WO2024000709A1 - 一种基于自注意力机制与u型结构结合的地震相自动化识别方法 - Google Patents

一种基于自注意力机制与u型结构结合的地震相自动化识别方法 Download PDF

Info

Publication number
WO2024000709A1
WO2024000709A1 PCT/CN2022/108319 CN2022108319W WO2024000709A1 WO 2024000709 A1 WO2024000709 A1 WO 2024000709A1 CN 2022108319 W CN2022108319 W CN 2022108319W WO 2024000709 A1 WO2024000709 A1 WO 2024000709A1
Authority
WO
WIPO (PCT)
Prior art keywords
seismic
module
self
sub
function
Prior art date
Application number
PCT/CN2022/108319
Other languages
English (en)
French (fr)
Inventor
王治国
陈宇民
杨阳
高照奇
李振
王倩楠
高静怀
Original Assignee
西安交通大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 西安交通大学 filed Critical 西安交通大学
Priority to US18/321,632 priority Critical patent/US11922679B2/en
Publication of WO2024000709A1 publication Critical patent/WO2024000709A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/26—Government or public services
    • G06Q50/265—Personal security, identity or safety
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/22—Image preprocessing by selection of a specific region containing or referencing a pattern; Locating or processing of specific regions to guide the detection or recognition
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/32—Normalisation of the pattern dimensions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks

Definitions

  • the invention belongs to the field of seismic exploration technology, and relates to an automatic seismic phase identification technology based on a combination of self-attention mechanism and U-shaped structure, especially a technology that combines Segformer self-attention segmentation network, UNet network structure and Hypercolumn semantic segmentation technology.
  • Seismic image semantic segmentation method, and the seismic image semantic segmentation method is used for automatic recognition and classification of seismic phases of seismic data.
  • oil and gas exploration mainly uses seismic exploration methods, that is, post-stack seismic data is obtained through artificial seismic wave reflection, and the contained underground structure, lithology, oil and gas properties and other information are analyzed through multi-disciplinary knowledge mining, so as to locate the distribution of underground oil and gas reservoirs.
  • Traditional seismic facies classification schemes are manually interpreted by interpreters or employ some mathematical methods to semi-automatically extract features and segment seismic facies.
  • these traditional methods are highly subjective.
  • the semi-automatic methods are not accurate enough and have poor timeliness. They cannot accurately locate oil and gas reservoirs in complex underground structures and depositional conditions. How to use computer resources to implement an efficient automated seismic phase identification method has become a current problem that needs to be solved.
  • seismic facies classification methods based on deep learning have been proposed by many researchers.
  • This method is to learn a seismic data to seismic facies in an end-to-end manner based on the existing labeled seismic data.
  • the advantage of this method is that it can save a lot of labor costs and directly conduct end-to-end seismic facies classification, and improve the interpretation effect to a certain extent.
  • Jesper (2018) migrated the VGG16 network pretrained by ImageNet to manually annotated seismic data, and used the sliding window method to identify the seismic phase in the center of the sliding window, thereby realizing automatic classification of seismic phases.
  • Zhao used a CNN network with an encoder-decoder structure to classify seismic facies.
  • Di marked 4 Inline sections and used a network similar to UNet to realize automatic identification of seismic phases.
  • the purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art and propose an automatic seismic phase identification method based on the combination of self-attention mechanism and U-shaped structure.
  • This method introduces coding based on the semantic segmentation network of the self-attention mechanism.
  • -Decode the U-shaped structure use the semantic segmentation network as the encoder module, and introduce a block expansion module based on fully connected layer upsampling in the decoder, and use hypercolumn technology to fuse features, thereby obtaining a low computational load and ability to extract Global attention seismic phase segmentation method using multi-scale features.
  • the present invention provides an automatic seismic phase identification method based on the combination of attention mechanism and U-shaped structure.
  • the method includes the following steps:
  • the encoder is constructed using an overlapping block merging module with downsampling function and an efficient self-attention transformer module capable of global modeling representation;
  • the seismic phase recognition model includes a hypercolumn unitary divider
  • Construct a hybrid loss function use the training verification set in the sample training verification data set to iteratively train the seismic phase identification model; input test data to obtain the seismic phase of the test seismic data.
  • Obtaining post-stack seismic data and preprocessing the post-stack seismic data volume to construct a sample training verification data set specifically includes the following steps:
  • each slice block consists of a first sub-block and a second sub-block, and the first sub-block is used as a training
  • the set and the second sub-block are used as the verification set; where N is a positive integer greater than 2, and the number of the first sub-block is not less than the number of the second sub-block;
  • the use of an overlapping block merging module with downsampling function and an efficient self-attention transformer module capable of global modeling to construct a model encoder specifically includes the following steps:
  • LN is the layer normalization function
  • MHSA is the multi-head self-attention calculation function
  • L 1 and L 2 are two fully connected functions
  • cv is the convolution layer function.
  • the use of the block expansion module with linear upsampling function, the efficient self-attention transformer module and the skip connection module capable of fusing high and low layer features to construct a model decoder specifically includes the following steps:
  • the use of the encoder, the decoder and the hypercolumn module to construct a seismic phase identification model specifically includes the following steps:
  • the method of constructing a hybrid loss function and iteratively training the seismic phase identification model using the training verification set in the sample training and verification data set; inputting test data to obtain the seismic phase of the test seismic data specifically includes the following steps:
  • the seismic phase identification model is trained using a hybrid loss of the cross-entropy loss function and Dice Loss.
  • the calculation formula of the loss function is:
  • y is the real seismic phase label of the seismic image
  • p is the prediction mask of the seismic image
  • y i, j represents the one-hot encoding label corresponding to the pixel of the seismic image at i, j, Indicates the class probability that the pixel at position i and j of the seismic image is predicted to be the kth class seismic phase.
  • the method also includes the following steps:
  • a batch stochastic gradient descent algorithm is used to iteratively update and learn the parameters of the seismic phase identification model
  • the batch stochastic gradient descent algorithm is used to iteratively update and learn the parameters of the model.
  • the batch stochastic gradient descent algorithm is used to iteratively update and learn the parameters of the seismic phase identification model, which specifically includes the following steps:
  • the parameters of the seismic phase identification model are updated along the negative direction of the gradient to achieve a continuous decrease of the loss function.
  • the present invention combines the self-attention mechanism with the U-shaped structure, and uses the idea of hyper-column technology segmentation to automatically segment and identify seismic phases. At the same time, it introduces a hybrid loss function, which can make the model pay more attention to the continuity of segmentation.
  • the seismic phase identification network model proposed by the present invention not only enables the computer to have a lower calculation amount, but also can obtain a higher seismic phase identification accuracy. Furthermore, seismic phase identification results with high accuracy can be used to more effectively predict the location and structure of underground sedimentary environments, providing favorable technical support for oil and gas exploration.
  • Figure 1 is a structural diagram of an encoder according to an embodiment of the present invention.
  • Figure 2 is a structural diagram of a block expansion module according to an embodiment of the present invention.
  • Figure 3 is a HUSeg network structure diagram according to an embodiment of the present invention.
  • Figure 4 is a comparison diagram of different semantic segmentation methods used for seismic phase identification according to an embodiment of the present invention, wherein Figure 4(a) is a seismic profile image; Figure 4(b) is a seismic phase label map corresponding to the seismic profile; Figure 4(c) is a SegNet seismic phase identification result diagram; Figure 4(d) is a SegNet seismic phase identification result diagram; Figure 4(e) is a SegNet seismic phase identification result diagram; Figure 4(f) is a seismic phase proposed by the present invention. Phase identification result diagram;
  • Figure 5 is a post-stack three-dimensional seismic data volume diagram of the Bohai Bay Basin according to an embodiment of the present invention
  • Figure 6 is a diagram of seismic data layer slicing and identification results according to an embodiment of the present invention.
  • Figure 6(a) is a seismic data layer slicing diagram;
  • Figure 6(b) is a layer of seismic phase identification proposed by the present invention.
  • the seismic facies interpretation of post-stack seismic data can be regarded as semantic segmentation of seismic images, which divides seismic images into different areas, thus reflecting different underground depositional environments, and can be used to locate oil and gas reservoirs.
  • image segmentation methods to divide seismic facies. These methods do not model the correlation between seismic image pixels, and have poor effects on describing the boundary details of some seismic facies categories. Therefore, the deep neural network model with the help of the attention mechanism can be used to learn the regional correlation and continuity of seismic images, so as to depict the details of seismic phase boundaries more accurately.
  • the present invention proposes an automatic seismic phase identification method based on a combination of self-attention mechanism and U-shaped structure, which is used to predict the corresponding seismic phases of seismic data, thereby performing seismic interpretation.
  • the present invention provides an automatic seismic phase identification method based on a combination of self-attention mechanism and U-shaped structure.
  • the method includes the following steps:
  • the encoder is constructed using an overlapping block merging module with downsampling function and an efficient self-attention transformer module capable of global modeling representation;
  • the decoder is constructed using a block expansion module with linear upsampling function, the efficient self-attention transformer module and a skip connection module capable of fusing high- and low-level features;
  • the seismic phase recognition model includes a hypercolumn unitary divider
  • Construct a hybrid loss function use the training verification set in the sample training verification data set to iteratively train the seismic phase identification model; input test data to obtain the seismic phase of the test seismic data.
  • acquisition equipment can be used to collect seismic waves to obtain shot-collected seismic data.
  • the shot-collected seismic data can be superimposed to form a post-stack seismic data volume.
  • the shot collection here is a proper term for seismic data processing, which is excited by a primary earthquake source. And a type of seismic data received.
  • the efficient Transformer module is a semantic segmentation network.
  • obtaining a post-stack seismic data volume and preprocessing the post-stack seismic data volume to construct a sample training and verification data set specifically includes the following steps:
  • I, C, and D are respectively the number of main survey lines (Inline), the number of contact survey lines (Crossline), and the number of time sampling points, and the amplitude of the entire body is normalized to [0, 1].
  • the entire seismic volume training data is equally divided into 10 slice blocks along the Crossline direction; the first 70% of the sub-blocks in the Crossline direction within each slice block are taken as the training set, and the last 30% of the sub-blocks are taken as the verification set.
  • the seismic image formed by the slice inside each sub-block is resized, and the height and width (H, W) of the seismic image are transformed to the nearest multiple of 16 using the linear interpolation method, that is, the resolution is transformed into Finally, the sliced seismic image obtained after left and right flipping and Gaussian noise transformation is used as the input sample data set of the model.
  • constructing a model encoder using an overlapping block merging module with downsampling function and an efficient self-attention transformer module capable of global modeling representation specifically includes the following steps:
  • LN is layer normalization
  • MHSA is the multi-head self-attention calculation function
  • L 1 and L 2 are two fully connected functions
  • cv is the convolution layer function for position encoding.
  • constructing a model decoder using the block expansion module with linear upsampling function, the efficient self-attention transformer module and the skip connection module that can fuse high and low layer features specifically includes the following steps:
  • Multi-stage encoding features for encoder encoding Construct a decoder function f d such that The decoding process is symmetrical to the encoding process.
  • the decoder function f d includes 4 second sub-functions, and 4 second sub-functions There are 4 stages that make up the decoder.
  • the "skip connection" structure is used to connect the encoder and decoder features, so that the decoder receives the features of the encoder from the same stage for fusion to fuse the shallow coarse-grained and High-level fine-grained different semantic features.
  • constructing a seismic phase identification model using the encoder, the decoder and the hypercolumn module specifically includes the following steps:
  • Upample[ 2 i N C is the number of seismic phase categories.
  • the constructed seismic phase recognition model is shown in Figure 3.
  • the encoder and decoder form a U-shaped structure.
  • the left side of the U is the encoder and the right side is the decoder.
  • the output of each stage of the decoder is classified by super-structure fusion features for seismic phase classification.
  • the present invention names this network as a Hypercolumns-U-Segformer (HUSeg) network.
  • HUSeg Hypercolumns-U-Segformer
  • Pixel-level cross-entropy loss (CE) is selected as the main optimization goal of model training.
  • Dice loss is used to assist, so the specific loss function is:
  • y is the real seismic phase label of the seismic image
  • p is the prediction mask of the seismic image
  • y i, j represents the one-hot label corresponding to the pixel of the seismic image at i, j, Indicates the class probability that the pixel at position i and j of the seismic image is predicted to be the kth class seismic phase.
  • Adam optimizer with a batch size of 8, an initial learning rate of 1e-3, and a weight decay of 1e-4 to train and learn the model.
  • the method also includes the following steps:
  • the size of the seismic profile image is adjusted to a multiple of 16 to input the seismic phase recognition model to predict the seismic phase category.
  • the use of a batch stochastic gradient descent algorithm to iteratively update and learn the parameters of the model specifically includes the following steps:
  • the parameters of the seismic phase identification model are updated along the negative direction of the gradient to achieve a continuous decrease of the loss function.
  • the linear interpolation method is used to adjust the size of the seismic profile image to 16 times, and the model is input for prediction.
  • the size of the model output tensor is [H, W, N C ].
  • the largest probability index in each dimension is taken as the predicted seismic phase category, and the final output is a seismic phase matrix of size [H, W], where the value of each position represents the seismic phase category at the corresponding position of the input seismic image.
  • the public seismic data set from the F3 block of the Dutch North Sea annotated by Alaudah et al. (2019) was used for verification.
  • FIG 4 it is the seismic phase segmentation result of the seismic phase identification model proposed by the present invention on the test set inline400 seismic profile.
  • This experiment compared SegNet, UNet, Segformer and the HUSeg seismic phase automatic identification method proposed by the present invention respectively.
  • Figure 4(a) and (b) are respectively the test set inline400 seismic profile images and the corresponding seismic phase labels.
  • Figure 4(c)-(f) are respectively the earthquakes of SegNet, UNet, Segformer and the HUSeg model proposed by the present invention. Phase identification results.
  • the accuracy of the SegNet test set is 0.851
  • the accuracy of the UNet test set is 0.861
  • the accuracy of the Segformer test set is 0.903
  • the accuracy of the HUeg test set proposed by the present invention is 0.931.
  • the calculation amounts of UNet, Segformer and the HUSeg model proposed by the present invention are compared. Due to the lower accuracy of the SegNet model, there is no comparison of the calculation amount of the SegNet model here.
  • the calculation amounts of UNet, Segformer and the HUSeg model proposed by the present invention are 13.65 (billion), 4.41 (billion) and 4.37 (billion) respectively.
  • the HUSeg network model proposed by the present invention not only enables the computer to have a lower calculation amount, but also can obtain a higher seismic phase identification accuracy.
  • SegNet introduces some shallow seismic phase noise in the deep seismic phase area of the seismic image, and introduces a large amount of deep seismic phase noise in the middle seismic phase area.
  • UNet introduces a small amount of adjacent seismic phase noise in the shallow seismic phase area, and introduces a variety of seismic phase noise in the intermediate seismic phase area.
  • Segformer only introduces a large number of shallow seismic phases in the intermediate seismic phase area.
  • the HUSeg model proposed by the present invention is better than the results of all the above models. It not only removes the noise in the middle layer seismic phase, but also reduces a large amount of other seismic phase noise in the deep seismic phase area, and the boundary details of the seismic phase are more detailed. Smooth, lateral continuity is better.
  • the effectiveness of the present invention is further verified on a post-stack seismic data in the Bohai Bay Basin.
  • the main frequency of post-stack seismic data is 29Hz and the frequency band is 6-52Hz.
  • the HUSeg model of the present invention is used to identify seismic phases on all inline sections of the seismic body. The results are spliced into seismic phase bodies, and layer slices are taken to interpret sedimentary phases.
  • Figure 5 shows the three-dimensional seismic data, in which the Inline number is 7596 and the Xline number is 3180.
  • the stratigraphic slices and seismic facies identification results can be interpreted as sedimentary microfacies such as delta plains, swamps, distributary channels, natural dykes, underwater distributary bays, and semi-deep lakes. .

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • Evolutionary Computation (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Multimedia (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Human Resources & Organizations (AREA)
  • Strategic Management (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Economics (AREA)
  • Tourism & Hospitality (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Business, Economics & Management (AREA)
  • Development Economics (AREA)
  • Marketing (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Game Theory and Decision Science (AREA)
  • Operations Research (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Quality & Reliability (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Computer Security & Cryptography (AREA)
  • Educational Administration (AREA)
  • Primary Health Care (AREA)
  • Geophysics And Detection Of Objects (AREA)

Abstract

一种基于自注意力机制和U型结构的地震相自动识别方法，获得叠后地震数据体，并对叠后地震数据体进行预处理，以构建样本训练验证数据集；利用具有下采样功能的重叠块合并模块与能够全局建模表示的高效自注意力变换器模块构造编码器；利用具有线性上采样功能的块扩张模块、高效自注意力变换器模块与能够融合高低层特征的跳跃连接模块构造解码器；利用编码器、解码器与超列模块构造地震相识别模型，地震相识别模型包括超列酉分割器；构建混合损失函数，利用样本训练验证数据集中的训练验证集迭代训练地震相识别模型；输入测试数据，以获得测试地震数据地震相，通过公开数据和实际数据实验，对比传统卷积神经网络的地震相识别模型，具有较少的噪声点，边界细节更加明显、横向连续性更强，能够更加准确的解释地下不同的沉积环境。

Description

一种基于自注意力机制与U型结构结合的地震相自动化识别方法 技术领域
本发明属于地震勘探技术领域，涉及一种基于自注意力机制与U型结构结合的地震相自动化识别技术，尤其是一种利用Segformer自注意力分割网络、UNet网络结构和Hypercolumn语义分割技术相结合的地震图像语义分割方法，并将该地震图像语义分割方法用于地震数据的地震相自动化识别分类。
背景技术
随着人们对石油和天然气需求的日益增长与人工智能技术的快速发展，现代油气勘探技术正逐渐向智能与自动化发展。目前油气勘探主要是采用地震勘探方法，即通过人工地震波反射获取到叠后地震数据，并通过多学科知识挖掘分析出所蕴涵的地下构造、岩性和含油气性等信息，从而定位地下油气藏分布。传统的地震相划分方案是由解释人员手动解释或采用一些数学方法来半自动化地提取特征和分割地震相。然而这些传统方法具有很大的主观性，半自动方法准确性不够高，且时效性较差，无法在地下构造和沉积条件复杂的情况下准确定位油气储层。如何借助计算机资源，实现高效的自动化地震相识别方法，成为了当前需要解决的问题。
为了改进地震相自动化识别方法，基于深度学习的地震相分类方法被许多研究者们提了出来，该类方法是根据已有标签的地震数据，以端到端的形式学习一个从地震数据到地震相标签的非线性映射，并将其应用到未见过的地震数据上进行地震相分类。这种方法的优点是可以省去大量人工成本直接进行端到端的地震相分类，并在一定程度上提升解释效果。Jesper(2018)将ImageNet预训练好的VGG16网络迁移到手工标注的地震数据上，通过滑窗的方式来识别滑窗中心的地震相，从而实现地震相自动化分类。Zhao(2019)使用编码-解码结构的CNN网络来对地震相分类。Di(2019)标注了4个Inline剖面并采用一个类似于UNet的网络实现地震相自动化识别。上述基于深度学习的地震相识别方法虽然可以提高地震相识别精度，但是这些方法存在以下缺点，从而导致无法获得更加高效、准确的地震相解释结果。
以上技术具有如下缺点：
(1)将地震相识别任务当作单区域分类问题，而不是像素级语义分割问题， 因此计算量较大，存在大量的重复计算。
(2)上述方法都是基于窗口实现的，利用窗口内像素点组成块作为输入来预测中心一个点的地震相类型，只用到了局部特征，未用到全局信息。
发明内容
本发明的目的在于克服上述现有技术的缺点，提出一种基于自注意力机制与U型结构结合的地震相自动化识别方法，该方法在自注意力机制的语义分割网络的基础上，引入编码-解码U型结构，将该语义分割网络作为编码器模块，并在解码器中引入基于全连接层上采样的块扩张模块，使用超列技术融合特征，从而获得具有较低计算量、能够提取多尺度特征的全局注意力地震相分割方法。
本发明提供了一种基于注意力机制与U型结构结合的地震相自动识别方法，所述方法包括以下步骤：
获得叠后地震数据，并对所述叠后地震数据体进行预处理，以构建样本训练验证数据集；
利用具有下采样功能的重叠块合并模块与能够全局建模表示的高效自注意力变换器模块构造编码器；
利用具有线性上采样功能的块扩张模块、所述高效自注意力变换器模块与能够融合高低层特征的跳跃连接模块构造解码器；
利用所述编码器、所述解码器与超列模块构造地震相识别模型，所述地震相识别模型包括超列酉分割器；
构建混合损失函数，利用所述样本训练验证数据集中的训练验证集迭代训练所述地震相识别模型；输入测试数据，以获得测试地震数据地震相。
另外，根据本发明的基于自注意力机制与U型结构结合的地震相自动识别方法，还可以具有以下的技术特征：
所述获得叠后地震数据，并对所述叠后地震数据体进行预处理，以构建样本训练验证数据集具体包括以下步骤：
采集原始地震数据，进行预处理后得到叠后地震数据体，并将所述叠后地震数据的幅值归一化至[0,1]；
将所述叠后地震数据体沿着联络测线方向均分为N个切面块，其中每个所述切 面块由第一子块与第二子块构成，将所述第一子块作为训练集、第二子块作为验证集；其中，N为大于2的正整数，且第一子块的数量不少于第二子块的数量；
利用线性插值的方法将地震剖面图像的尺寸调整为16的倍数；
经过左右翻转、高斯噪声变换对所述地震剖面图像进行数据增广。
所述利用具有下采样功能的重叠块合并模块与能够全局建模表示的高效自注意力变换器模块构造模型编码器具体包括以下步骤：
对于一个输入高度、宽度分别为H、W的地震图像
Figure PCTCN2022108319-appb-000001
构造一编码器复合函数
Figure PCTCN2022108319-appb-000002
使得
Figure PCTCN2022108319-appb-000003
且对于特征图
Figure PCTCN2022108319-appb-000004
有第一子函数
Figure PCTCN2022108319-appb-000005
其中C i为第i个编码器输出特征图的通道数,所述第一子函数
Figure PCTCN2022108319-appb-000006
由1个所述重叠块合并模块和2个所述高效自注意力变换器模块组成，所述第一子函数
Figure PCTCN2022108319-appb-000007
包括4个，4个所述第一子函数组成所述编码器的4个连续的阶段，其中，所述重叠块合并模块通过步长小于核大小的卷积层实现，所述高效自注意力变换器模块包含自注意力子模块和前馈神经网络子模块，所述自注意力子模块以及所述前馈神经网络子模块的计算公式为：
sAtt(x)＝MHSA(LN(x))+x，     (1)
FFN(x)＝L 2(cv(L 1(LN(x))))+x    (2)
其中，LN为层归一化函数，MHSA为多头自注意力计算函数，L 1，L 2为两个全连接函数，cv为卷积层函数。
所述利用具有线性上采样功能的块扩张模块、所述高效自注意力变换器模块与能够融合高低层特征的跳跃连接模块构造模型解码器具体包括以下步骤：
构造一个解码器函数f d，使得
Figure PCTCN2022108319-appb-000008
所述解码器函数f d包括4个第二子函数，且4个第二子函数
Figure PCTCN2022108319-appb-000009
组成解码器的4个阶段，对于编码特征图
Figure PCTCN2022108319-appb-000010
与解码特征图
Figure PCTCN2022108319-appb-000011
有
Figure PCTCN2022108319-appb-000012
其中i＝{2，3，4}，d (4)＝x (4)，concat(·，·)为沿着通道维度张量拼接操作，最后一个阶段的特征图
Figure PCTCN2022108319-appb-000013
所述第二子函数
Figure PCTCN2022108319-appb-000014
由所述块扩张模块与2个所述高效Transformer模块构成；当输入特征图为x，所述块扩张模块的计算公式为：
x＝Linear[C，2C](x)，      (3)
Figure PCTCN2022108319-appb-000015
其中，Linear为全连接层，Reshape为维度重塑操作；
在编码器与解码器的对应阶段中通过所述跳跃连接模块连接，以使所述解码器接收来自相同阶段的所述编码器的特征进行融合。
所述利用所述编码器、所述解码器与超列模块构造地震相识别模型具体包括以下步骤：
引入超列结构对所述解码器的每个阶段输出特征图
Figure PCTCN2022108319-appb-000016
进行融合，并对融合后的特征图进行像素级别的地震相分类，其计算公式为：
Figure PCTCN2022108319-appb-000017
Figure PCTCN2022108319-appb-000018
Figure PCTCN2022108319-appb-000019
M＝Linear{C，N C](d f)      (8)
其中，Upsample[2 i×]表示上采样2 i倍的双线性插值操作，Concat为沿着通道维度拼接的操作，Linear[C，N C]为从维度C到维度N C的线性映射，N C为地震相类别数；所述编码器与所述解码器组成一个U型结构，所述U型结构的左边为所述编码器，所述U型结构的右边为所述解码器，所述解码器的各个阶段输出由所述超列结构融合特征进行地震相分类。
所述构建混合损失函数，利用所述样本训练验证数据集中的训练验证集迭代训练所述地震相识别模型；输入测试数据，以获得测试地震数据地震相具体包括以下步骤：
使用交叉熵损失函数与Dice Loss的混合损失对所述地震相识别模型进行训练，其损失函数的计算公式为：
Loss＝0.7*CE+0.3*Dice      (9)
其中，
Figure PCTCN2022108319-appb-000020
y为地震图像真实地震相标签，p为地震图像的预测掩码，y i，j表示地震图像在i，j处的像素对应的独热编码标签，
Figure PCTCN2022108319-appb-000021
表示地震图像在i，j处的像素预测为第k类地震相的类别概率。
所述方法还包括以下步骤：
训练所述地震相识别模型时，采用批量随机梯度下降算法来对所述地震相识别模型的参数进行迭代更新学习；
所述采用批量随机梯度下降算法来对模型的参数进行迭代更新学习采用批量随机梯度下降算法来对所述地震相识别模型的参数进行迭代更新学习具体包括以下步骤：
通过计算所述损失函数的梯度，沿着梯度负方向更新所述地震相识别模型的参数，以实现所述损失函数不断下降。
本发明的方法具有以下有益效果：
本发明将自注意力机制与U型结构结合，并采用超列技术分割的思想进行地震相自动化分割识别，同时引入混合损失函数，能够使得模型更关注分割的连续性。与已有的UNet和Segformer模型对比，本发明所提的地震相识别网络模型不仅使得计算机具有较低的计算量，而且能够获得较高的地震相识别准确率。进一步，利用具有高准确率的地震相识别结果更有效地预测地下沉积环境的位置和结构，为油气勘探提供有利的技术支撑。
附图说明
图1是本发明的一个实施例的编码器结构图；
图2是本发明的一个实施例的块扩张模块结构图；
图3是本发明的一个实施例的HUSeg网络结构图；
图4是本发明的一个实施例的不同语义分割方法用于地震相识别的对比图，其中，图4(a)为地震剖面图像；图4(b)为地震剖面对应的地震相标签图；图4(c) 为SegNet地震相识别结果图；图4(d)为SegNet地震相识别结果图；图4(e)为SegNet地震相识别结果图；图4(f)为本发明提出的地震相识别结果图；
图5为本发明的一个实施例的渤海湾盆地叠后三维地震数据体图；
图6为本发明的一个实施例的地震数据层位切片与识别结果图，其中，图6(a)为地震数据层位切片图；图6(b)本发明提出的地震相识别的层位切片图。
具体实施方法
下面结合附图1-图6对本发明提供的基于自注意力机制与U型结构的地震相自动化识别方法做进一步的详细描述：
叠后地震数据的地震相解释可以当作地震图像语义分割，将地震图像分割为不同区域，从而反映出地下的不同沉积环境，可以用来定位油气储层。但是，直接运用图像分割方法来划分地震相难以获得很好的效果，这些方法没有建模地震图像像素间的关联性，对于部分地震相类别的边界细节刻画效果较差。因此，借助于注意力机制深度神经网络模型可以用来学习地震图像的区域关联与连续性，从而对地震相边界细节刻画更加精确。本发明提出一种基于自注意力机制与U型结构结合的地震相自动识别方法，用于预测地震数据的对应地震相，从而进行地震解释。
本发明提供了一种基于自注意力机制与U型结构结合的地震相自动识别方法，所述方法包括以下步骤：
获得叠后地震数据体，并对所述叠后地震数据体进行预处理，以构建样本训练验证数据集；
利用具有下采样功能的重叠块合并模块与能够全局建模表示的高效自注意力变换器模块构造编码器；
利用具有线性上采样功能的块扩张模块、所述高效自注意力变换器模块与能够融合高低层特征的跳跃连接模块构造解码器；
利用所述编码器、所述解码器与超列模块构造地震相识别模型，所述地震相识别模型包括超列酉分割器；
构建混合损失函数，利用所述样本训练验证数据集中的训练验证集迭代训练所述地震相识别模型；输入测试数据，以获得测试地震数据地震相。
在具体实施中，可以采用采集设备对地震波进行采集得到炮集地震数据，炮集 地震数据可以叠加形成叠后地震数据体
Figure PCTCN2022108319-appb-000022
在地震勘探中，为了便于观测和分析，通常需要对叠后地震数据体进行处理以提高叠后地震数据体的分辨率，这里的炮集是地震数据处理的专有名词，是由一次震源激发而接收到的一类地震数据。
在具体实施中，所述高效Transformer模块为一种语义分割的网络。
在具体实施中，获得叠后地震数据体，并对所述叠后地震数据体进行预处理，以构建样本训练验证数据集具体包括以下步骤：
采集原始地震资料数据，并对原始地震资料数据进行预处理，得到叠后地震资料数据，记为
Figure PCTCN2022108319-appb-000023
其中I，C，D分别为主测线(Inline)数、联络测线(Crossline)数、时间采样点数，并将整个体的幅值归一化至[0，1]。将整个地震体训练数据沿着Crossline方向均分为10个切面块；取每个切面块内的Crossline方向前70％的子块作为训练集、后30％的子块作为验证集。对每个子块内部的切面形成的地震图像进行尺寸调整，利用线性插值的方法将地震图像的高宽(H，W)变换为最近的16的倍数，即分辨率变换为
Figure PCTCN2022108319-appb-000024
最后经过左右翻转、高斯噪声变换后得到的切面地震图像作为模型的输入样本数据集。
在具体实施中，利用具有下采样功能的重叠块合并模块与能够全局建模表示的高效自注意力变换器模块构造模型编码器具体包括以下步骤：
对于一个输入高度、宽度分别为H、W的地震图像
Figure PCTCN2022108319-appb-000025
构造一个编码器复合函数
Figure PCTCN2022108319-appb-000026
使得
Figure PCTCN2022108319-appb-000027
且对特征图
Figure PCTCN2022108319-appb-000028
有第一子函数
Figure PCTCN2022108319-appb-000029
其中C i为编码器第i个子函数输出特征图的通道数。如图1所示，编码器的每个第一子函数
Figure PCTCN2022108319-appb-000030
由1个重叠块合并模块和2个高效Transformer模块组成，称之为子编码块，每个子编码块都能够使得输入特征图的空间维度减半，从而4个子编码块组成编码器的4个连续的阶段，可以得到4个不同尺度的特征图，以提供不同抽象等级特征。
在每个子编码块中，重叠块合并模块通过有重叠的滑动窗口方式对输入数据进行矩阵乘法，从而实现特征图线性嵌入，能够将输入数据的空间分辨率维度下采样2倍。高效Transformer模块包含自注意力子模块和前馈神经网络子模块，用以学习切 面图像中不同位置间的全局特征与融合特征。对于第i个阶段的输入特征图x (i)，该子编码块计算可以表示为：
Figure PCTCN2022108319-appb-000031
其中，ο为复合函数运算符，Conv为使用卷积运算实现的重叠块合并模块，sAtt为自注意力子模块计算函数，FFN为高效前馈神经网络子模块计算函数。两者的计算公式如下：
sAtt(x)＝MHSA(LN(x))+x，      (1)
FFN(x)＝L 2(cv(L 1(LN(x))))+x      (2)
其中LN为层归一化，MHSA为多头自注意力计算函数，L 1，L 2为两个全连接函数，cv为卷积层函数，用以位置编码。
在具体实施中，利用具有线性上采样功能的块扩张模块、所述高效自注意力变换器模块与能够融合高低层特征的跳跃连接模块构造模型解码器具体包括以下步骤：
对于编码器编码的多阶段编码特征
Figure PCTCN2022108319-appb-000032
构造一个解码器函数f d，使得
Figure PCTCN2022108319-appb-000033
解码过程采用与编码过程对称的方式，解码器函数f d包括4个第二子函数，且4个第二子函数
Figure PCTCN2022108319-appb-000034
组成解码器的4个阶段。对于编码特征图
Figure PCTCN2022108319-appb-000035
与解码特征图
Figure PCTCN2022108319-appb-000036
有
Figure PCTCN2022108319-appb-000037
其中i＝{2，3，4}，d (4)＝x (4)，[·，·]为沿着通道维度张量拼接操作。最后一个阶段的特征图
Figure PCTCN2022108319-appb-000038
其中，所述第二子函数
Figure PCTCN2022108319-appb-000039
通过1个基于全连接层上采样的块扩张模块与2个高效Transformer模块一起构成，称之为子解码块。如图2所示为一个块扩张模块，其思想是用通道维度换空间维度，通过一个全连接层将数据的通道维度扩大两倍后进行维度重塑操作。具体操作为将其通道拆分成四部分，每两部分沿着空间维度进行交错拼接，从而将空间维度增倍同时通道维度减半。假设输入特征图为x，所述的块扩 张模块可以描述为：
x＝Linear[C，2C](x)，     (3)
Figure PCTCN2022108319-appb-000040
其中，Linear为全连接层，Reshape为维度重塑操作。
同时在编码器与解码器的对应阶段中通过“跳跃连接”结构连接，拼接编码器与解码器特征，从而让解码器接收来自相同阶段的编码器的特征进行融合，以融合浅层粗粒度和高层细粒度的不同语义特征。
在具体实施中，利用所述编码器、所述解码器与超列模块构造地震相识别模型具体包括以下步骤：
引入超列结构对解码器的每个阶段输出特征图
Figure PCTCN2022108319-appb-000041
进行融合，然后对融合后的特征图进行像素级别的地震相分类。这样综合使用多个层次的特征进行分类能够提升密集预测任务的效果，其计算过程为：
Figure PCTCN2022108319-appb-000042
Figure PCTCN2022108319-appb-000043
Figure PCTCN2022108319-appb-000044
M＝Linear[C，N C](d f)      (8)
其中，Upample[2 i×]表示上采样2 i倍的双线性插值操作，Concat为沿着通道维度拼接的操作，Linear[C，N C]为从维度C到维度N C的线性映射，N C为地震相类别数。构建的地震相识别模型如图3所示，编码器与解码器组成一个U型结构，U左边为编码器，右边为解码器，解码器的各个阶段输出由超列结构融合特征进行地震相分类，本发明将该网络命名为超列酉分割器(Hypercolumns-U-Segformer，HUSeg)网络。
在具体实施中，构建混合损失函数，利用所述样本训练验证数据集中的训练验证集迭代训练所述地震相识别模型；输入测试数据，以获得测试地震数据地震相，具体包括以下步骤：
选择像素级别的交叉熵损失(CE)作为模型训练的主要优化目标，同时为了让模型考察区域相关性，加以Dice损失进行辅助，所以具体的损失函数为：
Loss＝0.7*CE+0.3*Dice      (9)
其中，
Figure PCTCN2022108319-appb-000045
y为地震图像真实地震相标签，p为地震图像的预测掩码，y i，j表示地震图像在i，j处的像素对应的one-hot标签，
Figure PCTCN2022108319-appb-000046
表示地震图像在i，j处的像素预测为第k类地震相的类别概率。训练模型时，采用批大小为8、初始学习率为1e-3、权重衰减为1e-4的Adam优化器对模型进行训练学习。
在具体实施中，所述方法还包括以下步骤：
训练所述地震相识别模型时，采用批量随机梯度下降算法来对所述地震相识别模型的参数进行迭代更新学习；
所述地震相识别模型训练完毕后，将所述地震剖面图像的尺寸调整至16倍数，以输入所述地震相识别模型预测地震相类别。
在具体实施中，所述采用批量随机梯度下降算法来对模型的参数进行迭代更新学习采用批量随机梯度下降算法来对所述地震相识别模型的参数进行迭代更新学习具体包括以下步骤：
通过计算所述损失函数的梯度，沿着梯度负方向更新所述地震相识别模型的参数，以实现所述损失函数不断下降。
当模型训练完毕后，对于任意地震剖面，利用线性插值方法将地震剖面图像的尺寸调整至16倍数，输入模型进行预测，模型输出张量的大小为[H，W，N C]，在第三个维度取最大的概率索引作为预测的地震相类别，最终的输出一个大小为[H，W]的地震相矩阵，其中每个位置的值代表输入地震图像对应位置的地震相类别。
数值仿真结果
首先，为了验证本发明的有效性，采用Alaudah等人(2019)标注的来自荷兰北海F3区块的公开地震数据集用于验证。选取其中Inline 300-700、Crossline 300-1000的部分作为训练集，其余作为测试集。按照说明书具体实施方法的步骤(2)进行预处理，其次按照步骤(6)进行训练测试，得到的训练好的模型，并进行评估。
如图4所示，为本发明提出的地震相识别模型在测试集inline400地震剖面上的 地震相分割结果，本实验分别对比了SegNet、UNet、Segformer和本发明提出的HUSeg地震相自动识别方法。图4(a)和(b)分别为测试集inline400地震剖面图像与及对应的地震相标签，图4(c)-(f)分别为SegNet、UNet、Segformer和本发明提出的HUSeg模型的地震相识别结果。SegNet测试集的准确率为0.851，UNet测试集的准确率为0.861，Segformer测试集的准确率为0.903，本发明提出的HUeg测试集的准确率为0.931。为了进一步说明本发明所提网络模型的优势，对比UNet、Segformer和本发明提出的HUSeg模型的计算量。由于SegNet模型的准确率较低，这里没有对比SegNet模型的计算量。当输入样本为128X128的图像时，UNet、Segformer和本发明提出的HUSeg模型的计算量分别为13.65(十亿)、4.41(十亿)和4.37(十亿)。也就是说，本发明所提的HUSeg网络模型不仅使得计算机具有较低的计算量，而且能够获得更高的地震相识别准确率。SegNet在地震图像的深层地震相区域引入了部分浅层地震相噪声，且在中间的地震相区域引入了大量深层地震相噪声。UNet在浅层地震相区域引入了少量相邻的地震相噪声，且在中间的地震相区域引入了多种地震相噪声。Segformer则只在中间层地震相区域引入了大量浅层地震相。而本发明提出的HUSeg模型效果优于以上所有模型的结果，其不仅去除了中间层地震相中的噪声，而且在深层地震相区域也减少了大量其他地震相噪声，其地震相的边界细节更加平滑、横向连续性更好。
最后，在渤海湾盆地的一个叠后地震数据上来进一步验证本发明的有效性。叠后地震资料主频为29Hz、频带为6-52Hz，使用本发明的HUSeg模型对地震体所有inline剖面进行地震相识别，将结果拼接成地震相体，并取层切片解释沉积相。图5显示了三维地震数据，其中Inline编号为7596,Xline编号为3180。如图6所示的层切片与地震相识别结果，沿沙三中亚段的地层切片可解释为三角洲平原、沼泽、分流水道、天然堤、水下分流间湾和半深湖等沉积微相。

Claims (8)

  1. 一种基于自注意力机制与U型结构结合的地震相自动识别方法，其特征在于，包括以下步骤：
    获得叠后地震数据体，并对所述叠后地震数据体进行预处理，以构建样本训练验证数据集；
    利用具有下采样功能的重叠块合并模块与能够全局建模表示的高效自注意力变换器模块构造编码器；
    利用具有线性上采样功能的块扩张模块、所述高效自注意力变换器模块与能够融合高低层特征的跳跃连接模块构造解码器；
    利用所述编码器、所述解码器与超列模块构造地震相识别模型，所述地震相识别模型包括超列酉分割器；
    构建混合损失函数，利用所述样本训练验证数据集中的训练验证集迭代训练所述地震相识别模型；输入测试数据，以获得测试地震数据地震相。
  2. 根据权利要求1所述的基于自注意力机制与U型结构结合的地震相自动识别方法，其特征在于，所述获得叠后地震数据体，并对所述叠后地震数据体进行预处理，以构建样本训练验证数据集具体包括以下步骤：
    采集原始地震数据，进行预处理后得到叠后地震数据体，并将所述叠后地震数据的幅值归一化至[0,1]；
    将所述叠后地震数据体沿着联络测线方向均分为N个切面块，其中每个所述切面块由第一子块与第二子块构成，将所述第一子块作为训练集、第二子块作为验证集；其中，N为大于2的正整数，且第一子块的数量不少于第二子块的数量；
    利用线性插值的方法将地震剖面图像的尺寸调整为16的倍数；
    经过左右翻转、高斯噪声变换对所述地震剖面图像进行数据增广。
  3. 根据权利要求2所述的基于自注意力机制与U型结构结合的地震相自动识别方法，其特征在于，所述利用具有下采样功能的重叠块合并模块与能够全局建模表示的高效自注意力变换器模块构造模型编码器具体包括以下步骤：
    对于一个输入高度、宽度分别为H、W的地震图像
    Figure PCTCN2022108319-appb-100001
    构造一编码器复合函数
    Figure PCTCN2022108319-appb-100002
    使得
    Figure PCTCN2022108319-appb-100003
    且对于特征图
    Figure PCTCN2022108319-appb-100004
    有第一子函数
    Figure PCTCN2022108319-appb-100005
    其中C i为第i个编码器输出特征图的通道数,所述第一子函数
    Figure PCTCN2022108319-appb-100006
    由1个所述重叠块合并模块和2个所述高效自注意力变换器模块组成，所 述第一子函数
    Figure PCTCN2022108319-appb-100007
    包括4个，4个所述第一子函数组成所述编码器的4个连续的阶段，其中，所述重叠块合并模块通过步长小于核大小的卷积层实现，所述高效自注意力变换器模块包含自注意力子模块和前馈神经网络子模块，所述自注意力子模块以及所述前馈神经网络子模块的计算公式为：
    sAtt(x)＝MHSA(LN(x))+x，    (I)
    FFN (x)＝L 2(cv(L 1(LN(x))))+x    (2)
    其中，LN为层归一化函数，MHSA为多头自注意力计算函数，L 1，L 2为两个全连接函数，cv为卷积层函数。
  4. 根据权利要求3所述的基于自注意力机制与U型结构结合的地震相自动识别方法，其特征在于，所述利用具有线性上采样功能的块扩张模块、所述高效自注意力变换器模块与能够融合高低层特征的跳跃连接模块构造模型解码器具体包括以下步骤：
    构造一个解码器函数f d，使得
    Figure PCTCN2022108319-appb-100008
    所述解码器函数f d包括4个第二子函数，且4个第二子函数
    Figure PCTCN2022108319-appb-100009
    组成解码器的4个阶段，对于编码特征图
    Figure PCTCN2022108319-appb-100010
    与解码特征图
    Figure PCTCN2022108319-appb-100011
    有
    Figure PCTCN2022108319-appb-100012
    其中i＝{2，3，4}，d (4)＝x (4)，concat(·，·)为沿着通道维度张量拼接操作，最后一个阶段的特征图
    Figure PCTCN2022108319-appb-100013
    所述第二子函数
    Figure PCTCN2022108319-appb-100014
    由所述块扩张模块与2个所述高效Transformer模块构成；当输入特征图为x，所述块扩张模块的计算公式为：
    x＝Linear[C,2C](x),    (3)
    Figure PCTCN2022108319-appb-100015
    其中，Linear为全连接层，Reshape为维度重塑操作；
    在编码器与解码器的对应阶段中通过所述跳跃连接模块连接，以使所述解码器接收来自相同阶段的所述编码器的特征进行融合。
  5. 根据权利要求4所述的基于自注意力机制与U型结构结合的地震相自动识别方 法，其特征在于，所述利用所述编码器、所述解码器与超列模块构造地震相识别模型具体包括以下步骤：
    引入超列结构对所述解码器的每个阶段输出特征图
    Figure PCTCN2022108319-appb-100016
    进行融合，并对融合后的特征图进行像素级别的地震相分类，其计算公式为：
    Figure PCTCN2022108319-appb-100017
    Figure PCTCN2022108319-appb-100018
    Figure PCTCN2022108319-appb-100019
    M＝Linear[C，N c](d f)    (8)
    其中，Upsample[2 i×]表示上采样2 i倍的双线性插值操作，Concat为沿着通道维度拼接的操作，Linear[C，N c]为从维度C到维度N c的线性映射，N c为地震相类别数；所述编码器与所述解码器组成一个U型结构，所述U型结构的左边为所述编码器，所述U型结构的右边为所述解码器，所述解码器的各个阶段输出由所述超列结构融合特征进行地震相分类。
  6. 根据权利要求5所述的基于自注意力机制与U型结构结合的地震相自动识别方法，其特征在于，所述构建混合损失函数，利用所述样本训练验证数据集中的训练验证集迭代训练所述地震相识别模型；输入测试数据，以获得测试地震数据地震相具体包括以下步骤：
    使用交叉熵损失函数与Dice Loss的混合损失对所述地震相识别模型进行训练，其损失函数的计算公式为：
    Loss＝0.7*CE+0.3*Dice    (9)
    其中，
    Figure PCTCN2022108319-appb-100020
    y为地震图像真实地震相标签，p为地震图像的预测掩码，y i，j表示地震图像在i，j处的像素对应的独热编码标签，
    Figure PCTCN2022108319-appb-100021
    表示地震图像在i，j处的像素预测为第k类地震相的类别概率。
  7. 根据权利要求6所述的基于自注意力机制与U型结构结合的地震相自动识别方法，其特征在于，所述方法还包括以下步骤：
    训练所述地震相识别模型时，采用批量随机梯度下降算法来对所述地震相识别模型的参数进行迭代更新学习；
    所述地震相识别模型训练完毕后，将所述地震剖面图像的尺寸调整至16倍数，以输入所述地震相识别模型预测地震相类别。
  8. 根据权利要求7所述的基于自注意力机制与U型结构结合的地震相自动识别方法，其特征在于，所述采用批量随机梯度下降算法来对模型的参数进行迭代更新学习采用批量随机梯度下降算法来对所述地震相识别模型的参数进行迭代更新学习具体包括以下步骤：
    通过计算所述损失函数的梯度，沿着梯度负方向更新所述地震相识别模型的参数，以实现所述损失函数不断下降。
PCT/CN2022/108319 2022-06-30 2022-07-27 一种基于自注意力机制与u型结构结合的地震相自动化识别方法 WO2024000709A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US18/321,632 US11922679B2 (en) 2022-06-30 2023-05-22 Automatic seismic facies identification method based on combination of self-attention mechanism and u-shape network architecture

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210759364.2A CN115081719A (zh) 2022-06-30 2022-06-30 一种基于自注意力机制与u型结构结合的地震相自动化识别方法
CN202210759364.2 2022-06-30

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US18/321,632 Continuation US11922679B2 (en) 2022-06-30 2023-05-22 Automatic seismic facies identification method based on combination of self-attention mechanism and u-shape network architecture

Publications (1)

Publication Number Publication Date
WO2024000709A1 true WO2024000709A1 (zh) 2024-01-04

Family

ID=83254631

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/108319 WO2024000709A1 (zh) 2022-06-30 2022-07-27 一种基于自注意力机制与u型结构结合的地震相自动化识别方法

Country Status (2)

Country Link
CN (1) CN115081719A (zh)
WO (1) WO2024000709A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117668669A (zh) * 2024-02-01 2024-03-08 齐鲁工业大学(山东省科学院) 基于改进YOLOv7的管道安全监测方法及系统
CN117830788A (zh) * 2024-03-06 2024-04-05 潍坊科技学院 一种多源信息融合的图像目标检测方法

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110516740A (zh) * 2019-08-28 2019-11-29 电子科技大学 一种基于Unet++卷积神经网络的断层识别方法
CN111626355A (zh) * 2020-05-27 2020-09-04 中油奥博(成都)科技有限公司 一种基于Unet++卷积神经网络的地震数据初至拾取方法
US20210063594A1 (en) * 2019-09-03 2021-03-04 Chevron U.S.A. Inc. System and method for seismic imaging of subsurface volumes including complex geology
CN112711072A (zh) * 2020-12-23 2021-04-27 西安交通大学 基于Res U-net的三维地震数据断层识别方法
CN112731522A (zh) * 2020-12-14 2021-04-30 中国地质大学（武汉） 地震地层智能识别方法、装置、设备及存储介质
US20210247531A1 (en) * 2020-02-07 2021-08-12 POSTECH Research and Business Development Foundation Early earthquake detection apparatus and method
CN113296152A (zh) * 2020-02-21 2021-08-24 中国石油天然气集团有限公司 断层检测方法及装置
CN114660656A (zh) * 2022-03-17 2022-06-24 中国科学院地质与地球物理研究所 一种地震数据初至拾取方法及系统

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110516740A (zh) * 2019-08-28 2019-11-29 电子科技大学 一种基于Unet++卷积神经网络的断层识别方法
US20210063594A1 (en) * 2019-09-03 2021-03-04 Chevron U.S.A. Inc. System and method for seismic imaging of subsurface volumes including complex geology
US20210247531A1 (en) * 2020-02-07 2021-08-12 POSTECH Research and Business Development Foundation Early earthquake detection apparatus and method
CN113296152A (zh) * 2020-02-21 2021-08-24 中国石油天然气集团有限公司 断层检测方法及装置
CN111626355A (zh) * 2020-05-27 2020-09-04 中油奥博(成都)科技有限公司 一种基于Unet++卷积神经网络的地震数据初至拾取方法
CN112731522A (zh) * 2020-12-14 2021-04-30 中国地质大学（武汉） 地震地层智能识别方法、装置、设备及存储介质
CN112711072A (zh) * 2020-12-23 2021-04-27 西安交通大学 基于Res U-net的三维地震数据断层识别方法
CN114660656A (zh) * 2022-03-17 2022-06-24 中国科学院地质与地球物理研究所 一种地震数据初至拾取方法及系统

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117668669A (zh) * 2024-02-01 2024-03-08 齐鲁工业大学(山东省科学院) 基于改进YOLOv7的管道安全监测方法及系统
CN117668669B (zh) * 2024-02-01 2024-04-19 齐鲁工业大学(山东省科学院) 基于改进YOLOv7的管道安全监测方法及系统
CN117830788A (zh) * 2024-03-06 2024-04-05 潍坊科技学院 一种多源信息融合的图像目标检测方法
CN117830788B (zh) * 2024-03-06 2024-05-10 潍坊科技学院 一种多源信息融合的图像目标检测方法

Also Published As

Publication number Publication date
CN115081719A (zh) 2022-09-20

Similar Documents

Publication Publication Date Title
WO2024000709A1 (zh) 一种基于自注意力机制与u型结构结合的地震相自动化识别方法
CN112070779B (zh) 一种基于卷积神经网络弱监督学习的遥感影像道路分割方法
US11922679B2 (en) Automatic seismic facies identification method based on combination of self-attention mechanism and u-shape network architecture
CN107368787B (zh) 一种面向深度智驾应用的交通标志识别方法
CN112287807B (zh) 一种基于多分支金字塔神经网络的遥感影像道路提取方法
CN110516740A (zh) 一种基于Unet++卷积神经网络的断层识别方法
CN114943963B (zh) 一种基于双分支融合网络的遥感图像云和云影分割方法
CN112731522B (zh) 地震地层智能识别方法、装置、设备及存储介质
CN111080021B (zh) 一种基于地质信息库的砂体构型cmm神经网络预测方法
CN113160246A (zh) 一种基于深度监督的图像语义分割方法
Yang et al. Seismic fault detection based on 3D Unet++ model
CN116704137B (zh) 一种海上石油钻井平台点云深度学习逆向建模方法
CN116580241B (zh) 基于双分支多尺度语义分割网络的图像处理方法及系统
Wang et al. Seismic stratum segmentation using an encoder–decoder convolutional neural network
CN116452982A (zh) 一种基于改进YOLOv5的探地雷达地下排水管网图像检测方法
Wu et al. Variable seismic waveforms representation: Weak-supervised learning based seismic horizon picking
CN117173577A (zh) 一种基于改进Swin Transformer的遥感影像建筑物变化检测方法
Li et al. Fault-Seg-Net: A method for seismic fault segmentation based on multi-scale feature fusion with imbalanced classification
Zhang et al. Phase arrival picking for bridging multi-source downhole microseismic data using deep transfer learning
CN117251674A (zh) 一种地震相分类方法、装置、电子设备及介质
CN117031539A (zh) 一种自监督深度学习地震数据低频重建方法及系统
Liu et al. ASHFormer: axial and sliding window based attention with high-resolution transformer for automatic stratigraphic correlation
CN115880505A (zh) 一种目标边缘检测神经网络的低序级断层智能识别方法
CN115639605A (zh) 基于深度学习的高分辨率断层的自动识别方法和装置
CN115880455A (zh) 基于深度学习的三维智能插值方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22948804

Country of ref document: EP

Kind code of ref document: A1