CN-122017979-A - Method and device for identifying U-NET fracture based on transducer enhancement

CN122017979ACN 122017979 ACN122017979 ACN 122017979ACN-122017979-A

Abstract

The invention relates to the technical field of AI seismic attribute detection, and particularly discloses a method and a device for identifying U-NET fracture based on transform enhancement, wherein the method comprises the steps of performing artificial fault interpretation on original three-dimensional seismic data to obtain a fault pattern identification sample, inputting the fault pattern identification sample into a network of a U-NET architecture, and performing feature extraction, feature fusion, up-sampling, feature processing, loss calculation and parameter optimization to obtain a training model; and applying the optimal training model to actual work to perform fault identification on new seismic data. According to the method, after the network is fully trained, the fault positions in the seismic data can be accurately identified, particularly for obvious faults, the identification accuracy can reach a high level, and for the seismic data with a plurality of fault interlaces, irregular fault forms and noise interference, the characteristics of different faults can be distinguished, so that each fault can be accurately identified.

Inventors

ZHANG WEI
GAO QIANG
MA WENYU
LIU KEXING
HAN XUANYING
MA XIAOGANG

Assignees

中国石油化工股份有限公司
中石化石油物探技术研究院有限公司

Dates

Publication Date: 20260512
Application Date: 20241108

Claims (10)

1. The method for identifying the U-NET fracture based on the enhancement of the transducer is characterized by comprising the following steps of: Performing artificial fault interpretation on the original three-dimensional seismic data to obtain a fault pattern recognition sample; inputting the fault pattern recognition sample into a network of a U-NET architecture, and performing feature extraction, feature fusion, up-sampling, feature processing, loss calculation and parameter optimization to obtain a training model; Evaluating the training model, and adjusting and optimizing the training model according to the evaluation result to obtain an optimal training model; And applying the optimal training model to actual fault identification work of the seismic data, and carrying out fault identification on new seismic data.
2. The method of claim 1, wherein manually fault interpreting the raw three-dimensional seismic data to obtain fault pattern recognition samples comprises: converting the original three-dimensional seismic data into two-dimensional slices, namely an isochronous/deep slice and a seismic section; Performing artificial fault interpretation on the two-dimensional slice, and converting the results of the seismic section and the artificial fault interpretation into an image file; And taking the image file as a fault pattern recognition sample.
3. The method of claim 1, wherein the feature extraction comprises: Performing feature extraction on input data through a series of convolution operations in an encoder of the network, wherein each time the convolution operation is performed, the network extracts more abstract and representative features; And carrying out a maximum pooling operation, wherein the maximum pooling operation is used for keeping main characteristic information while reducing the data dimension so as to help the network to focus on more key characteristics gradually.
4. The method of claim 1, wherein the feature fusion, i.e., cross-layer connection, comprises: In the down sampling process of the network, the characteristics of different levels are fused, so that the network comprehensively utilizes the characteristic information of different abstract levels, and excessive loss of information in the dimension reduction process is avoided.
5. The method of claim 1, wherein the upsampling and feature processing comprises: Gradually restoring the dimension of the data through an up-sampling operation in a decoder of the network for making the dimension of the data approximate to the original input data; After the upsampling, further adjusting and processing features in combination with convolution operations for making feature information more accurate and targeted; the method adopted by the up-sampling is a deconvolution or interpolation method.
6. The method of claim 1, wherein the loss calculation and parameter optimization comprises: comparing the output of the network with the artificially marked labels, and calculating the difference between the predicted result and the real labels by adopting a cross entropy loss function to obtain a loss value; And according to the loss value, adopting a random gradient descent and a variant algorithm thereof to carry out back propagation and updating on the parameters of the network, and continuously adjusting the weight and bias of the network so as to enable the predicted result of the network to be more and more close to the real fault condition.
7. The method of claim 1, wherein evaluating the training model, and adjusting and optimizing the training model based on the result of the evaluation to obtain an optimal training model comprises: Evaluating the training model by using an independent test data set, and measuring the performance of the network on fault identification of the seismic data by calculating accuracy, recall rate and F1-score index; And adjusting and optimizing the structure and the super parameters of the network according to the evaluation result to obtain an optimal training model.
8. A fransformer-enhanced U-NET fracture identification device, comprising: the fault interpretation module is used for performing artificial fault interpretation on the original three-dimensional seismic data to obtain a fault pattern recognition sample; the model training module is used for inputting the fault pattern recognition sample into a network of a U-NET architecture, and carrying out feature extraction, feature fusion, up-sampling and feature processing, loss calculation and parameter optimization to obtain a training model; The model evaluation module is used for evaluating the training model, and adjusting and optimizing the training model according to the evaluation result to obtain an optimal training model; and the model application module is used for applying the optimal training model to the actual fault identification work of the seismic data and carrying out fault identification on the new seismic data.
9. An electronic device, the electronic device comprising: A memory storing executable instructions; A processor executing the executable instructions in the memory to implement a Transformer enhanced based U-NET break identification method according to any one of claims 1-7.
10. A computer readable storage medium, characterized in that it stores a computer program, which when executed by a processor implements the transformation-enhanced U-NET fracture identification method according to any one of claims 1-7.

Description

Method and device for identifying U-NET fracture based on transducer enhancement Technical Field The invention relates to the technical field of AI seismic attribute detection, in particular to a U-NET fracture identification method and device based on transform enhancement. Background Fracture prediction is one of the core works of seismic exploration interpretation, and is directly related to the efficiency and benefit of oil and gas field exploration and development. The main method for fracture prediction is to extract and analyze spatial attributes sensitive to fracture characterization by utilizing pre-stack and post-stack seismic data information. In the early days, seismic data fault identification was performed mostly by geologic specialist labor. This method is highly dependent on expert expertise and practical experience, and requires a lot of time and labor. With the development of machine learning technology, some methods based on traditional machine learning algorithms are applied to seismic data fault identification. Firstly, feature engineering is needed, namely, various features such as amplitude, frequency, phase and other attributes of seismic waves are manually extracted from seismic data, and some features are obtained based on texture analysis, and commonly used models are Support Vector Machines (SVMs), decision trees, random forests and the like. A relatively large, representative, and high quality data is typically required to train the model to achieve good results. Early deep learning methods, such as simple Convolutional Neural Networks (CNNs), can progressively extract features from low-level to high-level through multi-layer convolution and pooling operations. However, in terms of fusing different levels of features, it is relatively simple to rely on pooling and fully connected layers to integrate features, and there are limitations in processing seismic data, which is data with a complex structure. At present, the machine learning method has been applied to a certain extent in the field of seismic exploration, and is mainly suitable for two main types of methods, namely deep learning and pattern recognition, but the based samples are in a single-point form, and do not really take the form of fracture patterns as samples to participate in training, so that the predicted fracture detection results are scattered on the section and the plane, and even are opposite to the actual fracture direction. Based on the technical background, the invention researches a method and a device for identifying the U-NET fracture based on the enhancement of a transducer. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a U-NET fracture identification method and device based on the enhancement of a transducer, which can accurately identify fault positions in seismic data after a network is fully trained, particularly for obvious faults, the identification accuracy can reach a higher level, the trend and range of the faults can be clearly outlined, the situations of misjudgment and missed judgment are reduced, and for the seismic data with a plurality of fault interlaces, irregular fault forms and noise interference, the characteristics of different faults can be distinguished through a multi-layer characteristic extraction and fusion mechanism of the network, so that each fault can be accurately identified. To achieve the above object, a first aspect of the present invention provides a method for identifying U-NET fracture based on transducer enhancement, comprising: Performing artificial fault interpretation on the original three-dimensional seismic data to obtain a fault pattern recognition sample; inputting the fault pattern recognition sample into a network of a U-NET architecture, and performing feature extraction, feature fusion, up-sampling, feature processing, loss calculation and parameter optimization to obtain a training model; Evaluating the training model, and adjusting and optimizing the training model according to the evaluation result to obtain an optimal training model; And applying the optimal training model to actual fault identification work of the seismic data, and carrying out fault identification on new seismic data. A second aspect of the present invention provides a fransformer-enhanced U-NET fracture identification device, comprising: the fault interpretation module is used for performing artificial fault interpretation on the original three-dimensional seismic data to obtain a fault pattern recognition sample; the model training module is used for inputting the fault pattern recognition sample into a network of a U-NET architecture, and carrying out feature extraction, feature fusion, up-sampling and feature processing, loss calculation and parameter optimization to obtain a training model; The model evaluation module is used for evaluating the training model, and adjusting and optimizing the training model according to the evaluation