Low signal-to-noise ratio seismic data first arrival automatic picking method based on feature fusion network
Patent Information
- Application Number
- CN202410119223.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-29
- Publication Date
- 2026-10-20
- Estimated Expiration
- 2044-01-29
AI Technical Summary
但随着地震数据量的剧增，地震资料品质的下降，初至拾取效率和准确度明显下降，传统的自动拾取方法对资料的处理速度和精度渐渐无法满足实际需求
[0029]第一、本发明方法原理具有可靠性。主要使用理论基础较为成熟、应用效果较好的LSTM网络和UNet网络。同时在UNet的基础上引入了注意力通道和残差模块，构建了Re-AG-UNet。基于初至具备的单道边界特征和二维边界特征，充分利用深度学习网络挖掘大数据特征的能力，将两个分网络提取的不同特征进行融合，整合两个网络的优点以提高初至拾取任务的预测精度，在原理上具有较高的可靠性。同常规初至自动拾取网络相比，其中的特征融合扩展模块将不同初至特征进行融合能够更好地挖掘初至特征，使网络收敛更好，预测精度更高。
Smart Images
Figure CN118033730B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of seismic data processing technology, specifically to a method for automatic first arrival picking of low signal-to-noise ratio seismic data based on feature fusion networks. Background Technology
[0002] First arrival picking plays a crucial role in earthquake source location and static correction. The efficiency and accuracy of first arrival picking for low signal-to-noise ratio seismic data are of particular concern.
[0003] Over the past few decades, numerous scholars have conducted extensive analysis and research on the problem of automatic first-arrival picking, proposing various solutions. Traditional first-arrival picking methods mainly start from the time and frequency domains of seismic data, studying the characteristics of first arrivals in different domains, and achieving automatic first-arrival picking through mathematical algorithms. Stevenson et al. first proposed the long-short time window average energy ratio method, which determines the first arrival of seismic data by calculating the ratio of the short-time window average energy value to the long-time window average energy value of the seismic signal on a single seismic trace, thereby completing the automatic first-arrival picking. Many scholars have made improvements based on this. Another commonly used method is the fractal method. Fabio et al. proposed the fractal dimension method, analyzing the fractal dimension of seismic traces and achieving automatic first-arrival picking based on the change of fractal dimension over time. Many scholars have also conducted extensive research based on this. In addition, many scholars have contributed to first-arrival picking based on wavelet transform, S-transform, and multiple time windows.
[0004] In recent years, machine learning and deep learning have been widely applied in earthquake processing and interpretation, providing automation in some tasks, including first-arrival picking. Yuan Sanyi et al. used CNNs to classify first-arrival and non-first-arrival signals, and then automatically picked the first-arrival signal. Tsai et al. proposed using deep semi-supervised neural networks to complete automatic first-arrival picking. This method can be trained with less sample data and has the advantage of continuous learning. Akram et al. proposed the k-meansCF algorithm and compared feature functions based on amplitude and energy. This method achieved good detection results on noisy data.
[0005] Currently, traditional automatic first-arrival acquisition methods each have their own focus, but they utilize very limited deep features of seismic signals. For high-quality seismic data, traditional methods offer acceptable computational efficiency and accuracy. However, with the dramatic increase in seismic data volume and the decline in data quality, the efficiency and accuracy of first-arrival acquisition have significantly decreased, and traditional automatic acquisition methods are gradually failing to meet practical needs in terms of data processing speed and accuracy. Although artificial intelligence technologies such as deep learning have to some extent solved the shortcomings of traditional seismic first-arrival acquisition algorithms in terms of poor applicability and low efficiency, most current deep learning methods use relatively simple neural networks, which cannot fully explore the deep features of first-arrival waves from multiple perspectives. For seismic data with low signal-to-noise ratio and large first-arrival variations, problems such as poor acquisition results and poor network generalization ability still exist.
[0006] Based on this, the present invention designs an automatic first arrival picking method for low signal-to-noise ratio seismic data based on feature fusion network to solve the above problems. Summary of the Invention
[0007] To improve the accuracy of automatic first-arrival picking for low signal-to-noise ratio (SNR) seismic data, this invention provides an automatic first-arrival picking method based on a feature fusion network. This method improves the accuracy and continuity of first-arrival picking for seismic data with low SNR and significant first-arrival variations. The method analyzes the single-trace and two-dimensional boundary features of the first-arrival wave. Based on these two features, LSTM and Re-AG-UNet networks are used to extract the first-arrival wave features and location from the seismic data, respectively. Then, the concept of multi-network learning is introduced into the first-arrival picking process, and a feature fusion network is designed. Finally, the multi-view first-arrival features extracted by the LSTM and Re-AG-UNet networks are fused, effectively improving the accuracy of network picking and making it suitable for large-scale data processing.
[0008] To achieve the above objectives, the present invention provides the following technical solution:
[0009] An automatic first arrival picking method for low signal-to-noise ratio seismic data based on feature fusion networks includes the following steps:
[0010] First, labels are created for training data using seismic data. Then, outliers are removed, and the seismic data is preprocessed to obtain initial arrival feature values. A large number of sample data pairs are constructed, which are divided into training sample data pairs, validation sample data pairs, and test sample data pairs.
[0011] 2. Construct an LSTM network for extracting initial and final channel features, and train it based on the obtained training sample data to extract initial and final channel boundary features;
[0012] Third, construct a Re-AG-UNet network incorporating attention channels and residual modules for extracting initial and final two-dimensional boundary features, and train the network based on the obtained training sample data to extract initial and final two-dimensional boundary features.
[0013] Fourth, design a feature fusion module, construct a feature fusion network, fuse the initial single-path boundary features extracted in step two and the initial two-dimensional boundary features extracted in step three, and save the trained feature fusion network model.
[0014] Fifth, based on the feature fusion network model trained in step four, a more accurate initial arrival position is obtained, thus completing the automatic initial arrival picking.
[0015] Furthermore, the sample data pairs are divided into training sample data pairs, validation sample data pairs, and test sample data pairs, respectively, according to 80%, 10%, and 10%.
[0016] Furthermore, a series of preprocessing steps are performed on the seismic data, including max-min normalization, automatic gain control, anomaly trace extraction, and first arrival characteristic value calculation.
[0017] Furthermore, the formula for calculating the eigenvalues of the first arrival wave is as follows:
[0018]
[0019]
[0020] In the formula, S p Let A be the amplitude intensity at that point, A be the sum of the energies of the n points preceding the current point, and B be the sum of the energies of the n points following the current point; S i Let be the initial arrival feature value of the i-th sample point, BA be the boundary monitoring factor, and B / A be the boundary reinforcement factor.
[0021] Furthermore, the output dimension of the LSTM network is consistent with that of the UNet network.
[0022] Furthermore, in step four, the feature fusion extension module mainly includes an RSU module for fusing multi-scale features.
[0023] Furthermore, during the training process, three loss values are generated, denoted as L1, L2, and L3, which correspond to the loss values of the LSTM network, the Re-AG-UNet network, and the feature fusion network, respectively.
[0024] Furthermore, L1 and L2 are used to monitor the training progress of sub-networks, while L3 is used for quality control of the entire feature fusion network; the total loss function is defined as follows:
[0025]
[0026] BCELoss=-(label*log(P)+(1-label)*log(1-P))
[0027] Where P is the predicted value and label is the true value.
[0028] Beneficial effects
[0029] First, the method of this invention is reliable in principle. It primarily utilizes the LSTM and UNet networks, which have mature theoretical foundations and good application results. Furthermore, an attention channel and a residual module are introduced on top of UNet to construct Re-AG-UNet. Based on the single-channel and two-dimensional boundary features of the initial arrival, it fully leverages the ability of deep learning networks to mine big data features, fusing the different features extracted by the two sub-networks. This integrates the advantages of both networks to improve the prediction accuracy of the initial arrival picking task, demonstrating high reliability in principle. Compared to conventional automatic initial arrival picking networks, the feature fusion extension module in this invention better mines initial arrival features by fusing different initial arrival features, resulting in better network convergence and higher prediction accuracy.
[0030] Secondly, this invention is simple to operate and highly efficient. Commercial software using existing technology can quickly generate a large number of labels for direct training by picking up the first arrival and then removing abnormal first arrivals. Furthermore, the input data is preprocessed to calculate the first arrival feature values, which has a certain noise suppression effect, highlighting the first arrival position and effectively saving network training time, thus helping to improve network training accuracy.
[0031] Third, the picking results of this invention are reliable. During training, the network fully learns the single-channel and two-dimensional boundary features of the first arrival. For the LSTM network, there are not many outliers with large deviations. For the Re-AG-UNet network, the picking results are more continuous and stable in regions with good signal-to-noise ratios, and it can also supplement the picking results of empty channels and bad channels. The feature fusion network effectively combines the advantages of the two sub-networks, and can successfully pick more accurate first arrival positions. Compared with other automatic first arrival picking algorithms, this algorithm has higher processing efficiency and more accurate picking results. Attached Figure Description
[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0033] Figure 1The flowchart shows the automatic first arrival picking method for low signal-to-noise ratio seismic data based on feature fusion network according to the present invention.
[0034] Figure 2 This invention constructs an LSTM network for extracting first-to-last single-channel features.
[0035] Figure 3 The Re-AG-UNet network constructed for extracting initial two-dimensional features in this invention;
[0036] Figure 4 The feature fusion network constructed for this invention;
[0037] Figure 5 This invention provides typical earthquake data for region A.
[0038] Figure 6 This is typical earthquake data for region B in this invention;
[0039] Figure 7 This is a comparison of the pickup effect in region A of the present invention;
[0040] Figure 8 This refers to the initial pickup results in region A of the present invention;
[0041] Figure 9 This is the LSTM pickup result for region A in this invention;
[0042] Figure 10 This is the initial arrival picking result of Re-AG-UNet in region A of this invention;
[0043] Figure 11 This is a comparison of the pickup effect in region B of the present invention;
[0044] Figure 12 This refers to the initial pickup results in region B of the present invention;
[0045] Figure 13 This is the LSTM pickup result for region B in this invention;
[0046] Figure 14 This is the initial arrival picking result of Re-AG-UNet in region B of the present invention. Detailed Implementation
[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0048] The present invention will be further described below with reference to embodiments.
[0049] Example 1
[0050] Please refer to the instruction manual appendix. Figure 1 An automatic first arrival picking method for low signal-to-noise ratio seismic data based on feature fusion networks includes the following steps:
[0051] First, input earthquake data, use commercial software in the existing technology to create labels for training data, then manually remove the picked outliers, and perform a series of preprocessing on the earthquake data to obtain initial arrival feature values, thereby constructing a large number of sample data pairs. The sample data pairs are divided into training sample data pairs, validation sample data pairs, and test sample data pairs according to 80%, 10%, and 10% respectively.
[0052] The process involves a series of preprocessing steps for the seismic data, including maximum and minimum normalization, automatic gain control (AGC), anomaly trace extraction, and first arrival characteristic value calculation.
[0053] First-arrival eigenvalues can enhance the first-arrival characteristics of seismic data, facilitating accurate extraction of these characteristics using networks. The formula for calculating the first-arrival eigenvalues is as follows:
[0054]
[0055]
[0056] In the formula, S p Let A be the amplitude intensity at that point, A be the sum of the energies of the n points preceding the current point, and B be the sum of the energies of the n points following the current point; S i Let S be the initial arrival eigenvalue of the i-th sample point. i The larger the value, the stronger the initial arrival feature of point i. The point corresponding to the maximum initial arrival feature value is the initial arrival position. BA is the boundary monitoring factor, which is sensitive to the boundary. Its value is larger at the energy boundary and smaller at non-boundary locations. B / A is the boundary strengthening factor, which strengthens the boundary and weakens the non-boundary.
[0057] II. Figure 2 A good LSTM network is constructed for extracting first-to-last-channel features. The network is then trained based on the obtained training sample data to extract first-to-last-channel boundary features.
[0058] III. Figure 3 Construct a Re-AG-UNet network incorporating attention channels and residual modules for extracting initial-to-two-dimensional boundary features, and train it based on the obtained training sample data to extract initial-to-two-dimensional boundary features;
[0059] The initial emission exhibits different features from different perspectives, requiring different deep learning networks for training and extraction. From a single-channel perspective, the initial emission has one-dimensional boundary features, which are extracted using a trained LSTM network. From a single-shot perspective, the initial emission has two-dimensional boundary features, which are extracted using a trained Re-At-UNet network.
[0060] IV. Figure 4 Design a feature fusion module, construct a feature fusion network, fuse the initial single-path boundary features extracted in step two and the initial two-dimensional boundary features extracted in step three, and save the trained feature fusion network model.
[0061] By introducing the idea of multi-network learning, a feature fusion network is designed, adding a feature fusion extension module (whose core is an RSU module) to fuse the initial arrival features extracted by the two networks. This can better mine the initial arrival features, improve prediction accuracy, and is suitable for large-scale data processing.
[0062] Fifth, based on the feature fusion network model trained in step four, a more accurate initial arrival position is obtained, thus completing the automatic initial arrival picking.
[0063] This invention improves the accuracy and continuity of automatic first-arrival acquisition for data with low signal-to-noise ratio and drastic first-arrival changes, and has good stability, making it suitable for industrial production in seismic exploration.
[0064] Example 2
[0065] Based on Example 1, the following solution is added:
[0066] To better capture first-arrival boundary features and avoid gradient vanishing, the UNet network used for training introduces attention channels and residual modules to accelerate network convergence.
[0067] For feature fusion to be performed, the output dimension of the LSTM network must be consistent with that of the UNet network. The number of features in the LSTM network is consistent with the input_size, and the input of each neuron is one seismic data point. If each single shot has 128 seismic records, then the input_size should be 128.
[0068] In step four, the feature fusion extension module mainly includes an RSU module. The core of the RSU block is the fusion of multi-scale features. During training, three loss values are generated, denoted as L1, L2, and L3, corresponding to the LSTM network loss, the Re-AG-UNet network loss, and the feature fusion network loss, respectively. L1 and L2 are used to monitor the training progress of the sub-networks, while L3 is used for quality control of the entire feature fusion network. The total loss function is defined as follows:
[0069]
[0070] BCELoss=-(label*log(P)+(1-label)*log(1-P))
[0071] Where P is the predicted value and label is the true value.
[0072] In step five, the initial arrival position is directly predicted using the pre-trained feature fusion network model. If the data has extremely low signal-to-noise ratio and the results are inaccurate, outliers are detected and removed, and then interpolation is performed to obtain the final result.
[0073] Example 3
[0074] To verify the effectiveness of the invention, automatic initial arrival picking was performed on field data from two different work areas.
[0075] The terrain of Area A is generally higher in the east and lower in the west, and higher in the north and south and lower in the middle. The Gobi Desert is relatively flat, while the mountainous areas have significant undulations. The vast majority of shot sites in Area A were triggered by controlled-source seismic activity, with a few triggered by well shots. The seismic data sampling interval was 2 ms. Figure 5 This is typical earthquake data for Area A.
[0076] Area B has a large overall loess layer thickness, with drastic variations in surface structure, rapid changes in loess layer thickness and lithology, and significant variations in first arrival. The different absorption and attenuation rates of loess layers of varying thicknesses directly result in a low signal-to-noise ratio for the seismic data. The seismic data sampling interval is 2 ms. Figure 6 Typical seismic data for area B are presented.
[0077] This invention was used to perform first arrival picking on two sets of data, and the picking results were analyzed. 50,000 seismic data points were extracted from each of the two work areas, A and B, and a total of 2,000 first arrival picking labels of size 128×128 were created. The sample data pairs were divided into training, validation, and test data pairs, with 80%, 10%, and 10% respectively. After training, the saved feature fusion network model was used to directly predict the first arrival positions.
[0078] Figure 7-10The graph shows the processing results of different networks on the actual data in region A. By comparing the marked areas in the graph, it is clear that... Figure 8 The processing results effectively integrated Figure 9-10 The correct initial position is obtained, and the fused result is more consistent with the actual annotation.
[0079] Figure 11-14 This displays the processing results of different networks on the actual data of area B, and compares the overall results with... Figure 7-10 Comparison revealed that the processing results of this invention were least affected by noise and showed the best consistency with the label results. This demonstrates that the feature fusion network possesses a certain degree of noise resistance. Simultaneously, comparisons were made... Figure 9-10 and Figure 13-14 It can be observed that the sub-networks are greatly affected by the signal-to-noise ratio (SNR). For LSTM networks, the continuity between channels depends on the short-term memory input of the previous neuron, thus avoiding many outliers with large deviations; however, the picking performance of channels with lower energy deteriorates. For Re-AG-UNet networks, the continuity of the recognition results is determined by 2D boundary features. Therefore, there may be discrete points with large deviations in areas where boundaries cannot be identified, but the picking results are more continuous and stable in areas with good SNR, and can also supplement the picking results of empty and bad channels. This invention effectively combines the advantages of the two sub-networks, successfully picking more accurate first-arrival positions. Experimental results show that using a trained model to pick single-shot records of actual seismic data, a single LSTM and Re-AG-UNet network can roughly extract accurate first-arrival positions. However, due to the influence of strong noise, the first-arrival extraction positions in noisy channels are abnormal. For areas with strong noise interference, the method proposed in this invention, after optimization, can effectively improve the accuracy of first-arrival picking. The first-arrival picking results of two different sets of data demonstrate that this invention has strong applicability and can be effectively used for actual data of different qualities.
[0080] Experiments using real data from regions A and B revealed that the automatic first arrival picking method for multi-view low signal-to-noise ratio seismic data based on feature fusion networks outperforms the picking performance of single LSTM networks and Re-AG-UNet networks. This improves the accuracy and continuity of first arrival picking, demonstrating the practicality and reliability of the method and its significant application potential.
[0081] This invention has the following advantages:
[0082] First, the methodology is reliable. It primarily utilizes the theoretically mature and practically effective LSTM and UNet networks. Furthermore, an attention channel and residual module are introduced on top of UNet to construct Re-AG-UNet. Based on the single-channel and two-dimensional boundary features inherent in the initial arrival, it fully leverages the ability of deep learning networks to mine features from large datasets, fusing the different features extracted by the two sub-networks. This integrates the advantages of both networks to improve the prediction accuracy of the initial arrival picking task, demonstrating high reliability in principle. Compared to conventional automatic initial arrival picking networks, the feature fusion extension module integrates different initial arrival features, resulting in better feature mining, better network convergence, and higher prediction accuracy.
[0083] Secondly, it is simple to operate and highly efficient. Existing commercial software can quickly generate a large number of labels for direct training by picking up the first arrival and then removing abnormal first arrivals. The input data is preprocessed to calculate the first arrival feature values, which has a certain noise suppression effect, highlights the first arrival position, effectively saves network training time, and helps to improve the accuracy of network training.
[0084] Third, the picking results are reliable. During training, the network fully learns the single-channel and two-dimensional boundary features of the first arrival. For the LSTM network, there are not many outliers with large deviations. For the Re-AG-UNet network, the picking results are more continuous and stable in regions with good signal-to-noise ratios, and it can also supplement the picking results of empty channels and bad channels. The feature fusion network effectively combines the advantages of the two sub-networks, and can successfully pick more accurate first arrival positions. Compared with other automatic first arrival picking algorithms, this algorithm has higher processing efficiency and more accurate picking results.
[0085] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for automatic first arrival picking of low signal-to-noise ratio seismic data based on feature fusion networks, characterized in that, Includes the following steps: First, labels are created for training data using seismic data. Then, outliers are removed, and the seismic data is preprocessed to obtain initial arrival feature values. A large number of sample data pairs are constructed, which are divided into training sample data pairs, validation sample data pairs, and test sample data pairs.
2. Construct an LSTM network for extracting initial and final channel features, and train it based on the obtained training sample data to extract initial and final channel boundary features; Third, construct a Re-AG-UNet network incorporating attention channels and residual modules for extracting initial and final two-dimensional boundary features, and train the network based on the obtained training sample data to extract initial and final two-dimensional boundary features. Fourth, design a feature fusion module, construct a feature fusion network, fuse the initial single-path boundary features extracted in step two and the initial two-dimensional boundary features extracted in step three, and save the trained feature fusion network model. Fifth, based on the feature fusion network model trained in step four, a more accurate initial arrival position is obtained, thus completing the automatic initial arrival picking.
2. The method for automatic first arrival picking of low signal-to-noise ratio seismic data based on feature fusion network according to claim 1, characterized in that, The sample data pairs were divided into training sample data pairs, validation sample data pairs, and test sample data pairs, with 80%, 10%, and 10% respectively.
3. The method for automatic first arrival picking of low signal-to-noise ratio seismic data based on feature fusion network according to claim 2, characterized in that, A series of preprocessing steps are performed on the seismic data, including maximum and minimum normalization, automatic gain control, outlier and empty trace removal, and first arrival characteristic value calculation.
4. The method for automatic first arrival picking of low signal-to-noise ratio seismic data based on feature fusion network according to claim 3, characterized in that, The formula for calculating the eigenvalues of the first arrival wave is as follows: ； ； In the formula, The amplitude intensity at that point, Before the current point The sum of energy at each point; After the current point The sum of energy at each point; For the first The initial arrival eigenvalues of each sample point For boundary monitoring factors, This is the boundary reinforcement factor.
5. The method for automatic first arrival picking of low signal-to-noise ratio seismic data based on feature fusion network according to claim 4, characterized in that, The dimensionality of the LSTM network output is the same as that of the Re-AG-UNet network output.
6. The method for automatic first arrival picking of low signal-to-noise ratio seismic data based on feature fusion network according to claim 5, characterized in that, In step four, the feature fusion module includes an RSU module for fusing multi-scale features.
7. The method for automatic first arrival picking of low signal-to-noise ratio seismic data based on feature fusion network according to claim 6, characterized in that, During training, three loss values are generated, denoted as... , and , respectively, correspond to the loss values of the LSTM network, the Re-AG-UNet network, and the feature fusion network.
8. The method for automatic first arrival picking of low signal-to-noise ratio seismic data based on feature fusion network according to claim 7, characterized in that, , Used to monitor the training progress of subnetworks. Used for quality control of the entire feature fusion network; the total loss function is defined as follows: ； ； in, It is a predicted value. It is the actual value.