A method and related device for intelligently picking first arrival waves of seismic data

Abstract

The present invention discloses a method and related device for intelligent first arrival wave picking of seismic data. The method comprises: obtaining a linearly corrected seismic data set and a label data set; selecting a time window including the first arrival times of all seismic data; using the label data set to train a picking model and a binary classification model respectively; using the trained picking model to pick first arrivals to obtain a first first arrival set; using the trained binary classification model to classify the first arrivals in the first first arrival set to obtain reliable first arrivals and unreliable first arrivals; determining unstable first arrivals contained in the first first arrival set; eliminating first arrivals that belong to both the unreliable first arrivals and the unstable first arrivals; eliminating first arrivals whose fitting errors are greater than a preset error threshold; performing inverse linear correction processing to obtain first arrival wave intelligent picking results. The method can improve the accuracy and efficiency of first arrival wave picking, reduce the workload of manually modifying first arrivals, and help meet the requirements of first arrival picking of massive data.

Description

Technical Field

[0001] The present invention relates to the field of geophysical exploration seismic data processing, and in particular to a seismic data first arrival wave intelligent picking method and related devices. Background Art

[0002] The quality of first-arrival wave picking is related to the accuracy of near-surface model establishment, static correction effect, and migration imaging quality. Therefore, efficient and accurate first-arrival wave picking is very important in seismic data processing. With the continuous development of seismic exploration technology, the exploration area has gradually penetrated into mountainous areas, Gobi deserts, and hilly areas. Affected by surface conditions and different external environments, the first-arrival wave background noise of seismic data collected in the field is strong and the first-arrival wave information is relatively weak. Conventional first-arrival wave picking methods based on the instantaneous characteristics of seismic records include extreme value method, maximum amplitude method, energy ratio method, etc. Conventional first-arrival wave picking methods based on the overall characteristics of seismic records include correlation method, statistical method, fractal dimension algorithm, etc. Summary of the Invention

[0003] The inventors have found that the conventional first-arrival wave picking methods in the prior art, which are based on the instantaneous characteristics of seismic records, are difficult to accurately pick up the first-arrival waves when the noise seriously interferes with the first-arrival waves. The conventional first-arrival wave picking methods based on the overall characteristics of seismic records are affected by factors such as the similarity between seismic traces. For complex surface seismic data, the accuracy of first-arrival picking will also be affected. Therefore, for the picking of first-arrival waves of complex seismic data, the conventional first-arrival picking methods cannot well solve the first-arrival picking problems encountered in actual production. The first-arrival picking of these seismic data still relies on manual intervention to a large extent, especially for the first-arrival picking of weak energy, strong background noise interference and massive data. The workload of manual modification is very large. In order to at least partially solve the technical problems existing in the prior art, the inventors have made the present invention, and through specific implementation methods, provide a seismic data first-arrival wave intelligent picking method and related devices.

[0004] In a first aspect, an embodiment of the present invention provides a method for intelligently picking first arrival waves from seismic data, comprising:

[0005] Collecting a plurality of label data from a plurality of seismic data, and performing linear correction on the plurality of seismic data and the plurality of label data, respectively, to obtain a linearly corrected seismic data set and a label data set including the linearly corrected plurality of label data;

[0006] Based on the linearly corrected earthquake dataset and the label dataset, a time window including the first arrival time of all earthquake data is selected;

[0007] Within the time window, using the label dataset, respectively training the picking model and the binary classification model;

[0008] Within the time window, first arrival picking is performed on the linearly corrected seismic data set using the trained picking model to obtain a first first arrival set;

[0009] Using the trained binary classification model, the first arrivals in the first arrival set are classified into reliable first arrivals and unreliable first arrivals;

[0010] Cluster the first arrivals in the first arrival set to identify the unstable first arrivals contained therein;

[0011] Eliminate the first arrivals that belong to both the unreliable first arrivals and the unstable first arrivals from the first first arrival set to obtain a second first arrival set;

[0012] Fitting the first arrival times of the second first arrival set, eliminating first arrivals with fitting errors greater than a preset error threshold, to obtain a third first arrival set;

[0013] An inverse linear correction process is performed on the third first arrival set to obtain a first arrival wave intelligent picking result.

[0014] In some optional embodiments, collecting a plurality of label data from a plurality of seismic data includes:

[0015] Performing preliminary processing on a plurality of seismic data; the preliminary processing includes processing the seismic data by an observation system;

[0016] Label data is collected from the plurality of seismic data after preliminary processing.

[0017] In some optional embodiments, performing linear correction on the plurality of seismic data and the plurality of label data respectively includes:

[0018] Performing linear velocity analysis on the collected tag data including the first arrival time to obtain the first arrival apparent velocity of each tag data including the first arrival time;

[0019] Using each first-arrival apparent velocity, interpolating and extrapolating the plurality of seismic data, to obtain the first-arrival apparent velocity of each seismic trace in the plurality of seismic data;

[0020] Performing linear correction processing on the plurality of seismic data and label data using the first arrival apparent velocity of each seismic trace of the plurality of seismic data;

[0021] The first arrival apparent velocity of each seismic trace is used to calculate the corresponding correction time difference of each seismic trace, and the correction time difference is used to perform corresponding linear correction processing on the first arrival time of the collected tag data.

[0022] In some optional embodiments, using each first-arrival apparent velocity to interpolate and extrapolate the plurality of seismic data to obtain the first-arrival apparent velocity of each seismic trace in the plurality of seismic data includes:

[0023] Using a surface interpolation method, interpolating the first arrival apparent velocity of the tag data that does not include the first arrival time in the collected tag data to obtain the first arrival apparent velocity of the tag data that does not include the first arrival time;

[0024] The first arrival apparent velocity of the seismic trace adjacent to the seismic trace with unknown first arrival apparent velocity in the plurality of seismic data is used as the first arrival apparent velocity of the seismic trace with unknown first arrival apparent velocity.

[0025] In some optional embodiments, the preliminary processing of seismic data further includes:

[0026] At least one of noise attenuation, filtering, and static correction processing is performed on the seismic data.

[0027] In some optional embodiments, when the preliminary processing of seismic data includes the static correction processing, the method further includes:

[0028] An anti-static time difference process is performed on the third first arrival set.

[0029] In some optional embodiments, selecting a time window including the first arrival times of all seismic data based on the linearly corrected seismic dataset and the label dataset includes:

[0030] Select multiple time windows based on the linearly corrected earthquake dataset and the label dataset;

[0031] The first arrival times of the seismic data in the multiple time windows are interpolated using the first arrival times in the linearly corrected label data set.

[0032] In some optional embodiments, within the time window, using the label dataset to train the picking model and the binary classification model respectively includes:

[0033] Marking the label data of the seismic traces with first arrival time in the label data set as the first type of label data, and marking the label data of the seismic traces without first arrival time in the label data set as the second type of label data;

[0034] Use the first type of labeled data to train the picking model;

[0035] Use the first-class labeled data and the second-class labeled data to train a binary classification model.

[0036] In some optional embodiments, clustering the first arrivals in the first arrival set to determine unstable first arrivals included therein includes:

[0037] Calculating the first arrival time difference between adjacent seismic traces of each first arrival in the first first arrival set, performing clustering filtering on the first arrival time difference between adjacent seismic traces of each first arrival, and obtaining at least two first arrival clusters classified according to the first arrival time difference between adjacent seismic traces;

[0038] calibrating at least one first arrival cluster according to the order of change values ​​of the first arrival time differences of the first arrival clusters from large to small;

[0039] The first arrivals in the marked first arrival cluster are determined to be unstable first arrivals.

[0040] In some optional embodiments, performing fitting processing on the first arrival times of the second first arrival set and eliminating first arrivals with fitting errors greater than a preset error threshold to obtain a third first arrival set includes:

[0041] The first arrival time of the second first arrival set is fitted using the following polynomial:

[0042] p(x)=p1x n +p2x n-1 +...+p n x+p n+1

[0043] Where p is the fitting coefficient, n+1 is the number of fitting first arrivals, and x is the first arrival time;

[0044] The first arrivals whose fitting errors are greater than a preset error threshold are eliminated from the second first arrival set to obtain a third first arrival set.

[0045] In a second aspect, an embodiment of the present invention provides a device for intelligently picking first arrival waves of seismic data, comprising:

[0046] The label data set construction module 101 is used to collect a plurality of label data from a plurality of seismic data, and perform linear correction on the plurality of seismic data and the plurality of label data to obtain a linearly corrected seismic data set and a label data set containing the linearly corrected label data;

[0047] The model training module 102 is configured to select a time window including the first arrival times of all seismic data based on the linearly corrected seismic data set and the label data set; and to train the picking model and the binary classification model respectively within the time window using the label data set;

[0048] The first arrival wave picking module 103 is used to pick the first arrival of the linearly corrected seismic data set within the time window using the trained picking model to obtain a first first arrival set; use the trained binary classification model to classify the first arrivals in the first first arrival set to obtain reliable first arrivals and unreliable first arrivals; cluster the first arrivals in the first first arrival set to determine the unstable first arrivals contained therein; eliminate the first arrivals that belong to the unreliable first arrivals and the unstable first arrivals from the first first arrival set to obtain a second first arrival set; fit the first arrival time of the second first arrival set to eliminate the first arrivals whose fitting error is greater than a preset error threshold value to obtain a third first arrival set; perform anti-linear correction on the third first arrival set to obtain a first arrival wave intelligent picking result.

[0049] In a third aspect, an embodiment of the present invention provides a computer storage medium, wherein the computer storage medium stores computer executable instructions, and when the computer executable instructions are executed by a processor, the aforementioned method for intelligently picking first arrival waves of seismic data is implemented.

[0050] In a fourth aspect, an embodiment of the present invention provides a terminal device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the aforementioned method for intelligently picking up the first arrival wave of seismic data when executing the program.

[0051] The beneficial effects of the above technical solutions provided by the embodiments of the present invention include at least:

[0052] The embodiment of the present invention provides a method and related device for intelligent picking of first arrival waves of seismic data. Through linear correction, it is helpful to reduce noise interference, and can provide higher quality training data for model training, which helps to improve the model training effect and improve the accuracy of first arrival picking and first arrival classification using picking models and binary classification models. By comparing unreliable first arrivals obtained by the binary classification model with unstable first arrivals obtained by clustering, the accuracy of eliminating first arrivals with low quality is improved, and unreliable first arrivals can be effectively deleted, which helps to improve the accuracy of first arrival picking. By training an artificial intelligence model for first arrival picking, the accuracy and efficiency of first arrival wave picking can be improved, the seismic data processing cycle can be shortened, the workload of manual modification of first arrivals can be reduced, and it helps to adapt to the requirements of first arrival picking of massive data and improve the imaging quality of seismic data.

[0053] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purposes and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings.

[0054] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0056] Figure 1 This is a flow chart of a method for intelligently picking first arrival waves of seismic data in an embodiment of the present invention;

[0057] Figure 2 A schematic diagram of a picking model in an embodiment of the present invention;

[0058] Figure 3 Schematic diagram of a binary classification model in an embodiment of the present invention;

[0059] Figure 4 This is a plan view of the actual shot inspection distribution in the seismic data exploration area in an embodiment of the present invention;

[0060] Figure 5 This is a plan view of the label data inspection distribution in an embodiment of the present invention;

[0061] Figure 6a This is the image before the label data is manually picked up in the embodiment of the present invention;

[0062] Figure 6b This is an image after the label data is first manually picked up in an embodiment of the present invention;

[0063] Figure 7 This is a diagram showing the effect of linear correction of seismic data in an exploration area according to an embodiment of the present invention;

[0064] Figure 8 This is a schematic diagram of selecting a time window after linear correction in an embodiment of the present invention;

[0065] Figure 9 This is the first arrival graph of the first arrival set in an embodiment of the present invention;

[0066] Figure 10 This is the first arrival graph of the second first arrival set in an embodiment of the present invention;

[0067] Figure 11 This is the first arrival graph of the third first arrival set in an embodiment of the present invention;

[0068] Figure 12 This is a flowchart of a specific implementation process of a method for intelligently picking first arrival waves of seismic data in an embodiment of the present invention;

[0069] Figure 13 This is a block diagram of an intelligent device for picking up first arrival waves of seismic data in an embodiment of the present invention. DETAILED DESCRIPTION

[0070] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0071] In order to solve the problems existing in the prior art, an embodiment of the present invention provides a method and related device for intelligently picking first arrival waves of seismic data.

[0072] In the following embodiments, first arrival refers to the first wave. The first wave arriving on a seismic record is called the first arrival wave. The first arrival time is the time it takes for the signal to propagate from the shot point to the receiver point. Determining the first arrival time on a seismic trace allows us to determine the first arrival wave on that trace. Conversely, determining the first arrival wave on a seismic trace allows us to determine the first arrival time on that trace.

[0073] Example 1

[0074] The first embodiment of the present invention provides a method for intelligently picking first arrival waves of seismic data, the process of which is as follows: Figure 1 As shown, the following steps are included:

[0075] Step S101: Collect multiple label data from multiple seismic data, perform linear correction on the multiple seismic data and the multiple label data respectively, to obtain a linearly corrected seismic data set and a label data set containing the multiple label data after the linear correction; based on the linearly corrected seismic data set and the label data set, select a time window including the first arrival time of all seismic data.

[0076] Perform preliminary processing on seismic data; select a 3D seismic exploration area and collect seismic data for the entire area.

[0077] In some optional embodiments, an observation system is set up for seismic data of the entire work area.

[0078] In some optional embodiments, at least one of noise attenuation, filtering, and static correction is performed on the seismic data. Noise attenuation primarily involves performing strong energy suppression on individual or local seismic traces with strong background noise, increasing the energy intensity of the first arrival relative to the background noise and facilitating first arrival pickup. Filtering primarily involves picking up first arrivals from seismic data where there is a large frequency difference between the primary frequency of the background noise and the primary frequency of the first arrival. By filtering, the interference of background noise on the first arrival is reduced, facilitating first arrival pickup. Filtering primarily involves processing data where the primary frequency difference between the noise and the primary arrival is relatively large. Static correction is performed to reduce the impact of large elevation changes or large changes in the near-surface low velocity zone. The raw seismic data must undergo terrain correction, excitation depth correction, and low velocity zone correction. Noise attenuation, filtering, and static correction need to be performed selectively based on the collected seismic data; not every seismic data requires these processes.

[0079] Select label data from the seismic data after preliminary processing; select label data from the seismic data based on the characteristics of the plane distribution map of the exploration area's shot-inspection relationship and with reference to the elevation map of the work area, select more label data from the seismic data portion with large elevation changes, and select more label data from other key areas of concern, thereby collecting more first-arrival data from the above-mentioned areas, which will help the learning model extract more accurate features.

[0080] Collect the label data with the first arrival time. The selected label data is obtained by combining conventional first arrival picking and manual picking to obtain the first arrival time of the label data, and is manually modified to eliminate unreliable first arrival label data and retain reliable first arrival label data. Among them, whether the first arrival is reliable can be determined by a variety of rules. For example, a first arrival with an unclear or discontinuous waveform can be determined as an unreliable first arrival. The label data with the first arrival time picked up by the above method has high accuracy, can provide higher quality training data for model training, improve the training effect of the model, and help the model extract more accurate features, thereby helping to improve the accuracy of the application model for picking or classification.

[0081] Linear velocity analysis is performed on the first arrival time of the tag data to obtain the apparent velocity of each first arrival of the tag data. Apparent velocity is the propagation distance divided by the propagation time. In practice, it can be obtained by dividing the offset by the first arrival time. Vn = Sn / Tn, where Sn is the offset and Tn is the time it takes for the first arrival wave to propagate from the shot point to the receiving point. Vn is the apparent velocity of the corresponding first arrival. Sn and Tn are obtained using existing techniques.

[0082] The seismic data is interpolated and extrapolated using the apparent velocity of the first arrival of the label data to obtain the apparent velocity of the first arrival of each seismic trace in the seismic data; interpolation refers to using the known apparent velocity of the label data to estimate the apparent velocity of other seismic data with unknown apparent velocity through a certain mathematical method. In some optional embodiments, a surface interpolation method is used to interpolate the apparent velocity of the first arrival of the label data that does not contain the first arrival time in the collected label data to obtain the apparent velocity of the first arrival of the label data that does not contain the first arrival time; the apparent velocity of the seismic trace adjacent to the seismic trace with unknown first arrival apparent velocity in the multiple seismic data is used as the apparent velocity of the first arrival of the seismic trace with unknown first arrival apparent velocity. Among them, the surface interpolation method is a conventional method in the prior art. Interpolation is an important method for approximating discrete functions. It can be used to estimate the value of the function at other points based on the value of the function at a finite number of points.

[0083] Using the apparent velocity of the first arrival of each seismic trace in the seismic data, linear correction is performed on both the seismic data and the label data. The purpose of linear correction is to eliminate the difference in first arrival propagation time of seismic traces caused by offset, flatten the first arrival waves, and reduce the first arrival time difference between seismic traces. After linear correction of first arrivals, the peaks of the first arrivals ideally lie on the same horizontal line, which facilitates multi-trace deep learning and multi-trace first arrival picking in this embodiment.

[0084] The apparent velocity of each seismic trace's first arrival is used to calculate the corresponding corrected time difference for each seismic trace. This corrected time difference is used to perform a linear correction on the first arrival time of the tag data. The corrected time difference Tn1 = Sn / Vp, where Vp is the interpolated apparent velocity, Tn1 is the time difference to be corrected for each seismic trace, and Sn is the offset. This corrected time difference is used to perform a linear correction on the first arrival time of the tag data, ensuring that the first arrival time of the tag data is consistent with the first arrival of the tag data.

[0085] Linear correction can help reduce noise interference and provide higher-quality training data for model training, which helps improve the model training effect and the accuracy of first-arrival picking and first-arrival classification using the picking model and the binary classification model.

[0086] Based on the linearly corrected seismic data and label data, a time window is selected that includes the first arrival times of all seismic data, including:

[0087] Select multiple time windows based on the linearly corrected seismic data and label data;

[0088] The seismic data in the multiple time windows are interpolated using the first arrival times of the linearly corrected label data. However, the first arrival time interpolation is not performed on the label data in the multiple time windows.

[0089] Step S102: Within the time window, use the label data set to train the picking model and the binary classification model respectively. Label data with first arrival time in the label data set are marked as first-class label data, and label data without first arrival time in the label data set are marked as second-class label data. Use the first-class label data to train the picking model. Figure 2 As shown in the figure, a fully convolutional neural network model including a bidirectional long short-term memory (BILSTM) network is used as the picking model. Conventional neural networks only extract single-channel earthquake first arrival features, while the main function of BILSTM is to perform multi-channel cross-correlation and mine relevant information of multiple seismic waves. The fully convolutional neural network and BILSTM respectively extract the single-channel and multi-channel features of the first arrival as the input of the fully connected network. Figure 2 The function of the fully connected layer is to perform nonlinear fitting on the extracted seismic wave features and normalize the output shape size to the target shape size. The picking model is trained using the first type of label data.

[0090] Use the first and second label data to train a binary classification model. Figure 3 As shown in the figure, a fully convolutional neural network model including a bidirectional long short-term memory network (BILSTM) is used as a binary classification model. Conventional neural networks only extract single-channel earthquake first arrival features, while the main function of BILSTM is to perform multi-channel cross-correlation and mine relevant information of multiple seismic waves. The fully convolutional neural network and BILSTM extract single-channel and multi-channel features of the first arrival respectively. As the input of the fully connected network, Figure 3 The fully connected layer performs nonlinear fitting on the extracted seismic wave features, reducing the output shape size to the target shape size. The binary classification model is trained using the first and second labeled data to obtain a binary classification model. By inputting the first and second labeled data into the binary classification model for training, appropriate weights for classifying the first and second data are trained.

[0091] Step S103: Within the time window, first arrivals are picked from the linearly corrected seismic data set using the trained picking model to obtain a first first arrival set. For example, first arrivals are picked from the entire work area data after linear correction based on the obtained picking model, and the picked first arrivals constitute the first first arrival set. By using a trained artificial intelligence model, i.e., a picking model, to pick first arrivals from seismic data, the efficiency of first arrival picking can be improved, the manual picking workload can be reduced, the requirements for first arrival picking of massive data sets can be met, and the seismic data processing cycle can be shortened.

[0092] Step S104: Use the trained binary classification model to classify the first arrivals in the first arrival set to obtain reliable first arrivals and unreliable first arrivals. The neural network classification problem is mainly based on the perceptual characteristics of the human brain and uses the threshold unit as the transfer function, so it can only output two values. Assuming that P is the first arrival input vector, W is the weight matrix quantity, and b is the threshold vector, then the decision boundary of the perceptron is wp+b. When wp+b>0, it is determined to be a first-class reliable first arrival, otherwise it is a second-class unreliable first arrival. By using an artificial intelligence model, i.e., a binary classification model, to further screen the picked first arrivals for reliable ones, the accuracy of picking the first arrivals can be improved.

[0093] Step S105: clustering the first arrivals in the first arrival set to determine the unstable first arrivals contained therein.

[0094] The first arrival time differences between adjacent seismic traces in the first arrival set are calculated and clustering filtering is performed using these values. For example, the K-means algorithm can be selected for clustering. The K-means algorithm is an unsupervised machine learning method that continuously searches for the nearest mean value to a seed point until the error converges. It is a typical distance-based clustering algorithm, using distance as the evaluation criterion for similarity. Based on the principle that the closer the distance between the first arrival times, the greater the similarity, the K-means algorithm considers clusters to be composed of first arrival time differences between adjacent traces with close distances. Therefore, the ultimate goal is to obtain independent clusters with similar distances. Therefore, clustering in this embodiment is the process of dividing the set of adjacent trace first arrival time differences into similar object classes, so that the first arrival time differences between adjacent traces within the same cluster (or class) have high similarity, while the first arrival time differences between adjacent traces in different clusters have high dissimilarity. This allows the identification of unstable first arrivals, where the first arrival time differences between adjacent seismic traces vary significantly. By identifying unstable first arrivals from a clustering perspective, the quality of the picked first arrivals can be further improved.

[0095] Step S106: First arrivals that are both unreliable and unstable are eliminated from the first arrival set to obtain a second first arrival set. The arrival times of the second first arrival set are fitted, and arrivals with fitting errors greater than a preset error threshold are eliminated to obtain a third first arrival set. The third first arrival set is subjected to inverse linear correction to obtain the first arrival wave intelligent picking result.

[0096] By combining the two different perspectives of binary classification model and clustering, low-quality first arrivals are eliminated, thereby improving the accuracy of eliminating low-quality first arrivals and helping to improve the accuracy of first arrival picking in this embodiment.

[0097] For example, a 3D seismic exploration area is selected, such as Figure 4 The vertical lines in the diagram are detection lines, which represent the arrangement of 3D seismic data. Figure 4 The horizontally distributed points in the middle represent shot points. There are 498 shot records in the work area, with 36 shots per shot. First, the entire exploration seismic data is processed by field static correction (static correction data is collected in the field), the seismic data is corrected to the work area reference plane, and then noise attenuation processing is performed. The 498 shot seismic data after preliminary processing are processed in Figure 4 Select label data in the exploration area blast inspection distribution plan. The label data selection results are shown in Figure 5 The vertical lines in the figure are detection lines, which represent the arrangement of 3D seismic data. The horizontal points in the figure represent shot points. A total of 25 shot data were selected, and the 25 shot data were basically evenly distributed throughout the exploration area. The selected labeled data were manually picked for the first arrival, mainly picking the more reliable first arrivals. Unreliable first arrivals were not picked, and the first arrivals with unclear or discontinuous waveforms were judged as unreliable first arrivals. Figure 6a and Figure 6b , Figure 6a and Figure 6b The vertical axis represents time, and the horizontal axis represents position. Figure 6a This is the image before the label data is manually picked up. Figure 6b The image after the label data is first picked up manually, Figure 6b The clear and continuous waveform on the upper edge of the peak-like figure is determined to be a reliable first arrival and is depicted with a thicker line. The fuzzy and discontinuous waveform on the upper edge of the peak-like figure is determined to be an unreliable first arrival. By marking the first arrival, the label data with the first arrival time is obtained. Then, a linear velocity analysis is performed based on the first arrival time of the 25-shot label data, and the apparent velocity of each first arrival of the label data is calculated. Vn = shot spacing / first arrival time. The first arrival time is the time it takes to propagate from the shot point to the detection point. The apparent velocity of the first arrival of the label data is used to interpolate and extrapolate the seismic data of the entire exploration area to obtain the apparent velocity of the first arrival wave of each seismic channel in the entire work area. The interpolated apparent velocity is used to perform linear correction processing on the first arrival time of the 498-shot data, the 25-shot label data, and the 25-shot data (correction time difference Tn1 = shot offset / apparent velocity). The correction results are shown as follows: Figure 7 The linear correction effect diagram of seismic data in the exploration area is shown in the figure. Figure 7 The vertical axis represents time, and the horizontal axis represents position. Figure 7 Display the linear correction result of the 420th shot in the work area. After linear correction, select an appropriate first arrival time window and try to ensure that the first arrival of the entire exploration area is within the selected time window, such as Figure 8As shown in the figure, the vertical axis represents time, the horizontal axis represents position, and the time window is represented by a rectangular frame enclosed by a thick line. Based on whether the seismic traces of the labeled data have a first arrival time, the labeled data after the selected time window of 25 shots are subjected to a first arrival binary classification process. Seismic traces with first arrival times are calibrated as first-category labeled data, and seismic traces without first arrival times are calibrated as second-category labeled data. The 25-shot data and the first arrivals within the given time window are tested and trained on a full convolutional neural network model for multiple seismic traces. In this embodiment, five seismic traces are selected to obtain a picking model.

[0098] The full convolutional neural network model for multiple seismic traces is a full convolutional neural network model including a bidirectional long short-term memory network BILSTM. The structure of the model is shown in Figure 2 In the figure, the seismic data of multiple seismic channels are input into the fully convolutional neural network. BILSTM extracts the relevant information of the multi-channel seismic waves, and the fully convolutional neural network and BILSTM extract the features of the first arrival single channel and multiple channels respectively as the input of the fully connected network. Figure 2 The function of the fully connected layer is to perform nonlinear fitting on the extracted seismic wave features and normalize the output shape size to the target shape size. The picking model is trained with the first type of label data. The full convolutional neural network binary classification model is tested and trained on the first type of label data and the second type of label data. The neural network structure is shown in Figure 3 The first arrival set was obtained by using the first arrival picking training model to pick the first arrival of the 498-shot seismic data in the entire work area after linear correction. The first arrival picking results are shown in Figure 9 Then, the data is classified into two categories according to the binary classification model, and the second category of unreliable first arrival seismic traces is calibrated. In an arrangement of the three-dimensional exploration shot set, the first arrival time difference between seismic traces is calculated, and the first arrival time difference clustering filtering is performed using the first arrival time difference to calibrate the first arrival seismic traces with large and unstable first arrival time differences between seismic traces. The calibrated seismic traces that happen to belong to the second category of unreliable first arrival seismic traces are deleted to obtain the second first arrival set. The first arrival picking effect after deletion is shown in Fig. Figure 10 Then, fitting is performed on the remaining first arrival times after removing the first arrival times. In this embodiment, the fitting error threshold is set to 10 milliseconds, and the number of fitting channels is 30 adjacent seismic channels. After removing the first arrival times with fitting errors greater than 10 milliseconds, the third first arrival set is obtained. The first arrival picking effect of the third first arrival set is shown in FIG. Figure 11 Finally, the first arrival times of the third first arrival set are subjected to inverse linear time difference correction and inverse static time difference correction, until all arrangements of 498 guns have completed all steps of first arrival picking in this embodiment, and the first arrival wave intelligent picking results are obtained.

[0099] The intelligent picking method for first arrival waves of seismic data provided in this embodiment helps to reduce noise interference through linear correction, can provide higher quality training data for model training, helps to improve the model training effect, and improves the accuracy of first arrival picking and first arrival classification using picking models and binary classification models. By comparing unreliable first arrivals obtained by the binary classification model with unstable first arrivals obtained by clustering, the accuracy of eliminating first arrivals with low quality is improved, and unreliable first arrivals can be effectively deleted, which helps to improve the accuracy of first arrival picking. By training an artificial intelligence model for first arrival picking, the accuracy and efficiency of first arrival wave picking can be improved, the seismic data processing cycle can be shortened, the workload of manual modification of first arrivals can be reduced, it helps to adapt to the requirements of first arrival picking of massive data, and improve the imaging quality of seismic data.

[0100] Example 2

[0101] The second embodiment of the present invention provides a specific implementation process of a method for intelligently picking first arrival waves of seismic data, and the process is as follows: Figure 12 As shown, the following steps are included:

[0102] Step S201: Collect multiple label data from multiple seismic data, perform linear correction on the multiple seismic data and the multiple label data respectively, to obtain a linearly corrected seismic data set and a label data set containing the multiple label data after the linear correction; based on the linearly corrected seismic data set and the label data set, select a time window including the first arrival time of all seismic data.

[0103] Preliminary processing is performed on a plurality of seismic data; the preliminary processing includes processing the seismic data using an observation system; a three-dimensional seismic exploration area is selected, and seismic data of the entire area is collected.

[0104] An observation system is set up for the seismic data of the entire work area.

[0105] In some optional embodiments, the seismic data undergoes at least one of noise attenuation, filtering, and static correction. Noise attenuation primarily involves applying strong energy suppression to individual or local seismic traces with strong background noise, increasing the energy intensity of the first arrival relative to the background noise and facilitating first arrival pickup. Filtering primarily involves picking up first arrivals from seismic data where there is a significant frequency difference between the primary frequency of the background noise and the primary frequency of the first arrivals. By filtering, background noise interference on the first arrivals is reduced, facilitating first arrival pickup. Filtering primarily involves processing data with a significant frequency difference between the primary frequency of the noise and the primary arrivals. Static correction is performed to reduce the impact of large elevation changes or large changes in the near-surface low velocity zone. The raw seismic data undergoes terrain correction, excitation depth correction, and low velocity zone correction. Noise attenuation, filtering, and static correction are selectively performed based on the collected seismic data; not all seismic data require these processes. However, when the initial processing of the seismic data includes static correction, it also includes performing inverse static correction on the third first arrival set.

[0106] Collect label data from the multiple seismic data after preliminary processing; select label data from the seismic data based on the characteristics of the plane distribution map of the exploration area's blasting and inspection relationship, and refer to the elevation map of the work area. Select more label data from the seismic data part with large elevation changes. You can also select more label data in other key areas, so as to collect more first arrival data in the above-mentioned areas, which will help the training model extract more accurate features.

[0107] The label data with the first arrival time included in the collected label data are subjected to linear velocity analysis to obtain the first arrival apparent velocity of each label data with the first arrival time. The collected label data are obtained by combining conventional first arrival picking and manual picking to obtain the first arrival time of the label data, and are manually modified to eliminate unreliable first arrival label data and retain reliable first arrival label data. Among them, whether the first arrival is reliable can be determined by a variety of rules. For example, a first arrival with an unclear or discontinuous waveform can be determined as an unreliable first arrival. The label data with the first arrival time picked up by the above method has high accuracy, can provide higher quality training data for model training, improve the training effect of the model, and help the model extract more accurate features, thereby helping to improve the accuracy of the application model for picking or classification.

[0108] Linear velocity analysis is performed on the first arrival time of the tag data to obtain the apparent velocity of each first arrival of the tag data. Apparent velocity is the propagation distance divided by the propagation time. In practice, it can be obtained by dividing the offset by the first arrival time. Vn = Sn / Tn, where Sn is the offset and Tn is the time it takes for the first arrival wave to propagate from the shot point to the receiving point. Vn is the apparent velocity of the corresponding first arrival. Sn and Tn are obtained using existing techniques.

[0109] Using each of the first-arrival apparent velocities, the multiple seismic data are interpolated and extrapolated to obtain the first-arrival apparent velocity of each seismic trace in the multiple seismic data; interpolation refers to using the known apparent velocity of the label data to estimate the apparent velocity of other seismic data with unknown apparent velocities through a certain mathematical method. In some optional embodiments, a surface interpolation method is used to interpolate the first-arrival apparent velocities of the label data that do not include the first-arrival time in the collected label data to obtain the first-arrival apparent velocity of the label data that does not include the first-arrival time; the first-arrival apparent velocity of the seismic trace adjacent to the seismic trace with unknown first-arrival apparent velocity in the multiple seismic data is used as the first-arrival apparent velocity of the seismic trace with unknown first-arrival apparent velocity.

[0110] Using the apparent first-arrival velocity of each seismic trace in the multiple seismic data, linear correction is performed on the multiple seismic data and the label data. The purpose of linear correction is to eliminate the first-arrival propagation time differences of the seismic traces caused by the offset, flatten the first-arrival waves, and reduce the first-arrival time differences between seismic traces. After linear correction, the first-arrival peaks ideally lie on the same horizontal line, which facilitates multi-trace deep learning and multi-trace first-arrival picking in this embodiment.

[0111] The corrected time difference for each seismic trace is calculated using the apparent velocity of the first arrival of each seismic trace. This corrected time difference is used to perform a linear correction on the first arrival time of the collected tag data. The corrected time difference Tn1 = Sn / Vp, where Vp is the interpolated apparent velocity, Tn1 is the time difference to be corrected for each seismic trace, and Sn is the offset. This corrected time difference is used to perform a linear correction on the first arrival time of the tag data, ensuring that the first arrival time of the tag data is consistent with the first arrival of the tag data.

[0112] Linear correction can help reduce noise interference and provide higher-quality training data for model training, which helps improve the model training effect and the accuracy of first-arrival picking and first-arrival classification using the picking model and the binary classification model.

[0113] According to the linearly corrected earthquake dataset and label dataset, a time window including the first arrival time of all earthquake data is selected.

[0114] In some optional embodiments, a plurality of time windows are selected based on the linearly corrected seismic dataset and the label dataset;

[0115] The first arrival times of the seismic data in the multiple time windows are interpolated using the first arrival times in the linearly corrected label data set. However, the first arrival time interpolation is not performed on the label data in the multiple time windows.

[0116] Step S202: within the time window, use the label data set to train the picking model.

[0117] The label data of the seismic traces in the label data set with the first arrival time are calibrated as the first type of label data, and the label data of the seismic traces in the label data set without the first arrival time are calibrated as the second type of label data. The first type of label data is used to train the picking model. In some optional embodiments, referring to Figure 2 As shown in the figure, a fully convolutional neural network model including a bidirectional long short-term memory (BILSTM) network is used as the picking model. Conventional neural networks only extract single-channel earthquake first arrival features, while the main function of BILSTM is to perform multi-channel cross-correlation and mine relevant information of multiple seismic waves. The fully convolutional neural network and BILSTM respectively extract the single-channel and multi-channel features of the first arrival. As the input of the fully connected network, Figure 2 The function of the fully connected layer is to perform nonlinear fitting on the extracted seismic wave features and normalize the output shape size to the target shape size. The picking model is trained using the first type of label data to obtain the picking model.

[0118] Step S203: Within the time window, use the labeled data set to train the binary classification model.

[0119] Use the first-class labeled data and the second-class labeled data to train a binary classification model.

[0120] In some optional embodiments, referring to Figure 3 As shown in the figure, a fully convolutional neural network model including a bidirectional long short-term memory network (BILSTM) is used as a binary classification model. Conventional neural networks only extract single-channel earthquake first arrival features, while the main function of BILSTM is to perform multi-channel correlation and mine relevant information of multiple seismic waves. The fully convolutional neural network and BILSTM extract single-channel and multi-channel features of the first arrival respectively. As the input of the fully connected network, Figure 3 The fully connected layer performs nonlinear fitting on the extracted seismic wave features, reducing the output shape size to the target shape size. The binary classification model is trained using the first and second labeled data to obtain a binary classification model. The first and second labeled data are input into the binary classification model for training, resulting in appropriate weights for the first and second classifications.

[0121] Step S204: Within the time window, first arrivals are picked from the linearly corrected seismic data set using the trained picking model to obtain a first first arrival set. For example, first arrivals are picked from the entire work area data after linear correction based on the obtained picking model, and the picked first arrivals constitute the first first arrival set. By using a trained artificial intelligence model, i.e., a picking model, to pick first arrivals from seismic data, the efficiency of first arrival picking can be improved, the manual picking workload can be reduced, the requirements for first arrival picking of massive data sets can be met, and the seismic data processing cycle can be shortened.

[0122] Step S205: Use the trained binary classification model to classify the first arrivals in the first arrival set to obtain reliable first arrivals and unreliable first arrivals. The neural network classification problem is mainly based on the perceptual characteristics of the human brain and uses the threshold unit as the transfer function, so it can only output two values. Assuming that P is the first arrival input vector, W is the weight matrix quantity, and b is the threshold vector, then the decision boundary of the perceptron is wp+b. When wp+b>0, it is determined to be a first-class reliable first arrival, otherwise it is a second-class unreliable first arrival. By using an artificial intelligence model, i.e., a binary classification model, to further screen the picked first arrivals for reliable ones, the accuracy of picking the first arrivals can be improved.

[0123] Step S206: performing clustering processing on the first arrivals in the first arrival set to determine the unstable first arrivals contained therein.

[0124] In some optional embodiments, the first arrival time difference between adjacent seismic traces of each first arrival in the first first arrival set is calculated, and clustering filtering is performed on the first arrival time difference between adjacent seismic traces of each first arrival to obtain at least two first arrival clusters classified according to the first arrival time difference between adjacent seismic traces;

[0125] calibrating at least one first arrival cluster according to the order of change values ​​of the first arrival time differences of the first arrival clusters from large to small;

[0126] The first arrival in the marked first arrival cluster is determined to be an unstable first arrival.

[0127] For example, the K-means algorithm can be selected for clustering. The K-means algorithm is an unsupervised machine learning method that continuously searches for the nearest mean value to a seed point until the error converges. It is a typical distance-based clustering algorithm, using distance as the evaluation criterion for similarity. Based on the principle that the closer the distance between the first arrival times, the greater the similarity, the K-means algorithm considers clusters to be composed of first arrival time differences of adjacent traces with close distances. Therefore, the ultimate goal is to obtain independent clusters with similar distances. Therefore, clustering in this embodiment is the process of dividing the collection of first arrival time differences of adjacent traces into similar object classes, so that the first arrival time differences of adjacent traces within the same cluster (or class) have high similarity, while the first arrival time differences of adjacent traces in different clusters have high dissimilarity. This allows the identification of unstable first arrivals, where the first arrival time differences between adjacent seismic traces vary significantly. By identifying unstable first arrivals from a clustering perspective, the quality of the picked first arrivals can be further improved.

[0128] Step S207: First arrivals that are both unreliable and unstable are removed from the first arrival set to obtain a second first arrival set. The arrival times of the second first arrival set are fitted, and arrivals with fitting errors greater than a preset error threshold are removed to obtain a third first arrival set. The third first arrival set is subjected to inverse linear correction to obtain the first arrival wave intelligent picking result.

[0129] In some optional embodiments, fitting is performed on the first arrival times of the second first arrival set, and first arrivals with fitting errors greater than a preset error threshold are eliminated to obtain a third first arrival set, including:

[0130] The first arrival time of the second first arrival set is fitted using the following polynomial:

[0131] p(x)=p1x n +p2x n-1 +...+p n x+p n+1

[0132] Where p is the fitting coefficient, n+1 is the number of fitting first arrivals, and x is the first arrival time;

[0133] The first arrivals whose fitting errors are greater than a preset error threshold are eliminated from the second first arrival set to obtain a third first arrival set.

[0134] For example, a 3D seismic exploration area is selected, such as Figure 4The vertical lines in the figure are detection lines, which represent the arrangement of 3D seismic data. The horizontal points in the figure represent shot points. There are 498 shot records in the work area, with 36 shots per shot. First, the entire exploration seismic data is processed by field static correction (static correction data is collected in the field), and the seismic data is corrected to the work area reference plane. Then, noise attenuation processing is performed. The 498 shot seismic data after preliminary processing are processed in the Figure 4 Select label data in the exploration area blast inspection distribution plan. The label data selection results are shown in Figure 5 The vertical lines in the figure are detection lines, which represent the arrangement of 3D seismic data. The horizontal points in the figure represent shot points. A total of 25 shot data were selected, and the 25 shot data were basically evenly distributed throughout the exploration area. The selected labeled data were manually picked for the first arrival, mainly picking the more reliable first arrivals. Unreliable first arrivals were not picked, and the first arrivals with unclear or discontinuous waveforms were judged as unreliable first arrivals. Figure 6a and Figure 6b ,The vertical direction in the figure represents time, and the horizontal direction represents position. Figure 6a This is the image before the label data is manually picked up. Figure 6b The image after the label data is first picked up manually, Figure 6b The clear and continuous waveform on the upper edge of the peak-like figure is determined to be a reliable first arrival and is depicted with a thicker line. The fuzzy and discontinuous waveform on the upper edge of the peak-like figure is determined to be an unreliable first arrival. By marking the first arrival, the label data with the first arrival time is obtained. Then, a linear velocity analysis is performed based on the first arrival time of the 25-shot label data, and the apparent velocity of each first arrival of the label data is calculated. Vn = shot spacing / first arrival time. The first arrival time is the time it takes to propagate from the shot point to the detection point. The apparent velocity of the first arrival of the label data is used to interpolate and extrapolate the seismic data of the entire exploration area to obtain the apparent velocity of the first arrival wave of each seismic channel in the entire work area. The interpolated apparent velocity is used to perform linear correction processing on the first arrival time of the 498-shot data, the 25-shot label data, and the 25-shot data (correction time difference Tn1 = shot offset / apparent velocity). The correction results are shown as follows: Figure 7 The linear correction effect diagram of seismic data in the exploration area is shown in the figure. The vertical axis represents time and the horizontal axis represents position. Figure 7 Display the linear correction result of the 420th shot in the work area. After linear correction, select an appropriate first arrival time window and try to ensure that the first arrival of the entire exploration area is within the selected time window, such as Figure 8As shown in the figure, the vertical direction represents time, the horizontal direction represents position, and the time window is represented by a rectangular frame surrounded by a thick line. Based on whether the seismic trace of the label data has the first arrival time, the label data after the selected time window of 25 shots is subjected to the first arrival two-classification processing. The label data of the seismic trace with the first arrival time is calibrated as the first category of label data, and the label data of the seismic trace without the first arrival time is calibrated as the second category of label data. The 25-shot data and the first arrival are tested and trained on the full convolutional neural network model of multiple seismic traces within the given time window. In this embodiment, 5 seismic traces are selected to obtain the picking model. The full convolutional neural network model of multiple seismic traces is a full convolutional neural network model including a bidirectional long short-term memory network BILSTM. The structure of the model is shown in Figure 2 In the figure, the seismic data of multiple seismic channels are input into the fully convolutional neural network. BILSTM extracts the relevant information of the multi-channel seismic waves, and the fully convolutional neural network and BILSTM extract the features of the first arrival single channel and multiple channels respectively as the input of the fully connected network. Figure 2 The function of the fully connected layer is to perform nonlinear fitting on the extracted seismic wave features and normalize the output shape size to the target shape size. The picking model is trained with the first type of label data to obtain the picking model. The full convolutional neural network binary classification model is tested and trained on the first type of label data and the second type of label data. The neural network structure is shown in Figure 3 , and obtain the training model. The first arrival picking is performed on the 498-shot seismic data of the entire work area after linear correction using the obtained first arrival picking training model to obtain the first first arrival set. The first arrival picking results are shown in Figure 9 Then, the data is classified into two categories according to the binary classification model, and the second category of unreliable first arrival seismic traces is calibrated. In an arrangement of the three-dimensional exploration shot set, the first arrival time difference between seismic traces is calculated, and the first arrival time difference clustering filtering is performed using the first arrival time difference to calibrate the first arrival seismic traces with large and unstable first arrival time differences between seismic traces. The calibrated seismic traces that happen to belong to the second category of unreliable first arrival seismic traces are deleted to obtain the second first arrival set. The first arrival picking effect after deletion is shown in Fig. Figure 10 Then, fitting is performed on the remaining first arrival times after removing the first arrival times. In this embodiment, the fitting error threshold is set to 10 milliseconds, and the number of fitting channels is 30 adjacent seismic channels. After removing the first arrival times with fitting errors greater than 10 milliseconds, the third first arrival set is obtained. The first arrival picking effect of the third first arrival set is shown in FIG. Figure 11 Finally, the first arrival times of the third first arrival set are subjected to inverse linear time difference correction and inverse static time difference correction, until all arrangements of 498 guns have completed all steps of first arrival picking in this embodiment, and the first arrival wave intelligent picking results are obtained.

[0135] The intelligent picking method for first arrival waves of seismic data provided in this embodiment helps to reduce noise interference through linear correction, can provide higher quality training data for model training, helps to improve the model training effect, and improves the accuracy of first arrival picking and first arrival classification using picking models and binary classification models. By comparing unreliable first arrivals obtained by the binary classification model with unstable first arrivals obtained by clustering, the accuracy of eliminating first arrivals with low quality is improved, and unreliable first arrivals can be effectively deleted, which helps to improve the accuracy of first arrival picking. By training an artificial intelligence model for first arrival picking, the accuracy and efficiency of first arrival wave picking can be improved, the seismic data processing cycle can be shortened, the workload of manual modification of first arrivals can be reduced, it helps to adapt to the requirements of first arrival picking of massive data, and improve the imaging quality of seismic data.

[0136] Example 3

[0137] The third embodiment of the present invention provides a seismic data first arrival wave intelligent picking device, the structure of which is as follows: Figure 13 Shown, including:

[0138] The label data set construction module 101 is configured to collect a plurality of label data from a plurality of seismic data, perform linear correction on the plurality of seismic data and the plurality of label data, respectively, to obtain a linearly corrected seismic data set and a label data set including the plurality of linearly corrected label data; and select a time window including the first arrival times of all seismic data based on the linearly corrected seismic data set and the label data set.

[0139] A model training module 102 is configured to train a picking model and a binary classification model using the label dataset within the time window;

[0140] The first arrival wave picking module 103 is used to pick the first arrival of the linearly corrected seismic data set within the time window using the trained picking model to obtain a first first arrival set; use the trained binary classification model to classify the first arrivals in the first first arrival set to obtain reliable first arrivals and unreliable first arrivals; cluster the first arrivals in the first first arrival set to determine the unstable first arrivals contained therein; eliminate the first arrivals that belong to the unreliable first arrivals and the unstable first arrivals from the first first arrival set to obtain a second first arrival set; fit the first arrival time of the second first arrival set to eliminate the first arrivals whose fitting error is greater than a preset error threshold value to obtain a third first arrival set; perform anti-linear correction on the third first arrival set to obtain a first arrival wave intelligent picking result.

[0141] In this embodiment, linear correction is used to help reduce noise interference, provide higher quality training data for model training, help improve model training effects, and improve the accuracy of first arrival picking and first arrival classification using picking models and binary classification models. By comparing unreliable first arrivals obtained by the binary classification model with unstable first arrivals obtained by clustering, the accuracy of eliminating low-quality first arrivals is improved, and unreliable first arrivals can be effectively deleted, which helps to improve the accuracy of first arrival picking. By training artificial intelligence models for first arrival picking, the accuracy and efficiency of first arrival wave picking can be improved, the seismic data processing cycle can be shortened, and the workload of manual modification of first arrivals can be reduced, which helps to adapt to the requirements of first arrival picking of massive data and improve the imaging quality of seismic data.

[0142] Based on the same inventive concept, an embodiment of the present invention further provides a computer storage medium, wherein the computer storage medium stores computer executable instructions, and when the computer executable instructions are executed by a processor, the aforementioned method for intelligently picking first arrival waves of seismic data is implemented.

[0143] Based on the same inventive concept, an embodiment of the present invention also provides a terminal device, including: a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the program, the aforementioned method for intelligently picking up the first arrival wave of seismic data is implemented.

[0144] Regarding the apparatus in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0145] It should be understood that the specific order or hierarchy of steps in the disclosed processes is an example of an exemplary method. Based on design preferences, it should be understood that the specific order or hierarchy of steps in the process can be rearranged without departing from the scope of the present disclosure. The accompanying method claims present elements of the various steps in an exemplary order and are not intended to be limited to the specific order or hierarchy described.

[0146] In the foregoing detailed description, various features are grouped together in a single embodiment to simplify the disclosure. This method of disclosure should not be interpreted as reflecting an intention that embodiments of the claimed subject matter require more features than are expressly recited in each claim. On the contrary, as reflected in the appended claims, the invention comprises less than all the features of any individual disclosed embodiment. The appended claims are therefore hereby expressly incorporated into the detailed description, with each claim standing on its own as a separate preferred embodiment of the invention.

[0147] The steps of the methods or algorithms described in conjunction with the embodiments herein may be directly embodied as hardware, software modules executed by a processor, or a combination thereof. The software module may be located in a RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, register, hard disk, removable disk, CD-ROM, or any other form of storage medium well known in the art. An exemplary storage medium is connected to the processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium may also be an integral part of the processor. The processor and storage medium may be located in an ASIC. The ASIC may be located in a user terminal. Of course, the processor and storage medium may also be present in a user terminal as discrete components.

[0148] For software implementation, the techniques described in this application can be implemented using modules (e.g., procedures, functions, etc.) that perform the functions described in this application. These software codes can be stored in a memory unit and executed by a processor. The memory unit can be implemented within the processor or external to the processor. In the latter case, it is communicatively coupled to the processor via various means, which are well known in the art.

[0149] The foregoing description includes examples of one or more embodiments. Of course, it is not possible to describe all possible combinations of components or methods for the purposes of describing the above embodiments, but one of ordinary skill in the art will recognize that the various embodiments may be further combined and arranged. Therefore, the embodiments described herein are intended to encompass all such changes, modifications and variations that fall within the scope of the appended claims. Furthermore, to the extent the term "comprising" is used in the specification or claims, the term is intended to be encompassed in a manner similar to the term "including," as explained in terms of "including," used as a transitional word in the claims. Furthermore, any use of the term "or" in the specification of the claims is intended to mean a "non-exclusive or."

Claims

1. A method for intelligently picking first arrival waves from seismic data, characterized in that: include: Collecting a plurality of label data from a plurality of seismic data, and performing linear correction on the plurality of seismic data and the plurality of label data, respectively, to obtain a linearly corrected seismic data set and a label data set including the linearly corrected plurality of label data; Based on the linearly corrected earthquake dataset and the label dataset, a time window including the first arrival time of all earthquake data is selected; Within the time window, using the label dataset, respectively training the picking model and the binary classification model; Within the time window, first arrival picking is performed on the linearly corrected seismic data set using the trained picking model to obtain a first first arrival set; Using the trained binary classification model, the first arrivals in the first arrival set are classified into reliable first arrivals and unreliable first arrivals; Cluster the first arrivals in the first arrival set to identify the unstable first arrivals contained therein; Eliminate the first arrivals that belong to both the unreliable first arrivals and the unstable first arrivals from the first first arrival set to obtain a second first arrival set; Fitting the first arrival times of the second first arrival set, eliminating first arrivals with fitting errors greater than a preset error threshold, to obtain a third first arrival set; An inverse linear correction process is performed on the third first arrival set to obtain a first arrival wave intelligent picking result.

2. The method according to claim 1, wherein The step of collecting a plurality of label data from a plurality of seismic data comprises: Performing preliminary processing on a plurality of seismic data; the preliminary processing includes processing the seismic data by an observation system; Label data is collected from the plurality of seismic data after preliminary processing.

3. The method according to claim 1, wherein Performing linear correction on the plurality of seismic data and the plurality of label data respectively, including: Performing linear velocity analysis on the collected tag data including the first arrival time to obtain the first arrival apparent velocity of each tag data including the first arrival time; Using each first-arrival apparent velocity, interpolating and extrapolating the plurality of seismic data, to obtain the first-arrival apparent velocity of each seismic trace in the plurality of seismic data; Performing linear correction processing on the plurality of seismic data and label data using the first arrival apparent velocity of each seismic trace of the plurality of seismic data; The first arrival apparent velocity of each seismic trace is used to calculate the corresponding correction time difference of each seismic trace, and the correction time difference is used to perform corresponding linear correction processing on the first arrival time of the collected tag data.

4. The method according to claim 3, wherein The method of using each first arrival apparent velocity to interpolate and extrapolate the plurality of seismic data to obtain the first arrival apparent velocity of each seismic trace in the plurality of seismic data comprises: Using a surface interpolation method, interpolating the first arrival apparent velocity of the tag data that does not include the first arrival time in the collected tag data to obtain the first arrival apparent velocity of the tag data that does not include the first arrival time; The first arrival apparent velocity of the seismic trace adjacent to the seismic trace with unknown first arrival apparent velocity in the plurality of seismic data is used as the first arrival apparent velocity of the seismic trace with unknown first arrival apparent velocity.

5. The method according to claim 2, wherein The preliminary processing of the plurality of seismic data further includes: At least one of noise attenuation, filtering, and static correction processing is performed on the plurality of seismic data.

6. The method according to claim 5, wherein When the preliminary processing of seismic data includes the static correction processing, the method further includes: An anti-static time difference process is performed on the third first arrival set.

7. The method according to claim 1, wherein The method of selecting a time window including the first arrival times of all earthquake data based on the linearly corrected earthquake data set and the label data set includes: Select multiple time windows based on the linearly corrected earthquake dataset and the label dataset; The first arrival times of the seismic data in the multiple time windows are interpolated using the first arrival times in the linearly corrected label data set.

8. The method according to claim 1, wherein Within the time window, the label dataset is used to train the picking model and the binary classification model respectively, including: Marking the label data of the seismic traces with first arrival time in the label data set as the first type of label data, and marking the label data of the seismic traces without first arrival time in the label data set as the second type of label data; Use the first type of labeled data to train the picking model; Use the first-class labeled data and the second-class labeled data to train a binary classification model.

9. The method according to claim 1, wherein The clustering process of the first arrivals in the first arrival set to determine the unstable first arrivals contained therein includes: Calculating the first arrival time difference between adjacent seismic traces of each first arrival in the first first arrival set, performing clustering filtering on the first arrival time difference between adjacent seismic traces of each first arrival, and obtaining at least two first arrival clusters classified according to the first arrival time difference between adjacent seismic traces; calibrating at least one first arrival cluster according to the order of change values ​​of the first arrival time differences of the first arrival clusters from large to small; The first arrivals in the marked first arrival cluster are determined to be unstable first arrivals.

10. The method according to claim 1, wherein The fitting process is performed on the first arrival times of the second first arrival set, and first arrivals with fitting errors greater than a preset error threshold are eliminated to obtain a third first arrival set, including: The first arrival time of the second first arrival set is fitted using the following polynomial: p(x)=p1x n +p2x n-1 +...+p n x+p n+1 Where p is the fitting coefficient, n+1 is the number of fitting first arrivals, and x is the first arrival time; The first arrivals whose fitting errors are greater than a preset error threshold are eliminated from the second first arrival set to obtain a third first arrival set.

11. An intelligent device for picking up first arrival waves of seismic data, characterized in that: include: a label data set construction module, configured to collect a plurality of label data from a plurality of seismic data, and perform linear correction on the plurality of seismic data and the plurality of label data, respectively, to obtain a linearly corrected seismic data set and a label data set containing the linearly corrected plurality of label data; A model training module is used to select a time window including the first arrival time of all earthquake data based on the linearly corrected earthquake data set and the label data set; Within the time window, using the label dataset, respectively training the picking model and the binary classification model; The first arrival wave picking module is used to pick the first arrival of the linearly corrected seismic data set within the time window using the trained picking model to obtain a first first arrival set; use the trained binary classification model to classify the first arrivals in the first first arrival set to obtain reliable first arrivals and unreliable first arrivals; cluster the first arrivals in the first first arrival set to determine the unstable first arrivals contained therein; eliminate the first arrivals that belong to the unreliable first arrivals and the unstable first arrivals from the first first arrival set to obtain a second first arrival set; fit the first arrival time of the second first arrival set to eliminate the first arrivals whose fitting error is greater than a preset error threshold value to obtain a third first arrival set; perform anti-linear correction on the third first arrival set to obtain a first arrival wave intelligent picking result.

12. A computer storage medium, characterized in that The computer storage medium stores computer executable instructions, which, when executed by a processor, implement the method for intelligently picking first arrival waves of seismic data as described in any one of claims 1-10.

13. A terminal device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for intelligently picking first arrival waves of seismic data as described in any one of claims 1 to 10 is implemented.

Images