data under complex geological conditions.

The present invention proposes an end-to-end high-resolution fault automatic identification workflow based on deep learning technology. The workflow consists of two neural networks, one for seismic image quality enhancement and the other for seismic fault identification. First, the raw seismic images are fed into a well-trained seismic data quality improvement network to obtain high-quality seismic images. It is then fed into a well-trained fault recognition network to obtain high-resolution fault recognition results. Due to the lack of a large amount of labeled seismic data for the training and verification of the neural network, in order to overcome this problem, a large amount of training data is obtained through random parameter simulation technology. Then, the proposed method is applied to real seismic data in the field, and the results show that the method can obtain high-precision fault identification results from raw seismic images without any artificially labeled data. And the proposed method is compared with the conventional fault identification method. The results show that the proposed high-resolution fault identification method identifies faults that are more precise and clean, and the identified noise interference is less. In addition, when some adjacent faults are directly recognized through the original image, the recognition results are mixed together, but the recognition results of the proposed method are separated and well-characterized . More importantly, some small-scale faults were not identified due to the low resolution of the original seismic image and the high random noise contained in it. However, after the quality of the seismic data was improved, these small-scale faults were detected by the neural network for fault identification. OK detected. This means that the method will provide more accurate fault location information, which is of great help to the subsequent accurate geological modeling work. These all show the superiority of the proposed new method for high-resolution fault identification compared to existing fault identification methods, especially for low-quality seismic data under complex geological conditions.

FIG. 9 illustrates a schematic diagram of the physical structure of an electronic device. As shown in FIG. 9, the electronic device may include: a processor (processor) 910, a communication interface (Communications Interface) 920, a memory (memory) 930, and a communication bus 940, Wherein, the processor 910 , the communication interface 920 , and the memory 930 communicate with each other through the communication bus 940 . The processor 910 can call the logic instructions in the memory 930 to execute the method for automatic identification of high-resolution faults based on deep learning. The method includes: acquiring seismic images to be identified; inputting the seismic images to be identified into the quality improvement model, and outputting Seismic images after upgrading; wherein, the quality improvement model is obtained after training on a first training set and a first verification set constructed based on a plurality of sample low-quality seismic images and corresponding high-quality seismic image labels, and the first The training set and the first verification set are generated by random parameter simulation; the upgraded seismic image is input into a fault identification model, and a fault probability map is output; wherein, the fault identification model is based on a plurality of sample seismic images and corresponding faults obtained after training the second training set and the second verification set constructed with probability map labels.

In addition, the above-mentioned logic instructions in the memory 930 may be implemented in the form of software function units and be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the essence of the technical solution of the present invention or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including Several instructions are used to make a computer device (which may be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program codes. .

On the other hand, the present invention also provides a computer program product, the computer program product includes a computer program stored on a non-transitory computer-readable storage medium, the computer program includes program instructions, and when the program instructions are executed by a computer During execution, the computer can execute the deep learning-based automatic identification method of high-resolution faults provided by the above methods, the method includes: acquiring the seismic image to be identified; inputting the seismic image to be identified into the quality improvement model, and outputting the improved post-seismic image; wherein, the quality improvement model is obtained after training on a first training set and a first verification set constructed based on a plurality of sample low-quality seismic images and corresponding high-quality seismic image labels, and the first training The set and the first verification set are generated by random parameter simulation; the lifted seismic image is input into a fault identification model, and a fault probability map is output; wherein, the fault identification model is based on a plurality of sample seismic images and corresponding fault probability obtained after training on the second training set and the second verification set constructed with graph labels.

In yet another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it is implemented to perform the above-mentioned high-resolution tomography based on deep learning. An automatic identification method, the method comprising: acquiring a seismic image to be identified; inputting the seismic image to be identified into a quality improvement model, and outputting the enhanced seismic image; wherein, the quality improvement model is based on a plurality of samples of low-quality seismic images and The first training set and the first verification set constructed by corresponding high-quality seismic image labels are obtained after training, and the first training set and the first verification set are generated by random parameter simulation; A fault identification model is input, and a fault probability map is output; wherein, the fault identification model is obtained after training on a second training set and a second verification set constructed based on a plurality of sample seismic images and corresponding fault probability map labels.

The device embodiments described above are only illustrative, and the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in One place, or it can be distributed to multiple network elements. Part or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. It can be understood and implemented by those skilled in the art without any creative effort.

Through the above description of the implementations, those skilled in the art can clearly understand that each implementation can be implemented by means of software plus a necessary general hardware platform, and of course also by hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the prior art can be embodied in the form of software products, and the computer software products can be stored in computer-readable storage media, such as ROM/RAM, magnetic discs, optical discs, etc., including several instructions to make a computer device (which may be a personal computer, server, or network device, etc.) execute the methods described in various embodiments or some parts of the embodiments.

Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it can still be Modifications are made to the technical solutions described in the foregoing embodiments, or equivalent replacements are made to some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims (10)

1. A method for automatic identification of high-resolution faults based on deep learning, characterized in that, comprising: Obtain the seismic image to be identified; Input the seismic image to be identified into the quality improvement model, and output the enhanced seismic image; Wherein, the quality improvement model is obtained after training on a first training set and a first verification set constructed based on a plurality of sample low-quality seismic images and corresponding high-quality seismic image labels, and the first training set and the The first verification set is generated by random parameter simulation; Inputting the upgraded seismic image into a fault identification model, and outputting a fault probability map; Wherein, the fault recognition model is obtained after training on a second training set and a second verification set constructed based on a plurality of sample seismic images and corresponding fault probability map labels. 2. the automatic identification method of the high-resolution fault based on deep learning according to claim 1, is characterized in that, also comprises: input the seismic image to be identified into the fault identification model, and output the direct identification result of the seismic image; The fault probability map is compared with the seismic image direct identification result. 3. the automatic identification method of the high-resolution fault based on deep learning according to claim 1, is characterized in that, described first training set and described first validation set are generated by random parameter simulation, specifically comprising: Generate formation reflectance models according to preset rules; Generate different types of formation reflectance models through random parameter control; Convolve any stratum reflectivity model with low-frequency Reich wavelets and add random noise to the seismic data for twice down-sampling to obtain sample low-quality seismic images; The corresponding high-quality seismic image labels are obtained by convolving any stratum reflectivity model with the high-frequency Rake wavelet. 4. the automatic identification method of the high-resolution fault based on deep learning according to claim 3, it is characterized in that, the network structure of described quality improvement model in the training process is generation confrontation network, and described generation confrontation network comprises a A generator and a discriminator, wherein the generator includes a layer skip connection and multiple residual connections, and the discriminator includes a convolutional layer and seven convolutional blocks. 5. the automatic identification method of the high-resolution fault based on deep learning according to claim 4, is characterized in that, described quality promotion model is in training process, and described quality promotion model has the highest peak signal-to-noise ratio on verification set The corresponding network parameters are optimal model parameters, and the peak signal-to-noise ratio is the similarity between the high-quality seismic image reconstructed by the generative adversarial network and the real high-quality seismic image. 6. the automatic identification method of the high-resolution fault based on deep learning according to claim 1, is characterized in that, the model network structure used in the training process of described fault identification model is U-net network, and described U-net net network includes an encoding branch consisting of four convolution operations and downsampling, and a decoding branch consisting of four upsampling and convolution operations. 7. the automatic identification method of the high-resolution fault based on deep learning according to claim 6, it is characterized in that, in the loss function of described fault recognition model training process, the weight given to the pixel representing fault is greater than that representing non-fault The weights assigned to the pixels of . 8. An automatic identification device for high-resolution faults based on deep learning, characterized in that it comprises: an acquisition unit, configured to acquire a seismic image to be identified; a quality improvement unit, configured to input the seismic image to be identified into a quality improvement model, and output the enhanced seismic image; Wherein, the quality improvement model is obtained after training on a first training set and a first verification set constructed based on a plurality of sample low-quality seismic images and corresponding high-quality seismic image labels, and the first training set and the The first verification set is generated by random parameter simulation; a fault identification unit, configured to input the lifted seismic image into a fault identification model, and output a fault probability map; Wherein, the fault recognition model is obtained after training on a second training set and a second verification set constructed based on a plurality of sample seismic images and corresponding fault probability map labels. 9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and operable on the processor, characterized in that, when the processor executes the program, any of the following claims 1 to 7 are realized. Steps of a method for automatic identification of high-resolution tomograms based on deep learning. 10. A non-transitory computer-readable storage medium, on which a computer program is stored, characterized in that, when the computer program is executed by a processor, the deep learning-based Steps in the method for automatic identification of high-resolution tomograms.

Images