CN110609320A

CN110609320A - Pre-stack seismic reflection pattern recognition method based on multi-scale feature fusion

Info

Publication number: CN110609320A
Application number: CN201910802745.2A
Authority: CN
Inventors: 钱峰; 袁英淏; 胡光岷
Original assignee: University of Electronic Science and Technology of China
Current assignee: University of Electronic Science and Technology of China
Priority date: 2019-08-28
Filing date: 2019-08-28
Publication date: 2019-12-24
Anticipated expiration: 2039-08-28
Also published as: CN110609320B

Abstract

The invention discloses a pre-stack seismic reflection pattern recognition method based on multi-scale feature fusion, which is applied to the field of seismic reflection pattern recognition and aims to solve the problems that due to the fact that the data volume of pre-stack seismic signals is huge, tags are difficult to be added to each datum in a manual marking mode and single feature representation is insufficient in the existing pre-stack seismic emission pattern recognition; according to the method, a multi-scale feature fusion network is constructed, a countermeasure network is generated by introducing deep convolution, the network structure is improved, and low-level features and high-level features of the pre-stack seismic signals can be effectively extracted; and a fusion module is added on the basis of the improved convolution generation countermeasure network, and the complete representation of the reflection mode of the pre-stack seismic signal is obtained by carrying out multi-scale fusion on the high-level and low-level features.

Description

Pre-stack seismic reflection pattern recognition method based on multi-scale feature fusion

Technical Field

The invention belongs to the field of seismic reflection pattern recognition, and particularly relates to a prestack seismic reflection pattern recognition technology.

Background

With the emergence of a large number of emerging technologies such as deep learning, the traditional oil exploration industry also begins to adopt advanced technologies in a large scale to improve the production efficiency, so that the oil industry develops rapidly, and huge economic benefits are brought to national economy. However, with the increasing capacity of oil exploration, it becomes increasingly difficult to find new oil and gas fields under resource-limited conditions. Therefore, people need to know and understand the storage state of the underground oil and gas field in a more scientific way, improve the exploration prediction capability, and carry out the prediction research of the oil and gas field by mining more new information from the existing geophysical, geological, oil deposit development and other data.

The industry typically uses seismic signals to explore oil and gas fields. In order to characterize the lithology of the earth formations in detail and to explain the subsurface formations, it is desirable to obtain as much information as possible from the acquired seismic signals. The differences in seismic signals are mainly due to the differences in subsurface distribution of reservoir characteristics (presence of hydrocarbons) and structural characteristics (e.g., caverns, faults). The industry typically focuses on the seismic signals of a target hydrocarbon reservoir, finds the differences between them and non-target seismic signals, and classifies these signals according to the degree of similarity between the seismic signals. The phenomenon or characteristic of the same seismic signal different from that of other seismic signals is called as a reflection mode of the seismic signal, and the reflection mode mainly focuses on the difference between different seismic signals and the similarity between the same seismic signal. Therefore, by identifying the reflection mode of the seismic signals, the collected seismic signals can be distinguished and classified, and the distribution of various geological structures in the target work area is researched.

With the popularization of seismic stratigraphy in the exploration field, the method of utilizing seismic facies to conduct geological research is widely applied. The seismic facies reflect the sum of facies depositional representations on seismic sections of the same area. The seismic facies division is carried out according to the seismic facies type determined by seismic reflection pattern recognition on the section, and is a powerful method for researching the sedimentary facies. The method is an important method for dividing seismic facies by carrying out seismic reflection mode identification based on seismic signals. Nowadays, the technical level of acquiring seismic signals in the industry is continuously improved, and the seismic information contained in the seismic data is more abundant, so that the seismic facies are described by manually identifying the seismic reflection mode, and the method has great subjectivity and uncertainty. In order to realize the purpose of automatically and quantitatively analyzing the seismic facies, the seismic reflection mode needs to be quantitatively characterized and accurately identified by a mathematical representation method by means of a seismic data processing technology, an information processing technology and a computer technology. The accurate characteristics are the basis for completely representing the seismic reflection mode, so the method for enhancing the seismic reflection mode identification capability by intelligently extracting the characteristics of the seismic reflection signals can further improve the seismic facies division precision. The result can show the distribution of the underground geological structure, provide reference basis for the judgment of the exploration well azimuth, reduce the probability of well drilling failure and save the exploration cost.

The method is limited by the seismic signal property and the seismic signal processing technology of an actual work area, and seismic facies identification based on the post-stack waveform characteristics is the most main mode for automatically and quantitatively dividing seismic facies at present. However, as the seismic prestack data contains richer underground stratum information (such as information about variation of seismic reflection along with offset and information about variation of seismic reflection along with azimuth), the corresponding prestack seismic signal processing technology is becoming mature, and seismic facies division by using prestack seismic signals has become a popular method. Therefore, the characteristic extraction method of the pre-stack seismic signal and the post-stack seismic signal is respectively researched, and therefore the recognition capability of seismic reflection modes under different working conditions is improved. However, the existing pre-stack seismic signal feature extraction also has the following problems:

the data volume of the pre-stack seismic signals is huge, and a label is difficult to be added to each data in a manual marking mode;

the single feature characterization of the prior art is not sufficient.

Disclosure of Invention

In order to solve the technical problems, the invention provides a pre-stack seismic reflection mode identification method based on multi-scale feature fusion, which intelligently excavates features capable of representing multi-dimensional seismic reflection mode behaviors from uncertain (noise and other factors) multi-dimensional seismic data, intelligently identifies the seismic reflection modes under the unsupervised condition, and finally accurately divides the seismic facies. The method has very important significance for promoting the development of complex reservoir prediction towards refinement and quantification, improving the production efficiency, reducing the exploration and development cost, accelerating the progress of complex oil and gas reservoir exploration and development and the like.

The technical scheme adopted by the invention is as follows: a pre-stack seismic reflection pattern recognition method based on multi-scale feature fusion comprises the following steps:

s1, preprocessing the input pre-stack seismic signals to obtain a pre-stack seismic image set;

s2, constructing a multi-scale feature fusion network;

s3, training the multi-scale feature fusion network in the step S2 by adopting the pre-stack seismic image set in the step S1;

and S4, performing prestack seismic reflection mode recognition by adopting the multi-scale feature fusion network trained and completed in the step S3.

Further, step S1 is specifically: firstly, denoising an obtained pre-stack seismic signal by adopting a structure-oriented filtering algorithm; then, intercepting target data segments above and below the time dimension by referring to the position of the horizon; and finally, connecting the data sections with the same position but different azimuth angles to form a pre-stack seismic image set.

Further, the multi-scale feature fusion network of step S2 includes: fusing and generating a network; the fusion network is a network structure which extracts the low-level features and the high-level features of the pre-stack seismic signals and performs multi-scale fusion; the resulting network is a full convolutional network.

Further, the converged network includes three parts: the system comprises a multi-level feature extraction part, a multi-scale feature fusion part and a classifier part; the multilevel feature extraction part adopts four convolutional layers of Conv1, Conv2, Conv3 and Conv4 to extract multilevel features; the multi-scale feature fusion part extracts feature maps Fea2 and Fea4 from the convolutional layers Conv2 and Conv4 respectively and fuses the extracted feature maps to form new feature maps; and the classifier part judges the fused new feature map.

Further, the multi-scale fusion part is realized by the following steps:

(1) extracting characteristic maps Fea2 and Fea4 from the convolutional layers Conv2 and Conv4 respectively; fusing the feature maps Fea2 and Fea4 to form a new feature map;

(2) scaling the size, and adopting maximum pooling on the feature map Fea2 to ensure that the size of the feature map Fea2 after feature pooling is the same as that of the feature map Fea4 in the step (1);

(3) amplitude normalization, namely respectively performing amplitude normalization operation on the feature map Fea4 and the feature map Fea2 after size scaling;

(4) merging the Fea2 and the Fea4 processed in the step (3) according to the channel direction to form a new merged feature map Feaf;

(5) and (3) converting the dimension of the fused feature, and enabling the fused feature map Feaf to pass through two convolution layers: conv5 and Conv6 are converted into one-dimensional fused feature vectors, and finally input into a classifier for discrimination.

Further, the generation network generates the seismic images layer by adopting four deconvolution layers of Deconv1, Deconv2, Deconv3 and Deconv 4.

Further, the step S3 is specifically:

s31, randomly extracting m images from the pre-stack seismic image set to form a batch and normalizing the batch;

s32, extracting m 50-dimensional Gaussian noise vectors, inputting the vectors into a generation network, and generating m generated seismic images;

s33, keeping the generation network unchanged, randomly inputting the generated seismic image and the real seismic image into the fusion network, and updating parameters of the fusion network by adopting a random gradient descent method;

s34, extracting m 50-dimensional Gaussian random noises again and inputting the noises into a generation network to generate m new generated seismic images;

s35, keeping the parameters of the fusion network unchanged, and updating the parameters of the generation network by adopting a random gradient descent method;

s36, adding 1 to the iteration times until the network reaches the maximum iteration times M, and stopping training; and obtaining a multi-scale feature fusion network with complete training.

The invention has the beneficial effects that: the method for identifying the pre-stack seismic reflection mode has the following advantages:

(1) due to the fact that the data volume of the pre-stack seismic signals is huge, each piece of data is difficult to be labeled in a manual marking mode, and therefore the pre-stack seismic signals need to be analyzed in an unsupervised mode; the generated countermeasure network model is the latest unsupervised generation model and has the characteristics of strong detail learning capability, strong noise immunity and the like compared with other unsupervised models; according to the method, a deep convolution is introduced to generate a confrontation network, and the network structure is improved, so that the low-level features and the high-level features of the pre-stack seismic signals can be effectively extracted;

(2) the high-level features represent global abstract information of the pre-stack seismic signals, the low-level features represent local concrete information of the pre-stack seismic signals, and the high-level features and the low-level features respectively depict reflection modes of the pre-stack seismic signals on different levels; the characteristics of different levels have advantages and disadvantages in characterization, and aiming at the problem of insufficient single characteristic characterization of the traditional method, the invention adds a fusion module on the basis of an improved convolution generation countermeasure network, and obtains the complete characterization of the reflection mode of the pre-stack seismic signal by carrying out multi-scale fusion on the high-level and low-level characteristics.

Drawings

FIG. 1 is a flow chart of prior art seismic reflection pattern recognition;

FIG. 2 is a flow chart of a method of the present invention;

FIG. 3 is a schematic diagram of a converged network architecture;

FIG. 4 is a schematic diagram of a growth network structure;

FIG. 5 is a graph A of convolution versus deconvolution;

FIG. 6 is a graph B of convolution versus deconvolution;

FIG. 7 is a waveform of a pre-stack seismic signal for a work area;

wherein, fig. 7(a) is a real seismic waveform; FIG. 7(b) is a diagram for generating seismic waveforms;

FIG. 8 is a statistical chart of the multi-scale feature fusion network and the classification result differentiation of the self-encoder;

FIG. 9 is a prototype of physical model data;

FIG. 9(a) is an actual physical model, and FIG. 9(b) is a blueprint of the physical model;

FIG. 10 is a DCAE-based physical model prestack seismic facies plot;

FIG. 11 is a pre-stack seismic facies diagram of a physical model generated by the method of the present invention;

FIG. 12 is an amplitude profile of an LZB tool zone;

FIG. 13 is a pre-stack seismic phase plot of an LZB work area generated based on the DCAE method;

FIG. 14 is an LZB work area pre-stack seismic phase diagram generated by the method of the present invention;

FIG. 15 is an amplitude attribute of a brocade work area;

FIG. 16 is a pre-stack seismic facies diagram of a brocade work area generated based on the DCAE method;

FIG. 17 is a pre-stack seismic phase diagram of a brocade work area generated by the method of the invention.

Detailed Description

FIG. 1 is a flow chart of seismic reflection pattern recognition, which first preprocesses seismic signals and selects a suitable time window size on the horizon to obtain seismic signals of a target horizon segment. Then, the seismic signal reflection characteristics of the target interval are classified by combining a seismic signal characteristic extraction method and a machine learning classification algorithm, and corresponding seismic facies are distinguished in a class label mode, so that the distribution conditions of various target geological structures can be researched. In the process of identifying the seismic reflection mode, feature extraction and feature classification are two most critical steps, wherein the recognition of the seismic reflection mode is carried out on the premise that the features of a completely characterized target signal can be extracted, so that the seismic signal feature extraction method is the key point of the research of the invention. From the angle of input feature types, the seismic reflection pattern recognition can be classified into waveform-based classification and seismic attribute-based classification, so that the seismic reflection pattern recognition is reviewed from the two angles so as to comprehensively understand the current research situation and development dynamics at home and abroad. The following are two existing seismic reflection pattern recognition methods:

1. seismic reflection pattern recognition based on waveform classification

Seismic reflection pattern recognition based on waveform classification (hereinafter referred to as a waveform classification method) treats a single seismic trace as a waveform signal. The waveform classification method essentially identifies the patterns of different waveform signals, and is therefore called seismic waveform classification. The seismic waveform has intuitive interpretation significance, so a waveform classification technology is usually adopted in the seismic facies division process, and the waveform classification technology becomes the most key method in the current seismic reflection mode identification field. In the classical signal processing theory, the time series transform method is often used to extract the features of a waveform signal, which can be divided into time domain, frequency domain and time-frequency domain features, and we review the following three aspects:

(1) time domain waveform characteristics: the simplest time domain waveform feature is a direct waveform time sequence, Saggaf et al, Saudi America, directly input the waveform time sequence as the feature, and then use a competitive neural network to perform waveform pattern recognition. Song et al, national electronics science and technology university, further uses the multichannel waveform time series in the neighborhood as input for classification, and can reduce the influence of noise on the waveform feature extraction process by using multichannel characteristics in the time domain. France SNEA (P) science and technology center Dummy et al extracts statistical information such as peak value, zero crossing point, peak time and the like of a waveform as a time domain waveform, and compared with a direct time sequence, the time domain waveform is more accurate and more stable. The time domain waveform features have the defects of poor stability, strong uncertainty and sensitivity to the layer interpretation error.

(2) Frequency domain waveform characteristics: dummy et al extracted the total energy of the power spectrum in the power spectrum to 10%, 20%, 30%,. the time required for the total energy of the power spectrum was taken as the frequency domain waveform feature. The national institute of petroleum exploration and development, Xie, et al, think that seismic waves and voice signals have the same physical nature when propagating in a medium, so it is proposed to use a very successful Mel coefficient in voice recognition to represent seismic waveform characteristics.

(3) Time-frequency domain waveform characteristics: the time domain and frequency domain features are rough and not robust to the feature characterization of the waveform, so the research of the waveform classification method is focused on the time-frequency domain waveform feature extraction. Saraswat et al, the India mining institute, utilizes the superior dimension reduction characteristic of an artificial immune neural network to perform dimension reduction on wavelet transformation characteristics to remove redundant characteristics, and then utilizes the neural network to perform clustering, so that a very steady effect is obtained. The time-frequency domain waveform features based on wavelet transformation are not the best, and the university of domestic Dougu Du et al utilizes an empirical mode decomposition method to extract the time-frequency features of the waveform, which can retain more waveform features than a wavelet decomposition model, so that the waveform classification resolution is higher. The Italy Enni group R & D department Amendola et al and Xie et al hold the same view, namely the seismic waveform and the voice signal are similar in physical nature, except that the seismic waveform is more finely represented by the characteristics in the music field, the specific idea is to firstly adopt time-frequency transformation (such as wavelet transformation, S transformation and the like) to obtain the spectrogram of the seismic waveform, and then extract the music attribute of the seismic waveform from the spectrogram. Because the music attribute in the music field is finer and smoother than the singularity characteristic, the resolution of waveform classification can be improved. In China, Yangxilong and the like of Qinghua university screen the obtained earthquake phases firstly, and then classify the earthquake phases again, thereby researching the change details of the underground structure in the target exploration area. The method is characterized in that fractional derivatives are added in seismic reflection waveform classification research by Xushikun et al of university of Chengdu studys, a waveform set consisting of wavelet fractional derivatives is constructed, and each seismic waveform in seismic signals can be accurately matched, so that a seismic reflection mode is identified.

2. Seismic reflection pattern recognition based on seismic attributes

The biggest difference between seismic reflection pattern recognition based on seismic attributes (hereinafter referred to as seismic attribute method) and the method based on waveform classification is that the output result form is different: the seismic attribute method obtains a three-dimensional seismic phase and the waveform classification method obtains a two-dimensional seismic phase plane, so the seismic attribute method is also called as body waveform classification. Seismic attributes may be further divided into single seismic attributes and multiple seismic attributes based on the number of input seismic attributes.

(1) Single seismic attribute: herrmann et al, the American Massachusetts institute of technology, Inc. proposes a seismic facies attribute of scale invariant difference sharpness based on continuous wavelet transform, which can be used for depicting a primary seismic reflection mode and is insensitive to amplitude intensity. Gao et al, Marathon oil Inc. in USA, proposed a seismic facies attribute based on Texture Model Regression (TMR) from Regression model analysis. The specific idea is to perform linear least squares regression analysis between the actual seismic reflection signal and a preselected reference model to realize seismic phase analysis. Gao et al, university of west virginia, usa, further developed his previously proposed TMR model, the improvement being mainly embodied on a pre-selected reference model, which in his 2008 work was an adaptive phase, fixed frequency and amplitude cosine signal. In 2011 work, Gao trains cosine signals to obtain a reference model under the participation of well data, so that the resolution of seismic phase analysis can be effectively improved. The work of Liu et al of Qinghua university in China is novel and completely different from the work, and an image segmentation technology is adopted to extract seismic facies in an image with seismic attributes. The specific idea is as follows: firstly, seismic facies sampling points are marked on an image in a manual mode, and then a complete seismic facies is grown through a seed region growing technology, so that the spatial continuity of the seismic facies can be guaranteed to the maximum extent.

(2) Multiple seismic attributes: a single seismic attribute generally does not completely characterize more complex geological features, and multiple seismic attributes are often combined for analysis in describing a target reservoir. The most essential difference between the multi-seismic attributes and the single seismic attributes is that the multi-seismic attributes are also the pattern recognition problem, as well as the waveform classification. The us exxonmobil upstream research company West et al first proposed the use of multiple image textures in conjunction with a neural network classifier to perform quantitative analysis of seismic facies of three-dimensional seismic data. Linari et al, Argentine technotrol energy company, think that the time dimension information is lost by representing three-dimensional seismic data with two-dimensional image textures, and propose a new framework for three-dimensional seismic facies classification, which is used up to now after being determined. The three-dimensional seismic facies classification framework adopts a plurality of three-dimensional seismic attributes as input, and adopts algorithms such as principal component analysis and the like to perform feature fusion analysis and improve noise immunity, and finally, a hierarchical classification method is used for classification. Later researchers improve the framework continuously, Neves and Triebwasser, Saudi Armei oil company, expand the application range of the three-dimensional seismic facies classification framework of Linari and the like, and use new three-dimensional seismic attributes responding to cracks for predicting cracks of a carbonate reservoir. Sabeti et al, university of Yinbbier Jensenktechnology, tries to integrate frames of West and Linari to obtain a three-dimensional waveform classification frame capable of simultaneously processing two-dimensional seismic attributes and three-dimensional seismic attributes, and the core of the frame is to use a hierarchical clustering algorithm. However, due to the limitation of two-dimensional seismic attributes, the Sabeti framework is not superior to Linari, but rather increases complexity. A three-dimensional seismic facies classification framework of Linari et al is improved by Roy et al of Russian-Holoma university, the three-dimensional seismic facies classification framework is mainly embodied in that supervised and unsupervised body seismic facies classification is carried out on a plurality of three-dimensional seismic attribute data bodies by using a clustering algorithm with better performance such as a self-organizing neural network and GTM, and then quantitative analysis is systematically carried out on a silicalite reservoir stratum of an actual work area. In China, Zhuqiang and the like of university of Ottelian university describe the internal reservoir of geological structures such as carbonate deposition environment biological reefs by using a multi-attribute-body-based seismic reflection pattern recognition technology. The China Petroleum university Lang Xiaoling et al researches on the sand bodies of the trigonoceania in the Wanzhuang area of the North China oilfield by adopting a multi-attribute classification technology, and shows the direction of a development area in a favorable reservoir by depicting the three-dimensional spatial distribution of the sand bodies of the three-middle-section trigonoceania.

The flow of the pre-stack seismic reflection pattern recognition method based on multi-scale feature fusion provided by the invention is shown in fig. 2. Before extracting fusion characteristics, denoising the acquired pre-stack seismic signals by adopting a structure-oriented filtering algorithm, then intercepting target data segments above and below a time dimension by referring to the position of a horizon, and connecting the data segments with the same position but different azimuth angles to form a sample data set. Assuming that x represents one piece of data in a certain azimuth angle of the pre-stack seismic data, the dimensionality is v, and if the pre-stack seismic signals of a certain work area have n azimuth angles, combining the data of each piece of dataDimension n x v, which is the prestack seismic image set of the input multi-scale feature fusion networkWherein X_i∈R^n×v。

After the input data is preprocessed, the multi-scale feature fusion network is trained using the acquired set of pre-stack seismic images. After the entire network training converges, the fusion features are extracted for the prestack seismic signals using a subnetwork (fusion network). And finally, carrying out unsupervised classification on the fusion characteristics by adopting the FSOM, labeling each seismic signal, generating a corresponding pre-stack seismic facies map, and analyzing the pre-stack seismic facies division result according to the geological condition of the actual work area.

The method of the invention comprises the following steps:

s1 construction of multi-scale feature fusion network

In the field of computer vision in recent years, supervised learning by using the CNN has been applied successfully, and feature extraction is the greatest advantage of the CNN, because the CNN can directly input original training data, so that learning is performed from the training data implicitly and traditional explicit feature extraction is avoided, and the learned features can better represent the characteristics of the original data. Using a series of convolutional layers from lower to higher layers, CNNs automatically learn features from large-scale training data in an end-to-end manner. The features of different layers contain information of different layers, the features of the lower layer contain a lot of detail information, but the detail information is difficult to distinguish from the background, and the features of the higher layer are more abstract and pay less attention to the detail information.

In actual oil and gas field exploration, the pre-stack seismic signals acquired by an artificial seismic source are large in data volume and few in sample labels, so that an unsupervised learning method is needed for feature extraction. In the prestack seismic reflection pattern analysis, because the difference between the scales of geological structures concerned by people is very large, and the geological structures not only have large-area faults and riverways, but also have small-range karst caves and pores, all the structures cannot be accurately identified only by the characteristics of one scale. Therefore, the invention provides a DCGAN-based multi-scale feature fusion network for extracting the features of the pre-stack seismic signals by utilizing the characteristics of combining the advantages of unsupervised learning of the DCGAN and the advantages of feature extraction of the CNN and improving the network structure. The multi-scale feature fusion network structure is composed of a fusion network and a generation network. The fusion network is a network structure which extracts the low-level features and the high-level features of the pre-stack seismic signals and performs multi-scale fusion; the generated network is used for assisting the training process of the whole multi-scale fusion network and is a full convolution network. In order to train a network capable of accurately extracting the pre-stack seismic signal features, the two networks are combined together to perform training. In each iteration of the network training, in order for the fusion network to be able to accurately identify the true seismic signals from the generated seismic signals, the fusion network needs to be updated first. The resulting network is then updated so that it can generate enough "real" seismic signals to defeat the purpose of the converged network.

1. Converged network

The method includes acquiring each azimuth target data segment of the pre-stack seismic signals, and arranging and placing the seismic trace data (generally vectors) at the same position to form a data matrix, which can be regarded as a single-channel image, i.e., a seismic image corresponding to each seismic trace. The input of the fusion network designed by the invention is a real seismic image and a generated seismic image (generated by the generation network), and the size is generally 24 multiplied by 6(6 azimuth angles, 24 sampling points). As shown in fig. 3, a network structure diagram of the converged network is shown, which includes a convolutional layer and a pooling layer; the entire converged network can be divided into three parts: the system comprises a multi-level feature extraction part, a multi-scale feature fusion part and a classifier. The last layer is a classifier of a logistic regression, and the output is a label of whether the input sample is real data, namely 0 (input is used for generating the seismic image) and 1 (input is used for generating the real seismic image).

The multi-level feature extraction part of the fusion network extracts multi-level features by using a total of four convolutional layers of Conv1, Conv2, Conv3 and Conv4, wherein each convolutional layer is followed by a LeakyReLU function and a batch normalization layer, and the specification of each convolutional layer is shown in Table 1. The feature size of the image output by each layer is reduced by half, the number of channels is doubled, and the image is abstracted from the low-layer features to the high-layer features step by step.

TABLE 1 multistage feature extraction partial architecture parameters

Type	Kernel	Channels	Stride	Outputs
					input	-	-	-	24x6
Conv1	4x 3	1	2x 2	24x 6x 1
					Conv2	4x 2	16	2x 1	12x 3x 16
Conv3	2x 3	32	3x 3	6x 3x 32
					Conv4	2x 1	64	1x 1	2x 2x 64

The dimensions of the elements in the Output column in table 1 are: trace number time channels

The multistage feature extraction part utilizes the concept of CNN, the output of each convolution layer is the features of different layers of the input seismic image, the deeper the convolution layer is, the smaller the scale of the feature map is, the larger the receptive field is, the more the features are concentrated on the trend of the whole seismic image and a larger address structure, but the more serious the loss of detail information is. Inspired by the Inside-out Net and the characteristic cascade in the field of image target detection, the invention adds a multi-scale characteristic fusion part for a fusion network on the basis of multi-level characteristic extraction, and the part mainly has the following characteristics:

(1) a multi-level feature. Characteristic maps Fea2 and Fea4 were extracted from the convolutional layers Conv2 and Conv4, respectively. The Fea2 is a low-level feature with small scale and mainly describes detail information in the seismic image; fea4 is a high-level feature with a relatively large scale that tends to characterize the overall structural trend of the seismic image. The new characteristic diagram formed by fusing the two images can describe the whole seismic image from multiple layers, so that the representation is more sufficient.

(2) And (5) scaling the size. The two feature maps are not of the same size, and maximum pooling is used for Fea2 so that the pooled feature size is the same as Fea 4.

(3) And normalizing the amplitude. In the convolutional network structure, the amplitude of the feature map is greatly different in convolutional layers of different levels, and if the values are directly combined, the network training process becomes unstable, so that normalization operation needs to be added. The network uses L2 norm normalization.

(4) And fusing the feature maps. And combining the processed Fea2 and Fea4 according to the channel direction to form a new fusion characteristic map Feaf.

(5) And converting the dimension of the fusion features. The fused feature map Feaf is converted into a one-dimensional fused feature vector by two convolutional layers Conv5 and Conv6, and finally input into a classifier for discrimination.

2. Generating networks

The generation network provided by the invention is an auxiliary network and is used for assisting the training of the whole multi-scale feature fusion network. The whole network is constructed based on a full convolution network, and aims to learn the probability distribution of a real seismic image so as to simulate and generate a very vivid generated seismic image. The structure of the whole generated network model is shown in fig. 4, the input of the generated network is 50-dimensional gaussian random noise, and the output is a two-dimensional generated seismic image, and the size of the two-dimensional generated seismic image is consistent with that of the real seismic image input into the fusion network.

Unlike convolutional layers used in conventional CNNs, in the middle layer of the generation network, the generation network replaces the convolutional layers with deconvolution layers. In the generated network model, four deconvolution layers of Deconv1, Deconv2, Deconv3 and Deconv4 are used in total to generate seismic images layer by layer, and the specifications of the respective deconvolution layers are shown in table 2.

Table 2 generating network architecture parameters

Type	Kernel	Channels	Stride	Outputs
					input	-	-	-	50x 1
Deconv1	2x 1	64	1x 1	2x 2x 64
					Deconv2	2x 3	32	3x 3	6x 3x 32
Deconv3	4x 2	16	2x 1	12x 3x 16
					Deconv4	4x 3	1	2x 2	24x 6x 1

The dimensions of the elements in the Output column in table 2 are: trace number time channels

Each deconvolution layer is followed by a modified linear unit (including the activation function ReLU function and the batch normalization layer) in addition to the output layer. To normalize the output to the [0,1] range, the output layer consists of one convolutional layer and one Tanh function. As an "inverse" transformation process of conventional convolution operations, deconvolution is also referred to as micro-step convolution. Typically, the deconvolution transform is also constrained by the "forward" convolution transform parameters, i.e., Padding (Padding) and step size (Stride). The deconvolution layers in the generation network can be divided into two types according to the difference of the filling parameters (valid and same), and the following descriptions are respectively given by comparing the convolution layers with the same parameters:

(1) valid padding, step size not 1

Since the dimension of the input random noise needs to be increased first, the padding mode of the Deconv1 layer of the generation network is padding. In this case, the relationship between convolution and deconvolution is shown in fig. 5.

For convolution operations, assume the input as x ∈ R^4×4Convolution kernel size is 3 × 3, Stride is 2; padding selects valid, i.e., there is Padding. As schematically illustrated by the convolution operation in fig. 5. Then the output can be calculated according to equation (1):

y＝conv(x,ω,'valid')∈R^2×2 (1)

the corresponding deconvolution operation is shown in FIG. 5, with the input being p ∈ R^2×2The output is q ∈ R^4×4。

(2) same filling with step length not 1

The remaining deconvolution layers Deconv2, Deconv3 and Deconv4 of the generating network all use this parameter. In this case, the relationship between convolution and deconvolution is shown in fig. 6.

For convolution operations, the input is x ∈ R^4×4Setting Kernel _ size of the convolution Kernel (Kernel) to 3, that is, its size to 3 × 3, and Stride to 2; padding is same, i.e. no Padding. The output size can then be calculated by the following formula:

the input is y ═ conv (x, ω, 'same') ∈ R^2×2. As "The process of inverse convolution, which can explain how to fit y ∈ R^2×2Transformation to x ∈ R^4×4First, new Stride and Padding need to be calculated using the following equations:

can obtain new output x ∈ R^4×4。

Selecting a pre-stack seismic signal of a certain work area with a sampling rate of 2ms, wherein the waveform of the seismic signal in one trace set is shown as a figure 7(a), and inputting the seismic signal into a trained generation network, and then the figure 7(b) is the generated seismic signal waveform.

S2 multi-scale feature fusion network training algorithm

The framework of the multi-scale feature fusion network provided by the invention is introduced, on one hand, the feature extraction capability of the convolution network is very strong, and a series of features from a low layer to a high layer of the prestack seismic signal can be extracted, so that the prestack seismic reflection mode is characterized in a multi-scale manner; on the other hand, the training mode of generating antagonism among the sub-networks enables the networks to show good convergence when the marking data is scarce, and is suitable for unsupervised feature learning. Assuming input of a multi-scale feature fusion network as a pre-stack seismic imageWhere N is the total number of CDPs, N₁Is the number of sampling points in the time direction, N₂Is the gather number of a prestack CDP (CommonDepthPoint). The functional form of the generation network is represented by G (z), wherein z represents the random noise vector of the input, the effect of the generation network is to synthesize the generated seismic image by learning the distribution of the real seismic image, which means that the generation network can make the output through trainingAs similar as possible to input X. Similarly, a fusion network in functional form is denoted by f (X), where X denotes the input image, and the goal of the fusion network is to enable automatic learningTo complete characterization of X_nThe fusion characteristics of (1). In the converged network, the input prestack seismic image X is first processed_nPerforming convolution by using a series of convolution kernels W, wherein the size of W is c multiplied by l multiplied by k, wherein l represents the size of the convolution kernels, k represents the number of the convolution kernels, and c is the number of channels of an input image (the pre-stack seismic image is a single-channel image, so the value of c is always 1), which is equivalent to that W is matched with the X by using l multiplied by l convolution kernels to check the X_nK convolutions are calculated. Functionally for the input image X_nThe k-th feature image thereof can be expressed as:

Φ(X_n)＝σ(X_n*W^(k)+B_k) (4)

where σ (-) denotes the activation function, B_kRepresents the bias of the k-th feature image, represents a two-dimensional convolution operation. Next, the obtained low-level feature Φ (X)_i) And high layer characteristic phi (X)_j) And performing fusion, wherein i is less than j, and obtaining a fusion characteristic diagram capable of representing the input seismic image in a multi-scale mode. And finally, classifying the features by adopting a logistic regression operation to obtain the probability F (X) that the input seismic image is a real seismic image. The objective function of the multi-scale feature fusion network provided by the invention is obtained by improving the objective function of the GAN:

s3 multi-scale feature fusion network training algorithm process

Inputting: pre-stack seismic image gatherX_iA two-dimensional matrix is formed; the maximum number of training times M.

1: randomly extracting m images { x ] from training image set S⁽¹⁾,...,x^(m)Forming and normalizing a batch;

2: extracting m 50-dimensional Gaussian noise vectors { z⁽¹⁾,...,z^(m)Inputting a generation network to generate m generated seismic images;

3: keeping the generated network unchanged, randomly inputting the generated seismic image and the real seismic image into the fusion network, and updating parameters of the fusion network by adopting a random gradient descent method (SGD)

4: re-extracting m 50-dimensional Gaussian random noise x⁽¹⁾,...,x^(m)Inputting the images into a generation network to generate m new generated seismic images;

5: keeping the parameters of the converged network unchanged, and updating the parameters of the generated network by adopting a random gradient descent (SGD) method

6: the iteration times iter is iter +1, and the training is stopped until the network reaches the maximum iteration times M;

and (3) outputting: and generating a pre-stack seismic signal image.

Most practical work areas often lack logging or related geological data when providing data, so that the acquired seismic signals lack tag data, and a divided seismic phase diagram is usually submitted to a geological expert for identification after an experiment is finished, so that the effectiveness of the identification method is judged. In order to quantitatively explain the effect of the two proposed seismic reflection pattern recognition methods, the invention selects the discrimination index extracted from the angle of the algorithm to measure the recognition effect of the seismic reflection pattern, and the definition of the index is as follows:

degree of distinction

Wherein x is an input vector and n₁Denotes the center vector of the cluster in which x is located, n₂Represents the central vector of the cluster that is closest to x times, | | represents L₂And (4) norm. The discrimination metric depends on the relative distance in vector space from the input vector x to the two nearest cluster centers, which measures the identified reflection pattern against other classesDegree of discrimination of the reflection pattern. The higher the discrimination is, the more obvious the difference between the identified reflection pattern and the reflection patterns of other categories is, and thus the classification result is more reliable. Conversely, a lower degree of discrimination indicates a more uncertain result of seismic reflection pattern recognition.

Calculating the statistical distribution of the discrimination:

wherein n represents the total label number of the classification result, n_dAnd (4) indicating the number of labels with division values d in the classification result. To quantitatively verify the feature extraction capability of the method of the present invention, the following experimental data are combined for illustration:

(1) extracting a section of pre-stack seismic signals, and extracting fusion characteristics by using a characteristic fusion network;

(2) processing the same signal by using a traditional self-encoder algorithm to obtain corresponding characteristics;

(3) the features obtained by the two methods are homopolymerized into 8 types by using an FSOM algorithm, the discrimination index of each signal is calculated, and a discrimination statistical chart is drawn.

The statistical distribution of the discrimination of the multi-scale feature fusion network and the self-encoder is shown in fig. 8. Compared with the traditional self-encoder, the multi-scale feature fusion network has a larger percentage at high discrimination and a smaller percentage at low discrimination. The higher the discrimination degree is, the more reliable the classification result is, so the reliability of the final classification of the multi-scale feature fusion network provided by the invention is higher than that of a self-encoder, thereby verifying that the fusion features extracted by the multi-scale feature fusion network have stronger characterization capability on pre-stack seismic signals, and achieving higher identification capability of the pre-stack seismic facies reflection mode.

The feasibility of the algorithm provided by the invention is verified by selecting a piece of prestack physical model data, the research target of the physical model is a Jurassic dense reservoir, and a data prototype is shown in figure 9 and mainly simulates the structure of a karst cave and a fault. The data acquisition process comprises the following steps: firstly, a 10000:1 space model is made, a 2:1 speed scale factor is adopted, then the model is placed in a water tank to design a seismic signal acquisition system, and finally data are acquired by utilizing ultrasonic waves (the sampling rate is set to be 1 ms). The selected data contained six azimuths, 580 crosslines and 690 inlines, for a total of 400200 prestack seismic traces, i.e., 400200 sample images in the dataset. Each pass takes 24 samples, so the sample image dimension is 24 × 6.

In order to verify the improvement of the method on the seismic facies identification effect, the method uses the latest pre-stack seismic signal classification method based on a Deep Convolutional auto-coders (DCAE) for experimental comparison. Firstly, the DCAE is used for extracting the characteristics of the pre-stack seismic signals, and then the pre-stack seismic signals are divided into 10 types by using the FSOM, and as shown in FIG. 10, the pre-stack seismic phase diagram obtained by the DCAE method is shown. Fusion features of the pre-stack seismic signals are extracted by using a multi-scale feature fusion network, and are divided into 10 classes by using the FSOM, and the obtained pre-stack seismic phase diagram is shown in FIG. 11. Because the work area is a mature exploration work area, although a data prototype (figure 9) made by a geological expert in the later stage can be used as an evaluation standard figure 10 of the identification result of the pre-stack seismic reflection pattern, the faults and the karst caves with larger sizes can be well identified, the fineness of the karst caves with smaller sizes in an upper circle and a lower circle in the figure 10 is not enough, especially the complicated small faults in a middle circle are not clearly described, the higher exploration requirement is difficult to meet, and the practical application is not facilitated. On the contrary, fig. 11 not only well depicts the development trend of the fault, but also clearly identifies the karst cave and the fault with different scales, which is closer to fig. 9. By contrast, the method provided by the invention can more clearly identify geological structures such as karst caves, faults and the like, has a better identification effect on the pre-stack seismic reflection mode, and shows that the characteristics extracted by the multi-scale characteristic fusion network can better represent the pre-stack seismic reflection mode.

Actual work area data

(1) LZB work area

The LZB work area is positioned in the Sichuan basin, and a plurality of problems need to be overcome in the research of reservoir prediction, reservoir formation mechanism, exploration direction and the like of compact oil in the work area, so that the research on the work area by using the seismic reflection mode identification technology has very high practical value. The pre-stack seismic signals of the work area contain six azimuth angle data, the fracture and karst cave characteristics in the target layer are obvious, and fig. 12 shows the amplitude attribute of the work area.

The experimental work area was selected to be 950 inlines and 550 crosslines, for a total of approximately 522500 samples. The data sampling rate is 2ms, and 24 sampling points are selected for each seismic channel. The method is the same as the physical model data comparison experiment, the DCAE and the method provided by the invention are continuously adopted for comparison, the characteristics of the pre-stack seismic signals are respectively extracted by using the two methods, then the FSOM is used for classifying into 10 types, and finally the accuracy of the result is verified by comparing the obtained pre-stack seismic phase diagram with the diagram 12.

Fig. 13 is a pre-stack seismic phase diagram of the LZB work area generated by the DCAE method, in which a large fault trend can be seen, but the small fault structures in the two red circles are very blurred, and the karst cave in the middle red circle cannot be predicted, which is mainly caused by the small scale of the karst cave, so that the seismic data are poorly differentiated.

The pre-stack seismic phase diagram of the LZB work area generated by the method is shown in FIG. 14, and compared with FIG. 12, the overall geological structure trend of FIG. 13 is accurate, but is more clearly described on the small-scale fault and karst cave structure than that of FIG. 13, the improvement of the identification effect of the pre-stack seismic reflection mode of the method is verified, and meanwhile, the characteristic that the reflection mode of the multi-scale feature fusion network can be extracted and disclosed from the pre-stack seismic signal is also shown.

(2) Brocade work area

The amplitude properties of the brocade work area are shown in fig. 15, which includes a distinct fault distribution area. The experimental work area data sampling rate was chosen to be 1ms, containing 1051 inlines and 551 crosslines, and containing six azimuth data. Data for 24 points on each prestack seismic trace is selected along the horizon to form a sample set of input data. And (3) continuing to adopt an experimental method for comparing DCAE, and dividing the seismic signals into 6 classes by using FSOM after extracting the characteristics of the seismic signals before the superposition because the underground geological structure of the work area is relatively simple. Fig. 16 is a pre-stack seismic phase diagram of a brocade work area generated by a DCAE-based method, and fig. 17 is a pre-stack seismic phase diagram generated by the method of the invention. Compared with the actual geological structure in fig. 15, although the overall fault direction can be accurately described in fig. 16, for some detail information located at the end in the circle in fig. 16, only the description of fig. 17 is clearer and more continuous, so that the method provided by the invention has a better recognition effect on the pre-stack seismic reflection mode. The data of different work areas are used, but the identification effect is still good, so that the feature extraction method provided by the invention is very universal, and the requirement of identifying the pre-stack seismic signal reflection modes of different work areas in practical application can be met.

In the invention, inline is a main survey line seismic interpretation section, and crossline is a junctor seismic interpretation section.

It will be appreciated by those of ordinary skill in the art that the embodiments described herein are intended to assist the reader in understanding the principles of the invention and are to be construed as being without limitation to such specifically recited embodiments and examples. Various modifications and alterations to this invention will become apparent to those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A pre-stack seismic reflection pattern recognition method based on multi-scale feature fusion is characterized by comprising the following steps:

s2, constructing a multi-scale feature fusion network;

2. The method for identifying the pre-stack seismic reflection pattern based on the multi-scale feature fusion as claimed in claim 1, wherein the step S1 specifically comprises: firstly, denoising an obtained pre-stack seismic signal by adopting a structure-oriented filtering algorithm; then, intercepting target data segments above and below the time dimension by referring to the position of the horizon; and finally, connecting the data sections with the same position but different azimuth angles to form a pre-stack seismic image set.

3. The method for identifying the pre-stack seismic reflection pattern based on the multi-scale feature fusion as claimed in claim 1, wherein the multi-scale feature fusion network of step S2 comprises: fusing and generating a network; the fusion network is a network structure which extracts the low-level features and the high-level features of the pre-stack seismic signals and performs multi-scale fusion; the resulting network is a full convolutional network.

4. The method for identifying the pre-stack seismic reflection pattern based on the multi-scale feature fusion as claimed in claim 3, wherein the fusion network comprises three parts: the system comprises a multi-level feature extraction part, a multi-scale feature fusion part and a classifier part; the multilevel feature extraction part adopts four convolutional layers of Conv1, Conv2, Conv3 and Conv4 to extract multilevel features; the multi-scale feature fusion part extracts feature maps Fea2 and Fea4 from the convolutional layers Conv2 and Conv4 respectively and fuses the extracted feature maps to form new feature maps; and the classifier part judges the fused new feature map.

5. The method for identifying the pre-stack seismic reflection pattern based on the multi-scale feature fusion as claimed in claim 4, wherein the multi-scale fusion part is realized by the following steps:

6. The method as claimed in claim 5, wherein the generation network adopts four deconvolution layers of Deconv1, Deconv2, Deconv3 and Deconv4 to generate the seismic image layer by layer.

7. The method for identifying the pre-stack seismic reflection pattern based on the multi-scale feature fusion as claimed in claim 6, wherein the step S3 specifically comprises: