CN-122017988-A - Seismic data fracture identification method based on DENSEVNET multi-scale deep learning

CN122017988ACN 122017988 ACN122017988 ACN 122017988ACN-122017988-A

Abstract

The invention relates to the technical field of seismic data fracture identification and discloses a seismic data fracture identification method based on DENSEVNET multi-scale deep learning, which comprises the following steps of acquiring an actual work area seismic data body, inputting the actual work area seismic data body into a pre-trained nonlinear mapping relation model, and outputting a fracture identification result; the model comprises an input layer, a multi-scale feature extraction module, an encoder module, a decoder module and an output layer which are sequentially connected, wherein the multi-scale feature extraction module extracts and splices different-scale fracture features in data, the encoder module digs and reduces dimensions of the spliced features to generate multi-level coding features, the decoder module restores and fuses the features, and finally a result is output through the output layer.

Inventors

LU PENGFEI
JIANG SHUHAO
GUO AIHUA

Assignees

东华理工大学

Dates

Publication Date: 20260512
Application Date: 20260131

Claims (10)

1. A seismic data fracture identification method based on DENSEVNET multi-scale deep learning is characterized by comprising the following steps: Acquiring an actual work area seismic data body, inputting the actual work area seismic data body into a pre-trained nonlinear mapping relation model, and outputting a fracture identification result; The nonlinear mapping relation model comprises an input layer, a multi-scale feature extraction module, an encoder module, a decoder module and an output layer which are sequentially connected; The multi-scale feature extraction module is used for respectively extracting fracture features of different scales in an actual work area seismic data body input through the input layer, splicing the fracture features of different scales, excavating and dimension-reducing the spliced multi-scale features through the encoder module to generate multi-level coding features containing semantic information of different levels, and the decoder module is used for respectively restoring and fusing the generated multi-level coding features and finally outputting fracture identification results through the output layer.
2. The method for identifying the fracture of the seismic data based on DENSEVNET multi-scale deep learning according to claim 1, wherein the multi-scale feature extraction module comprises a first-scale convolution layer, a second-scale convolution layer and a third-scale convolution layer which are arranged in parallel, the first-scale convolution layer is used for extracting first fracture features, the second-scale convolution layer is used for extracting second fracture features, the third-scale convolution layer is used for extracting third fracture features, the sizes of the first fracture features, the second fracture features and the third fracture features are sequentially increased, the convolution kernel sizes of the first-scale convolution layer, the second-scale convolution layer and the third-scale convolution layer are 3, 5 and 7 respectively, the step sizes are 2, the filter numbers are 32, and the activation functions are ReLU.
3. The method for identifying the fracture of the seismic data based on DENSEVNET multi-scale deep learning according to claim 2, wherein the encoder module comprises three dense layers and two sampling layers, the first dense layer is used for receiving the spliced multi-scale features, the first downsampling layer carries out average pooling on the output result of the first dense layer to obtain a first feature, the first feature is input into the second dense layer, the second downsampling layer carries out average pooling again on the output result of the second dense layer to obtain a second feature, and the second feature is input into the third dense layer to obtain a third feature; The decoder module comprises four convolution blocks and three up-sampling layers, wherein a first feature is connected with the first convolution block to obtain a fourth feature, a second feature is connected with the second convolution block and then connected with the first up-sampling layer to obtain a fifth feature, a third feature is connected with the third convolution block and then connected with the second up-sampling layer to obtain a sixth feature, the fourth feature, the fifth feature and the sixth feature are spliced, the spliced result is connected with the fourth convolution block, the output result of the fourth convolution block is connected with the third up-sampling layer, and the up-sampled result is connected with the output layer.
4. The method for identifying the fracture of the seismic data based on DENSEVNET multi-scale deep learning according to claim 3, wherein the first dense layer comprises 2 layers of convolution cycles, the number of filters of the first dense layer is 64, the second dense layer comprises 3 layers of convolution cycles, the number of filters of the second dense layer is 128, the third dense layer comprises 4 layers of convolution cycles, the number of filters of the third dense layer is 256, and the convolution kernels of the inner convolution layers of the first dense layer, the second dense layer and the third dense layer are all 3 and are connected by jump.
5. A method of seismic data fracture identification based on DENSEVNET multi-scale deep learning as claimed in claim 3, wherein the pooling scale of the first and second downsampling layers is 2, the upsampling scale of the first and third upsampling layers is 2, and the upsampling scale of the second upsampling layer is 4.
6. The method for identifying the fracture of the seismic data based on DENSEVNET multi-scale deep learning according to claim 3 is characterized in that the first to fourth convolution blocks are formed by sequentially connecting two convolution layers, the first convolution block is used for carrying out feature refinement on a first feature, the two convolution layer parameters are respectively 64 in number, 1 in number of convolution kernels, 1 in number of steps and a ReLU in activation function, the second convolution layer filter number is 64, 3 in number of convolution kernels, 1 in number of steps and a ReLU in activation function, the second convolution block is used for carrying out feature refinement on the second feature, the two convolution layer parameters are respectively 128 in number of convolution layers, 1 in number of convolution kernels, 1 in number of steps, a ReLU in activation function, 128 in number of convolution kernels, 3 in number of steps and 1 in number of activation functions, the third convolution block is used for carrying out feature refinement on a third feature, the two convolution layer parameters are respectively 256 in number of convolution layers, 1 in number of convolution kernels, 3 in number of steps and a fourth convolution function, the second convolution block is a number of steps is 1 in number of convolution kernels, and the second convolution block is a fourth convolution block is a number of activation function, and the second convolution block is a fourth feature refinement, and the second convolution block is a feature refinement, and the second feature is a feature in number of 16.
7. The method for recognizing the fracture of the seismic data based on DENSEVNET multi-scale deep learning according to claim 1 is characterized in that a nonlinear mapping relation model is trained, a fracture synthetic seismic data set is divided into a training data set and a verification data set according to a preset proportion during training, after the nonlinear mapping relation model is trained based on the training data set, the verification data set is utilized again for training, and when a loss function of the nonlinear mapping relation model meets a preset precision requirement, training is finished, and a trained nonlinear mapping relation model is obtained.
8. The method for identifying fracture of seismic data based on DENSEVNET multi-scale deep learning as claimed in claim 7, wherein the process of acquiring the fracture synthetic seismic data set includes the steps of acquiring a fracture synthetic seismic data volume first, forming an initial fracture synthetic seismic data set, and then carrying out normalization processing and data enhancement processing on the initial fracture synthetic seismic data set, so as to obtain the fracture synthetic seismic data set after preprocessing, wherein the data enhancement processing adopts at least one of noise adding, inversion and mirroring.
9. The DENSEVNET multi-scale depth learning based seismic data fracture identification method of claim 7, wherein the loss function is a balanced cross entropy loss function determined based on: ; In the formula, For the Focal cross entropy loss value, In order to balance the cross entropy loss values, For the loss value after the Focal cross entropy loss value and the balance cross entropy loss value are fused, For the Focal cross entropy loss value weight, The value weights are lost for balanced cross entropy.
10. The method for identifying the fracture of the seismic data based on DENSEVNET multi-scale deep learning according to claim 9, wherein training of the nonlinear mapping relation model adopts an Adam optimizer, the learning rate is set to be 1e-4, the batch processing size is 1, the iteration number is 100 times or more, when the loss function of the verification set is iterated for 10 times continuously and the value of the loss function is not reduced, training is stopped, and the current nonlinear mapping relation model is stored.

Description

Seismic data fracture identification method based on DENSEVNET multi-scale deep learning Technical Field The invention relates to the technical field of seismic data fracture identification, in particular to a seismic data fracture identification method based on DENSEVNET multi-scale deep learning. Background The seismic data fracture identification is a key link in seismic interpretation and oil and gas exploration, and a fracture system not only controls the structure evolution and the sedimentation pattern, but also directly influences the migration, aggregation and preservation of oil and gas. The major and minor fractures determine the structural trend of the oil-containing fracture and the shape and the appearance of the fracture, and the minor fractures further control the local micro-amplitude structure, divide the oil-containing fracture and complicate the oil-water relationship of the oil-containing fracture. The large fracture transverse extension length exceeds 1km, the breaking distance is larger than 20m, the medium fracture transverse extension length is not larger than 1km, the breaking distance is smaller than 20m, the small fracture transverse extension length is not larger than 500m, and the breaking distance is smaller than 10m. As the exploration and development of oil fields further deepens, the targets of fracture identification have turned to configurations such as small fractures or fractured zones, i.e., small fractures are identified as much as possible in the case of large and medium fractures. The small fracture has three meanings, namely that the small fracture is difficult to find in the seismic data of conventional exploration by using the conventional fracture identification standard, the fracture distance and the extension length of the fracture are small in scale, and the extension on a plane is short, so that the small fracture can be used for conveying oil gas, and can also be used for laterally plugging the oil gas and controlling the oil gas to be stored. Small fractures on a seismic section generally have the characteristics of small dislocation, in-phase axis distortion, weak amplitude and the like, have stronger concealment and large identification difficulty, and have important influences on residual oil distribution, unconventional reservoir transformation and well pattern deployment. Currently, fracture identification methods are mainly classified into three categories: The first is a traditional identification method based on seismic attributes, and the method mainly comprises the technologies of coherence, curvature, variance, ant body and the like. These methods perform fracture detection based on geometric, amplitude or phase characteristics of the seismic signals. For example, the coherence body identifies discontinuous areas by calculating the similarity between adjacent channels, curvature attributes infer fracture positions by using formation curvature changes, and ant body technology tracks fracture tracks by simulating ant colony foraging behaviors. Although these methods have contributed to the detection of breaks to some extent, there are significant limitations. The ant body technology simulates ant foraging, is sensitive to small faults when applied to the small faults, and has the most interference, such as obvious noise interference and obvious polynaphy, the coherence body, curvature and variance are based on the difference between the same-phase axes of the earthquakes corresponding to the faults, a time window is needed in the process of calculating the difference, and the time window smoothing effect influences the resolution and precision of the small fault earthquake response characteristic identification of the same-phase axes with tiny dislocation and the like, so that the sensitivity to the small faults is poor. The second type is a fracture identification method based on edge detection and image processing, the method regards seismic data as an image, and fracture enhancement is carried out by adopting technologies such as edge detection, gradient enhancement, structure tensor and the like. Although the visual effect is achieved, the method often lacks space continuity constraint in geological sense, and is difficult to distinguish fracture response from lithology change, noise and other interferences. The third type is fracture recognition method based on deep learning, and along with the development of deep learning technology, convolutional Neural Network (CNN), U-Net and other structures have been primarily applied to seismic interpretation and fracture detection. Such methods are capable of automatically learning features from data, exhibiting superior performance over conventional methods over a portion of the public data set. However, the existing deep learning fracture recognition method still has the problem that the detection capability of small targets (such as small fractures) is insufficient and the detail loss is serious. I