Near-surface fault automatic identification method
By introducing multi-scale residual modules and feature fusion attention modules in U-Net networks, fault recognition capabilities are improved, the shortcomings of traditional U-Net in complex fault recognition are solved, and more efficient fault feature extraction and recognition effects are achieved. .
Patent Information
- Application Number
- CN202411204321.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-30
- Publication Date
- 2024-12-06
AI Technical Summary
Traditional U-Net has limited ability to extract and restore feature information in fault recognition, and cannot effectively identify complex faults, especially in the recognition of small-scale faults at near-surface.
Improved the U-Net network structure, introduced multi-scale residual module, residual jump connection and feature fusion attention module, built an encoder and decoder, extracted features through multi-scale residual module, and connected multiple connections after each operation. The scale feature fusion attention module uses residual jump connection instead of the traditional encoder and decoder, which improves the network's feature extraction capability.
The improved U-Net network can extract image features more effectively from all aspects, improve the recognition ability and training speed of near-surface faults, and can identify complex and tiny faults more clearly, significantly improving the efficiency and accuracy of fault recognition. .
Smart Images

Figure CN119087511A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of fault identification in seismic data interpretation, and in particular, relates to a method for automatic identification of near-surface faults. Background Art
[0002] In the process of near-surface Q modeling, the strata have lateral discontinuities due to the existence of faults, resulting in large lateral variations in Q, and the Q values between fault blocks need to be modeled separately. Therefore, accurately interpreting the development characteristics of near-surface faults is of great significance for near-surface Q modeling.
[0003] Traditional fault identification methods are easily affected by earthquake quality and interpretation experience. For small-scale fault identification methods such as near-surface faults, further improvement is needed. The commonly used U-Net has achieved satisfactory results in fault identification. However, the network has limited ability to extract and restore feature information and has no attention mechanism, which limits its ability to identify complex faults. Summary of the invention
[0004] In response to the problem that the ability to identify tiny faults is limited, an embodiment of the present application provides a method for automatic identification of near-surface faults, which transforms and improves the two-layer convolution, jump connection and attention mechanism of the traditional U-net network, and introduces a multi-scale residual module, residual jump connection and feature fusion attention module.
[0005] The embodiment of the present application provides a method for automatic identification of near-surface faults, including: S1, preparing post-stack seismic data; S2, building a neural network, adjusting parameters of each part, and the improved U-Net includes an encoder, a decoder, residual jump connections and an end-to-end image segmentation model; S3, constructing training data and producing label data; S4, performing neural network training; S5, using the trained network to apply actual data.
[0006] Among them, in step S2, the encoder on the left side of the improved U-Net includes a multi-scale residual module, an attention module and a maximum pooling; the multi-scale residual module is used to extract and restore features, and a multi-scale feature fusion attention module is connected after each multi-scale residual module.
[0007] Among them, in step S2, the decoder is composed of a multi-scale residual module and deconvolution, and upsampling is achieved through 4 deconvolution operations to restore the image resolution and finally output the segmented image.
[0008] Among them, in step S2, the residual jump connection module is used to replace the original jump connection between the encoder and the decoder; the convolutional layer of the residual connection replaces the encoder and the decoder that simply connect the same layer.
[0009] Wherein, in step S3, the synthetic seismic data is used to automatically generate labeled data and training data sets.
[0010] Among them, in step S4, compared with the training of U-Net, all training conditions, training data, initial network parameters, and hardware facilities are the same.
[0011] In step S4, the batch size is set to 2, the optimizer used is RMSProp, the initial learning rate is set to 1e-3, and the weight decay coefficient is set to 1e-8; the convolutional neural network is trained with M groups of 3D synthetic seismic data and fault labels, and the verification data set contains N groups of different synthetic data and fault labels, which do not participate in the network training process.
[0012] In step S4, M is 300 and N is 30.
[0013] Among them, in step S4, 300 training data sets are input into the improved U-Net network in batches, and 4 training data sets are merged into earthquake and fault labeling data; after multiple iterations, when the loss function loss value gradually decreases to the target value, the improved U-Net is trained.
[0014] Wherein, in step S4, the target value is 0.01.
[0015] The method for automatically identifying near-surface faults in the embodiment of the present application has the following beneficial effects:
[0016] This application uses a multi-scale residual module to remove the two-layer convolution of the traditional U-Net, replaces the traditional U-Net jump connection with a residual jump connection, and introduces an attention mechanism. The improved U-Net network model can solve the semantic loss problem caused by the fusion of high- and low-level semantic information, and can extract image features from all aspects. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 This is a schematic diagram of the process flow of the method for automatically identifying near-surface faults according to an embodiment of the present application;
[0018] Figure 2 The architecture of U-Net;
[0019] Figure 3 The improved U-Net architecture of the embodiment of the present application;
[0020] Figure 4 The original three-dimensional seismic data volume and the fault probability volume of the embodiment of the present application;
[0021] Figure 5 Identification and comparison of the 2598th fault in the embodiment of the present application;
[0022] Figure 6 This is a comparison of fault identification in horizontal time slices of an embodiment of the present application. DETAILED DESCRIPTION
[0023] The present application is further described below in conjunction with the accompanying drawings and embodiments.
[0024] In the following introduction, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance. The following introduction provides multiple embodiments of the present invention, and different embodiments can be replaced or combined, so this application can also be considered to include all possible combinations of the same and / or different embodiments recorded. Therefore, if one embodiment includes features A, B, and C, and another embodiment includes features B and D, then this application should also be considered to include embodiments containing one or more of all other possible combinations of features A, B, C, and D, even though the embodiment may not be clearly recorded in the following text.
[0025] Embodiment 1
[0026] The commonly used U-Net has achieved satisfactory results in fault identification. However, the network has limited ability to extract and restore feature information and has no attention mechanism, which limits its ability to identify complex faults.
[0027] like Figure 1 As shown, the method for automatic identification of near-surface faults of the present application includes: S1, preparing post-stack seismic data; S2, building a neural network, adjusting parameters of each part, and the improved U-Net includes an encoder, a decoder, residual jump connections and an end-to-end image segmentation model; S3, constructing training data and producing labeled data; S4, training the neural network; S5, using the trained network to apply actual data.
[0028] The present application relates to an improved U-Net network for automatic fault identification, which enables the network to extract features of input images from all aspects to the greatest extent, effectively improves the training speed and efficiency of near-surface fault identification, and has better micro-fault identification capabilities.
[0029] Embodiment 2
[0030] In a specific embodiment, the method for automatically identifying near-surface faults of the present application includes:
[0031] Step S1 prepares post-stack seismic data;
[0032] Step S2 builds a neural network and adjusts the parameters of each part. Similar to the traditional U-Net structure, the improved U-Net includes an encoder, a decoder, residual skip connections and an end-to-end image segmentation model. Figure 2is the U-Net network structure, such as Figure 2 As shown, Figure 3 For the improved U-Net network structure, Figure 3 As shown, the encoder on the left of the improved U-Net is mainly composed of a multi-scale residual module (MSRB), an attention module (AM) and maximum pooling. The present invention uses MSRB to extract and restore features, and connects a multi-scale feature fusion attention module after each MSRB, so that the extracted feature information can be efficiently distributed, which is conducive to focusing on key areas and suppressing interference information in other areas. The decoder consists of MSRB and deconvolution, and upsampling is achieved through 4 deconvolution operations to restore the image resolution and finally output the segmented image. The residual jump connection module (RSCM) is used to replace the original jump connection between the encoder and the decoder; the convolution layer of the residual connection replaces the encoder and decoder that simply connect the same layer;
[0033] Step S3 constructs training data and produces label data. The training data can optimize the neural network parameters and gradually converge the prediction error. In order to avoid the difficulty of manually creating label data, the present invention uses synthetic seismic data to automatically generate label data and training data sets.
[0034] Step S4 performs neural network training, which is similar to the training of U-Net. All training conditions, training data, initial network parameters, and hardware facilities are the same. The batch size is set to 2, the optimizer used is RMSProp, the initial learning rate is set to 1e-3, and the weight decay coefficient is set to 1e-8. The convolutional neural network is trained using 300 sets of three-dimensional synthetic seismic data and fault labels, and the verification data set contains 30 different sets of synthetic data and fault labels, which do not participate in the network training process. The 300 training data sets are input into the improved U-Net network in batches, of which 4 training data sets are merged into earthquake and fault label data. After multiple iterations, when the loss function loss value gradually decreases to 0.01, the improved U-Net is trained.
[0035] Step S5 uses the trained network to apply actual data.
[0036] In this application, a multi-scale residual module is used to remove the two layers of convolution of the traditional U-Net, the residual jump connection is used to replace the traditional U-Net jump connection, and an attention mechanism is introduced; the improved U-Net network model can solve the semantic loss problem caused by the fusion of high- and low-level semantic information, and can extract image features from various aspects.
[0037] Embodiment 3
[0038] The two-layer convolution, skip connection and attention mechanism of the traditional U-Net network were transformed and improved, and a multi-scale residual module, residual skip connection and feature fusion attention module were introduced to enable the extracted feature information to be efficiently distributed, which is conducive to focusing on key areas and suppressing interference information in other areas.
[0039] The improved U-Net is tested using actual data. Figure 4 As shown, the actual data is located onshore China and belongs to sandstone reservoirs. The area has frequent tectonic activities, developed faults, and diverse morphologies. Due to the quality of seismic data and the experience of interpreters, the interpretation of deep faults in this area is quite large and difficult. The actual data includes 350 main survey lines (inline) and 450 contact survey lines (Xline). Each primary survey line contains 450 common depth points. The log-meaning CDP time interval is 25m. The sampling interval is 2ms and the seismic trace length is 3s. Comparison of fault identification results of traditional U-Net, improved U-Net and fault coherence body Figure 5 and Figure 6 Shown are: (a) original profile, (b) coherence volume, (c) U-Net, and (d) improved U-Net.
[0040] like Figure 5 As shown in (a) of the figure, there are both large faults from top to bottom and micro faults between layers, especially in the middle and deep parts where the fault zone is concentrated, which poses a great challenge to fault identification and sorting. The results obtained by using deep learning methods are as follows Figure 5 (c) and Figure 5 As shown in (d) in the figure. Figure 5 As can be seen from (b) in the figure, the fault recognition result is very clear, and there is no similar small layer illusion. However, the fault details recognized by the improved U-Net are as follows Figure 5 (d) shows that the fault details are richer than those recognized by the traditional U-Net. This is closely related to the improvement of the recognition and extraction capabilities of U-Net features.
[0041] like Figure 6 As shown in Figure 2, the plane distribution patterns of fault results identified by the three methods are shown. Figure 6 Compared with (b) in Figure 6 (c) and Figure 6 The spatial distribution characteristics of the faults shown in (d) are clearer, and the combination of major and minor faults on the plane is more reasonable. Figure 6 As shown in (d) in the figure, the faults obtained by the improved U-Net are more detailed, and the fault plane distribution is more vivid and natural, which reflects the adaptability and superiority of this method for complex hierarchical recognition.
[0042] The actual data application of the present invention shows that the improved U-Net network model has good micro-fault recognition ability, which is of great significance for the characterization of near-surface faults. It enables the network to extract the features of the input image from all aspects to the greatest extent, effectively improving the training speed and efficiency of near-surface fault recognition.
[0043] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A method for automatically identifying near-surface faults, characterized in that: include: S1, prepare post-stack seismic data; S2, build a neural network and adjust the parameters of each part. The improved U-Net includes encoder, decoder, residual skip connection and end-to-end image segmentation model; S3, build training data and produce label data; S4, conduct neural network training; S5, use the trained network to apply actual data.
2. The method for automatic identification of near-surface faults according to claim 1, characterized in that: In step S2, the encoder on the left of the improved U-Net includes a multi-scale residual module, an attention module and a maximum pooling; a multi-scale residual module is used to extract and restore features, and a multi-scale feature fusion attention module is connected after each multi-scale residual module.
3. The method for automatically identifying near-surface faults according to claim 1 or 2, characterized in that: In step S2, the decoder consists of a multi-scale residual module and deconvolution, which performs upsampling through four deconvolution operations, restores the image resolution, and finally outputs the segmented image.
4. The method for automatically identifying near-surface faults according to claim 1 or 2, characterized in that: In step S2, the residual jump connection module is used to replace the original jump connection between the encoder and the decoder; the convolutional layer of the residual connection replaces the encoder and the decoder that simply connect the same layer.
5. The method for automatically identifying near-surface faults according to claim 1 or 2, characterized in that: In step S3, labeled data and training data sets are automatically generated using synthetic seismic data.
6. The method for automatically identifying near-surface faults according to claim 1 or 2, characterized in that: In step S4, compared with the training of U-Net, all training conditions, training data, initial network parameters, and hardware facilities are the same.
7. The method for automatic identification of near-surface faults according to claim 6, characterized in that: In step S4, the batch size is set to 2, the optimizer used is RMSProp, the initial learning rate is set to 1e-3, and the weight decay coefficient is set to 1e-8; the convolutional neural network is trained with M groups of 3D synthetic seismic data and fault labels, and the verification data set contains N groups of different synthetic data and fault labels, which do not participate in the training process of the network.
8. The method for automatically identifying near-surface faults according to claim 7, characterized in that: In step S4, M is 300 and N is 30.
9. The method for automatically identifying near-surface faults according to claim 6, characterized in that: In step S4, 300 training data sets are input into the improved U-Net network in batches, of which 4 training data sets are merged into earthquake and fault labeling data; after multiple iterations, when the loss function loss value gradually decreases to the target value, the improved U-Net is trained.
10. The method for automatic identification of near-surface faults according to claim 9, characterized in that: In step S4, the target value is 0.01.
