A method for intelligent identification of seismic phases based on multidimensional labeling model
Through the combination of multi-dimensional label model and neural network, the problems of high professional threshold and label complexity of artificial interpretation in intelligent seismic phase recognition are solved, and the fine division and accurate identification of seismic phases are realized, which reduces the complexity of supervised learning and forms a Simpler and clearer seismic phase label.
Patent Information
- Application Number
- CN202410180571.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-02-18
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-02-18
AI Technical Summary
The existing intelligent seismic phase identification methods have high professional thresholds and multi-solvency problems for manual interpretation, and the lack of effective label guidance in supervision and learning, resulting in insufficient accuracy and reliability of seismic phase division.
Using the intelligent seismic phase recognition method based on multi-dimensional label model, a single-dimensional seismic phase label of multiple dimensions is created, and a neural network model is trained to predict the reflective appearance, internal structure, amplitude, continuity and frequency characteristics of seismic data. , and finally combine the prediction results to form the final seismic phase prediction results.
Reduces the professional requirements for humans in seismic phase division, improves the accuracy and reliability of seismic phase recognition, reduces the complexity of labels in supervised learning, and forms a simpler and clearer seismic phase label, which can better The attribute characteristics of traditional seismic phases are retained.
Smart Images

Figure CN118050785B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a seismic phase intelligent identification method based on a multi-dimensional label model, and belongs to the technical field of seismic phase intelligent identification. Background Art
[0002] Seismic facies is the manifestation of underground sedimentary facies in seismic data, which can indirectly reflect the characteristics of underground sedimentation. Its accurate identification is of great significance for sedimentary facies research and oil and gas exploration and development. Seismic facies analysis technology is an important method for reservoir prediction and can be used to describe the distribution law of favorable sedimentary facies belts.
[0003] Traditional seismic phase identification and division is done manually through seismic phase analysis, which requires a lot of time and requires the interpreter to have sufficient experience and combine certain analysis methods to complete. The professional threshold is high, and there is a certain degree of multi-solution in manual interpretation, and the interpretation results are somewhat subjective. With the improvement of exploration accuracy and the application of high-density and wide-azimuth acquisition technology, seismic data has entered the era of big data, and how to extract geological feature information from massive seismic data has become a research hotspot.
[0004] Automatic identification of seismic phases can be regarded as a fitting problem, through which a nonlinear relationship between seismic records and sedimentary phase types is established. The essence of machine learning is to learn this nonlinear relationship from data driven by a large amount of data. Seismic records have a large amount of data and rich information, so machine learning methods can be used to obtain this nonlinear relationship, thereby identifying and dividing seismic phases. In recent years, with the rapid development of artificial intelligence and the vigorous promotion of oil and gas artificial intelligence, domestic and foreign scholars have proposed a variety of methods for intelligent identification of seismic phases. The methods of intelligent identification of seismic phases are summarized and summarized into three categories: unsupervised learning, supervised learning, and semi-supervised learning.
[0005] Unsupervised seismic phase division is to train unlabeled data samples and reveal the characteristic information inside the data through learning. Unsupervised learning mainly includes methods such as clustering and self-organizing map networks. K-means clustering is a clustering method commonly used for data set grouping, and has also been widely used in seismic phase division. This method divides the data set into K clusters (groups) and assigns each data point to the nearest cluster to achieve the purpose of aggregating data points with high similarity. Coléou (2023) applied K-means clustering to seismic phase analysis. The idea of self-organizing mapping is to project data into a low-dimensional space to obtain a more intuitive understanding, but it can only judge the number of categories based on experience and select the optimal seismic attributes to characterize the geological characteristics in seismic data. Zhang Gong et al. (2015) realized the automatic identification of seismic phases with preferred attributes based on K-means clustering based on self-organizing mapping and particle swarm optimization. Wang Qiannan (2024) proposed a multi-feature fusion autoencoder to extract implicit features of seismic data, so that the extracted implicit features can statistically classify seismic phases.
[0006] Unsupervised seismic facies classification lacks the guidance of prior knowledge and cannot be judged for its correctness. Seismic facies derived through unsupervised learning are often difficult to associate with existing geological units or structures, whereas supervised learning enables us to decide the number and type of seismic facies to include.
[0007] Supervised learning is to let the model learn the invisible features inside the data under the condition of defining labels, so as to predict similar data. Supervised learning algorithms include neural networks, support vector machines and random forests, among which neural networks are the most widely used. West (2002) combined image texture analysis with neural network classification to quantitatively map seismic phases in three-dimensional data. Alaudah et al. (2019) developed two baseline models for phase classification based on the deconvolution network architecture and applied machine learning methods to phase classification. At the same time, selecting seismic attributes that are sensitive to reservoir characteristics from a large number of seismic attributes is an important part of seismic phase analysis (Jie Tao et al., 2016). Qi et al. (2020) performed seismic phase classification by selecting seismic phase combinations. Lubo-Robles et al. (2018) applied an unsupervised multi-attribute technique to reduce dimensionality and extract the most valuable information from multiple spectral amplitude components to perform unsupervised seismic phase classification. For supervised learning, the performance of seismic phase classification can be improved by selecting input attributes related to the output category and avoiding using redundant attributes that are similar to each other to construct the best attribute subset.
[0008] The current intelligent seismic phase identification method under supervised learning is to train with labeled seismic data sets so that the model can learn the characteristics of seismic phases and make predictions on new unlabeled data. The labeled seismic phases, that is, seismic phase labels, are generally divided from multiple perspectives through human interpretation. Each seismic phase divided in this way contains multiple complex seismic information. At present, scholars do not pay much attention to seismic phase labels. They mostly use traditional geological division methods and focus on how to improve the accuracy of network model learning. Summary of the invention
[0009] In order to overcome the defects in the prior art, the present invention aims to provide a method for intelligent identification of seismic phases based on a multi-dimensional label model.
[0010] The technical solution provided by the present invention to solve the above technical problems is: a method for intelligent identification of seismic phases based on a multidimensional label model, comprising the following steps:
[0011] S1, making single-dimensional seismic phase labels of multiple dimensions corresponding to the original three-dimensional seismic data;
[0012] S2, establishing a data set based on the original three-dimensional seismic data and the single-dimensional seismic phase labels corresponding to multiple dimensions, and dividing the data set into a training set and a test set;
[0013] S3, taking the original seismic data and the single-dimensional seismic phase label as input, training the neural network model through the training set, and then testing the trained neural network model through the test set to obtain a trained neural network model;
[0014] S4, predicting the single-dimensional seismic phase prediction results corresponding to multiple dimensions of the seismic data to be identified through the trained neural network model;
[0015] S5. Combine the prediction results of multiple dimensions to form the final earthquake phase prediction result.
[0016] A further technical solution is that the multiple dimensions in step S1 include reflection shape dimension, internal structure dimension, amplitude dimension, continuity dimension, and frequency dimension.
[0017] A further technical solution is that the reflection shape dimension is divided into a mat shape and a hill shape according to the reflection shape.
[0018] A further technical solution is that the internal structure dimensions of the step are divided into parallel, oblique, S-shaped, messy, and top convex according to the internal structure.
[0019] A further technical solution is that the amplitude dimension is divided into strong amplitude, medium amplitude and weak amplitude according to the amplitude.
[0020] A further technical solution is that the continuity dimension is divided into high continuity, medium continuity, low continuity and discontinuity according to continuity.
[0021] A further technical solution is that the frequency dimension is divided into high frequency, medium frequency and low frequency according to the frequency.
[0022] A further technical solution is that the neural network model is a model that uses machine learning to perform intelligent identification of seismic phases among the VGG16 network model, the PSPNet network model, the Unet network model, and the DeepLabV3+ network model.
[0023] The present invention has the following beneficial effects: the present invention mainly aims at the problem of intelligent identification of seismic phases, and proposes a method for finely making seismic phase labels based on multiple dimensions of reflection shape, internal structure, amplitude, continuity and frequency, as well as a method for combining the seismic phase prediction results of each dimension after cyclic learning, training and testing of a single label in a VGG16 network. By dividing the labels in multiple dimensions, the production of traditional complex seismic phase labels becomes simpler. At the same time, the final combined seismic phase prediction results also retain the characteristics of comprehensive division of traditional seismic phases based on attributes such as reflection shape, internal structure, amplitude, continuity and frequency, which reduces the professional requirements for seismic phase division in supervised learning. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 is a flow chart of the present invention;
[0025] Figure 2 is the profile of the original seismic data;
[0026] Figure 3 It is the seismic phase label map of reflection shape dimension;
[0027] Figure 4 It is the seismic phase label map of the continuity dimension;
[0028] Figure 5 It is the seismic phase label map of internal structure dimension;
[0029] Figure 6 It is the frequency dimension seismic phase label map;
[0030] Figure 7 It is the seismic phase label map in amplitude dimension;
[0031] Figure 8 This is the VGG16 network model diagram;
[0032] Fig. 9 This is the result diagram of the prediction of the reflection shape dimension;
[0033] Fig.10 Prediction result graph for internal structural dimension;
[0034] Fig.11 This is the prediction result graph for the amplitude dimension;
[0035] Fig.12 Prediction result graph for continuous dimension;
[0036] Fig.13 Prediction result graph for frequency dimension;
[0037] Fig.14 This is the prediction result diagram of combined seismic phase labels. DETAILED DESCRIPTION
[0038] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0039] like Figure 1 As shown, a method for intelligently identifying seismic phases based on a multidimensional label model of the present invention comprises the following steps:
[0040] S1, making single-dimensional seismic phase labels of multiple dimensions corresponding to the original three-dimensional seismic data;
[0041] The reflection shapes are divided into sheet-like, hillock-like, and wedge-like; the internal structures are divided into parallel, oblique, S-shaped, messy, and top convex; the amplitude is divided into strong amplitude, medium amplitude, and weak amplitude; the continuity is divided into high continuous, medium continuous, low continuous, and discontinuous; the frequency is divided into high frequency, medium frequency, and low frequency, as shown in the seismic phase division parameter table.
[0042]
[0043] The following is the training data used in this invention. By analyzing the original seismic data in a single dimension, the reflection shape P is divided into two seismic phases: sheet-like and hillock-like. Each seismic phase is represented by a numerical value, 0 represents the sheet-like phase, and 1 represents the hillock-like seismic phase, that is, P[0,1]. Similarly, the internal structure dimension is divided into parallel, oblique, chaotic, and top convex, represented by S[0,1,2,3]; the amplitude dimension is divided into strong amplitude, medium amplitude, and weak amplitude, represented by A[0,1,2]; the continuity dimension is divided into four phase belts: high continuity, medium continuity, low continuity, and discontinuous, represented by C[0,1,2,3]; the frequency dimension is divided into high frequency, medium frequency, and low frequency, represented by F[0,1,2].
[0044] S2, establishing a data set based on the original three-dimensional seismic data and the single-dimensional seismic phase labels corresponding to multiple dimensions, and dividing the data set into a training set and a test set;
[0045] S3, using the original three-dimensional seismic data and the single-dimensional seismic phase label as input, training the neural network model through the training set, and then testing the trained neural network model through the test set to obtain a trained neural network model;
[0046] The neural network model is one of a VGG16 network model, a PSPNet network model, a Unet network model, and a DeepLabV3+ network model;
[0047] Fully convolutional network with VGG16 structure:
[0048] This network model is based on the idea of a fully convolutional network and is used for image semantic segmentation tasks. The main idea of a fully convolutional network is to replace the fully connected layers of a traditional convolutional neural network (CNN) with convolutional layers, so that the network can accept input images of any size and output predictions of the same size. In this particular model, VGG16 is used as the basic feature extractor, which is then gradually restored to the size of the original input image through deconvolution and upsampling operations. The model uses a pre-trained VGG16 model to initialize the convolutional layer parameters of the encoder part, which helps the model converge faster and achieves good results when trained on relatively small datasets.
[0049] The principle of the model is that the encoder extracts high-level features of the image through a series of convolution operations, gradually reducing the size of the feature map. The decoder gradually restores the size of the feature map through deconvolution and upsampling operations, and introduces skip connections to retain more spatial details. The output layer generates the prediction result of semantic segmentation through a convolution layer.
[0050] Encoder part:
[0051] The first five convolutional blocks of VGG16 are used to extract the features of the input image. These convolutional blocks gradually reduce the size of the feature map through convolution, batch normalization, ReLU activation function, and pooling operations.
[0052] Given input features Figure X And the convolution kernel (filter) W, the convolution operation can be expressed by the following mathematical formula:
[0053] Y[i,j]=∑ m ∑ n X[i+m,j+n]·W[m,n]+b
[0054] Among them, Y[i,j] is the value of position [i,j] in the feature map of the convolution output, X[i+m,j+n] is the value of position [i+m,j+n] in the input feature map, W[m,n] is the weight of position [m,n] in the convolution kernel, and b is the bias. For the operation of the entire convolution layer, the convolution operation can be performed on the input by sliding the convolution kernel to obtain the output feature map.
[0055] Decoder part:
[0056] Use the inverse operation of deconvolution and maximum pooling to upsample and gradually restore the size of the feature map. Each layer of the decoder has a corresponding index of the pooling layer, which is used for the inverse operation of the maximum pooling to restore the position information during pooling when upsampling. Finally, a convolution layer is used to output the result of semantic segmentation. The mathematical formula for deconvolution is as follows:
[0057] Y[i,j]=∑ m ∑ n X[im,jn]·W[m,n]+b
[0058] Among them, Y[i,j] is the value of position [i,j] in the feature map of the deconvolution output, X[im,jn] is the value of position [i+m,j+n] in the input feature map, W[m,n] is the weight of position [m,n] in the deconvolution kernel, and b is the bias.
[0059] The inverse operation of max pooling requires recording the index of the maximum value during the pooling operation and then putting these maximum values back to the corresponding positions.
[0060] Upsampling is an operation that enlarges the input feature map. The mathematical formula can be simply expressed as:
[0061]
[0062] Among them, X1 and X2 are two connected feature maps.
[0063] Skip connections:
[0064] A skip connection is introduced into the model, which adds the feature map of the corresponding layer of the encoder to the feature map of the corresponding layer of the decoder. This design can help retain more spatial details and slow down the information loss during feature map upsampling. This is especially important for tasks such as semantic segmentation, which require more refined spatial information.
[0065] Y[i,j]=X low [i,j]+X high [i,j]
[0066] Where X low is the feature map of the corresponding layer of the encoder, Xhigh is the feature map of the corresponding layer of the decoder.
[0067] Due to the different ways of dividing seismic phases, the weight of each seismic phase in different dimensions is different. In order to solve the problem of data imbalance, the calculation of category weights is added to the original neural network model.
[0068] The calculation of class weights is to deal with the problem of class imbalance. In the semantic segmentation task, the number of samples in different categories may vary greatly. In order to balance this difference, by calculating the weight of each category, the samples can be given greater weight during training, so as to better learn the minority category. By calculating the number of samples of each category in the training set, and then using class weights to balance the importance of each category. This is because in an unbalanced dataset, some categories may have fewer samples, causing the model to be more biased towards predicting categories with more samples. The introduction of class weights helps to balance the contribution of the loss function to each category, thereby improving the performance of the model on minority categories.
[0069]
[0070] S4, predicting the single-dimensional seismic phase prediction results corresponding to multiple dimensions of the seismic data to be identified through the trained neural network model;
[0071] S5, combining the prediction results of multiple dimensions to form the final earthquake phase prediction result;
[0072] The seismic phase labels of multiple dimensions are combined into new labels in the form of permutations and combinations. In the original seismic phase labels, the information is clear and single, and machine learning can better process the characteristic data corresponding to the labels. After the neural network test, the labels of five dimensions, namely reflection shape, internal structure, amplitude, continuity and frequency, are combined again. The combined seismic phase labels are more in line with the geological interpretation, forming a perfect closed loop, which can characterize the seismic phases in more detail.
[0073] The specific principles of earthquake phase combination are as follows:
[0074] P represents the reflection shape dimension: P0 represents mat shape, P1 represents hill shape, P i =[P0,P1], i=[0,1], a=2;
[0075] S represents the internal structure dimension: S0 represents parallel, S1 represents oblique, S2 represents disorderly, S3 represents convex, S j =[S0, S1, S2, S3], j=[0,1,2,3], b=4;
[0076] A represents the amplitude dimension: A0 represents strong amplitude, A1 represents medium amplitude, A2 represents weak amplitude, and A k =[A0,A1,A2], k=[0,1,2], c=3;
[0077] C represents the continuity dimension: C0 represents high continuity, C1 represents medium continuity, C2 represents low continuity, C3 represents discontinuity, and C n =[C0,C1,C2,C3], n=[0,1,2,3], d=4;
[0078] F represents the frequency dimension: F0 represents high frequency, F1 represents medium frequency, F2 represents low frequency, and F m =[F0,F1,F2], m=[0,1,2], e=3.
[0079]
[0080] P i S j A k C n F m =i×b×c×d×e+j×c×d×e+k×d×e+n×e+m
[0081] Among them, ijknm represent the seismic phases of reflection shape dimension, internal structure dimension, amplitude dimension, continuity dimension and frequency dimension respectively. It represents the combined seismic phase, a, b, c, d, and e represent the number of seismic phases in the reflection shape dimension, internal structure dimension, amplitude dimension, continuity dimension, and frequency dimension, respectively. The value of i×b×c×d×e+j×c×d×e+k×d×e+n×e+m is the numerical representation of the combined seismic phase;
[0082] For example, the sheet-like parallel medium-amplitude high-continuous medium-frequency seismic phase is expressed as P0S0A1C0F1=13, i.e. L 13 It indicates a sheet-like parallel medium-amplitude high-continuous medium-frequency seismic phase.
[0083] The above description is not intended to impose any form of limitation on the present invention. Although the present invention has been disclosed through the above embodiments, it is not intended to limit the present invention. Any technician familiar with the profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of the technical solution of the present invention.
Claims
1. A method for intelligent identification of seismic phases based on a multidimensional label model, characterized in that: The following steps are involved: S1, making single-dimensional seismic phase labels of multiple dimensions corresponding to the original three-dimensional seismic data; Multiple dimensions include reflection shape dimension, internal structure dimension, amplitude dimension, continuity dimension, and frequency dimension; By analyzing the original seismic data in a single dimension, the reflection shape P is divided into two seismic phases: sheet-like and hillock-like. Each seismic phase is represented by a numerical value, 0 represents the sheet-like phase, and 1 represents the hillock-like phase, that is, P[0,1]. Similarly, the internal structure dimension is divided into parallel, oblique, chaotic, and convex, represented by S[0,1,2,3]; the amplitude dimension is divided into strong amplitude, medium amplitude, and weak amplitude, represented by A[0,1,2]; the continuity dimension is divided into four phase belts: high continuity, medium continuity, low continuity, and discontinuous, represented by C[0,1,2,3]; the frequency dimension is divided into three types: high frequency, medium frequency, and low frequency, represented by F[0,1,2]. S2, establishing a data set based on the original three-dimensional seismic data and the single-dimensional seismic phase labels corresponding to multiple dimensions, and dividing the data set into a training set and a test set; S3, taking the original seismic data and the single-dimensional seismic phase label as input, training the neural network model through the training set, and then testing the trained neural network model through the test set to obtain a trained neural network model; S4, predicting the single-dimensional seismic phase prediction results corresponding to multiple dimensions of the seismic data to be identified through the trained neural network model; S5, combining the prediction results of multiple dimensions to form the final earthquake phase prediction result; The specific principles of earthquake phase combination are as follows: P represents the reflection shape dimension: P0 represents mat shape, P1 represents hill shape, P i =[P0,P1], i=[0,1], a=2; S represents the internal structure dimension: S0 represents parallel, S1 represents oblique, S2 represents disorderly, S3 represents convex, S j =[S0, S1, S2, S3], j=[0,1,2,3], b=4; A represents the amplitude dimension: A0 represents strong amplitude, A1 represents medium amplitude, A2 represents weak amplitude, and A k =[A0, A1, A2], k=[0,1,2], c=3; C represents the continuity dimension: C0 represents high continuity, C1 represents medium continuity, C2 represents low continuity, C3 represents discontinuity, and C n =[C0, C1, C2, C3], n=[0,1,2,3], d=4; F represents the frequency dimension: F0 represents high frequency, F1 represents medium frequency, F2 represents low frequency, and F m =[F0, F1, F2], m=[0,1,2], e=3; P i S j A k C n F m =i×b×c×d×e+j×c×d×e+k×d×e+n×e+m Among them, ijknm represent the seismic phases of reflection shape dimension, internal structure dimension, amplitude dimension, continuity dimension and frequency dimension respectively. represents the combined seismic phase, a, b, c, d, and e represent the number of seismic phases in the reflection shape dimension, internal structure dimension, amplitude dimension, continuity dimension, and frequency dimension, respectively; The value of i×b×c×d×e+j×c×d×e+k×d×e+n×e+m is the numerical representation of the combined seismic phase.
2. The method for intelligent identification of seismic phases based on a multidimensional label model according to claim 1, characterized in that: The neural network model is a model that uses machine learning to perform intelligent identification of seismic phases among the VGG16 network model, the PSPNet network model, the Unet network model, and the DeepLabV3+ network model.
Citation Information
Patent Citations
Three-dimensional seismic facies identification method and device based on convolutional neural network
CN113805235A
Pre-stack seismic inversion method and device
CN115130529A
Seismic facies intelligent identification and detection method based on TransUNet deep learning network
CN116559944A
Multi-scale seismic facies intelligent identification method based on combination of stacked generalization integration and convolutional neural network
CN117518248A
System and method for seismic facies identification using machine learning
US20180106917A1
