Frequency dispersion curve automatic extraction method based on deep learning and related device

Abstract

The invention discloses a frequency dispersion curve automatic extraction method based on deep learning and a related device, and relates to the technical field of geophysics and underground structure exploration, and the method comprises the steps: obtaining station pair data; performing cross-correlation processing on the station pair data to obtain a plurality of pieces of cross-correlation data; performing data processing on each piece of cross-correlation data to obtain processed cross-correlation data; according to the processed cross-correlation data, a tau-p conversion technology is adopted to obtain a frequency dispersion energy diagram; performing image recognition on the frequency dispersion energy diagram based on a deep learning model to obtain a frequency dispersion curve; the frequency dispersion curve is used for displaying the change rule of the frequency dispersion characteristic along with the periodic change speed; the deep learning model adopts an encoder-decoder framework, and the depth learning model adopts an encoder-decoder framework; in the deep learning model, a pooling layer is replaced by convolution with the step length being two, and up-sampling is carried out by adopting transposed convolution. Based on the method provided by the invention, the frequency dispersion curve can be efficiently and accurately extracted.

Description

Technical Field

[0001] The present application relates to the technical fields of geophysics and underground structure exploration, and particularly to a method and related device for automatically extracting dispersion curves based on deep learning. Background Art

[0002] The technology of probing underground structures using ambient noise inversion has a history of more than a decade. This technology not only plays a key role in the detection of underground structures in the field of earth science, but also plays a crucial role in traditional engineering fields, and has become one of the most critical technologies in the field of underground exploration. To reveal the underground structure, it is necessary to extract the dispersion energy map from the signals received by the instrument, which usually involves the processes of cross-correlation stacking and inversion. Extracting the dispersion curve from the dispersion energy map is the core step of this process. However, the traditional manual extraction method is time-consuming and labor-intensive. For example, to explore the underground structure of an area, it may be necessary to extract up to forty to fifty thousand dispersion curves. In the traditional manual way, even if a person works full-time, only one to two thousand dispersion curves can be extracted, which undoubtedly consumes a large amount of human resources. The emergence of the technology of automatically extracting dispersion curves based on deep learning effectively solves this problem, not only greatly reducing the manual requirements, but also the quantity and quality of the extracted dispersion curves are comparable to those of manual extraction, and even better in some cases.

[0003] Although predecessors have tried to automatically extract dispersion curves through machine learning technologies, such as models like Dispernet and Disperpicker, these methods have limitations that restrict their wide promotion in practical applications. Taking the recently proposed Disperpicker as an example, its model structure and parameter settings are not ideal, resulting in unsatisfactory recognition effects. In addition, its complex post-processing steps make the running time too long, affecting the efficiency and accuracy of dispersion curve extraction. Summary of the Invention

[0004] The purpose of the present application is to provide a method and related device for automatically extracting dispersion curves based on deep learning, which can efficiently and accurately extract dispersion curves.

[0005] To achieve the above purpose, the present application provides the following solutions:

[0006] In the first aspect, the present application provides a method for automatically extracting dispersion curves based on deep learning, including:

[0007] Obtain station pair data; the station pair data is seismic waveform data recorded by at least two seismic stations;

[0008] Perform cross-correlation processing on the station pair data to obtain a number of cross-correlation data;

[0009] Perform data processing on each of the cross-correlation data to obtain processed cross-correlation data;

[0010] According to the processed cross-correlation data, use the τ-p transformation technique to obtain a dispersion energy map; the dispersion energy map is used to display the variation law of the dispersion characteristics with the change speed of the period;

[0011] Perform image recognition on the dispersion energy map based on a deep learning model to obtain a dispersion curve; the dispersion curve is used to display the variation law of the dispersion characteristics with the change speed of the period; the deep learning model uses an encoder-decoder architecture; in the deep learning model, a convolution with a stride of two is used to replace the pooling layer, and transposed convolution is used for upsampling.

[0012] In a second aspect, the present application provides a deep learning-based automatic extraction system for dispersion curves, including:

[0013] A data acquisition module for acquiring station pair data; the station pair data is seismic waveform data recorded by at least two seismic stations.

[0014] A cross-correlation processing module for performing cross-correlation processing on the station pair data to obtain a number of cross-correlation data.

[0015] A data processing module for performing data processing on each of the cross-correlation data to obtain processed cross-correlation data.

[0016] A dispersion energy map construction module for obtaining a dispersion energy map according to the processed cross-correlation data by using the τ-p transformation technique; the dispersion energy map is used to display the variation law of the dispersion characteristics with the change speed of the period.

[0017] A dispersion curve extraction module for performing image recognition on the dispersion energy map based on a deep learning model to obtain a dispersion curve; the dispersion energy map is used to display the variation law of the dispersion characteristics with the change speed of the period; the deep learning model uses an encoder-decoder architecture; in the deep learning model, a convolution with a stride of two is used to replace the pooling layer, and transposed convolution is used for upsampling.

[0018] In a third aspect, the present application provides a computer device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, where the processor executes the computer program to implement a deep learning-based automatic extraction method for dispersion curves as described in any one of the above.

[0019] In a fourth aspect, the present application provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements a deep learning-based automatic extraction method for dispersion curves as described in any one of the above.

[0020] In a fifth aspect, the present application provides a computer program product, including a computer program which, when executed by a processor, implements a method for automatically extracting dispersion curves based on deep learning as described in any one of the above.

[0021] According to the specific embodiments provided by the present application, the following technical effects are disclosed:

[0022] The present application provides a method for automatically extracting dispersion curves based on deep learning and related devices. The method includes: First, obtaining station pair data is the first step in extracting dispersion curves, which involves collecting seismic waveform data recorded by at least two seismic stations. Then, perform cross-correlation processing on the station pair data. Cross-correlation processing can enhance signals and suppress noise, obtaining a number of cross-correlation data. Cross-correlation processing helps to extract the characteristics of seismic wave propagation between different stations. Perform data processing on the cross-correlation data to obtain processed cross-correlation data. This step ensures data quality and provides accurate input for subsequent analysis. Furthermore, use the τ-p transform technology to process the processed cross-correlation data to obtain a dispersion energy map. By analyzing the dispersion energy map, the propagation characteristics of seismic waves with different frequency components at different times can be visually seen. Based on a deep learning model, perform image recognition on the dispersion energy map to obtain a dispersion curve. The dispersion curve shows the variation law of dispersion characteristics with time and frequency and is an important tool for analyzing the change in seismic wave propagation speed. Among them, the deep learning model adopts an encoder-decoder architecture, and this structure is suitable for processing image recognition problems. In the model, a convolution with a stride of two replaces the pooling layer, and transposed convolution is used for upsampling, which helps to maintain the detailed information of the dispersion energy map and improve the accuracy of dispersion curve extraction. Through the above steps, the present application can extract dispersion curves more efficiently and accurately. Description of the Drawings

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0024] Figure 1 It is a flowchart of a method for automatically extracting dispersion curves based on deep learning provided by an embodiment of the present application.

[0025] Figure 2 It is a structural diagram of a Disperpicker model provided by an embodiment of the present application.

[0026] Figure 3An architecture diagram of an automatic picking dispersion curve model provided by an embodiment of the present application.

[0027] Figure 4 A structural diagram of a residual link module and a channel attention module provided by an embodiment of the present application.

[0028] Figure 5 A comparison chart of automatic and manual picking group velocities provided by an embodiment of the present application.

[0029] Figure 6 A comparison chart with manual extraction under the data of Area A provided by an embodiment of the present application.

[0030] Figure 7 A comparison chart with the Disperpicker model under the Namcha Barwa data provided by an embodiment of the present application.

[0031] Figure 8 A comparison chart with the Disperpicker model under the Namcha Barwa data provided by an embodiment of the present application.

[0032] Figure 9 A comparison chart of group velocity prediction errors provided by an embodiment of the present application.

[0033] Figure 10 A statistical comparison chart of group velocity prediction errors provided by an embodiment of the present application.

[0034] Figure 11 A comparison chart of phase velocity prediction errors provided by an embodiment of the present application.

[0035] Figure 12 A statistical comparison chart of phase velocity prediction errors provided by an embodiment of the present application.

[0036] Figure 13 A structural schematic diagram of a computer device provided by an embodiment of the present application. Detailed implementation manners

[0037] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0038] The currently latest proposed method for picking dispersion curves based on a convolutional neural network is called Disperpicker, as Figure 2As shown in the figure, it consists of three steps. First, a CNN simplified from U-net (Ronneberger et al., 2015) is used to complete the task of automatically extracting dispersion curves from the dispersion spectrograms in the cyclic velocity domain. The training dataset includes phase and group velocity dispersion spectrograms and the corresponding phase and group velocity dispersion curves. Using the pre-prepared dataset, a CNN with phase and group velocity dispersion spectrograms as inputs and a predicted probability image as output is trained. Then, a dispersion extraction strategy for automatically extracting dispersion curves is proposed. Finally, data quality control criteria are used to ensure the reliability of the extracted dispersion curves.

[0039] However, the model parameters are too simple, resulting in insufficient recognition ability. The lack of an attention mechanism limits the model's performance, and the post-processing of the probability map output by the model is too complicated, resulting in too long running time. This makes its extraction effect of dispersion curves poor.

[0040] The purpose of this application is to provide a method and related device for automatically extracting dispersion curves based on deep learning, which can efficiently and accurately extract dispersion curves.

[0041] To make the above objects, features, and advantages of this application more obvious and understandable, the following further details this application in conjunction with the accompanying drawings and specific embodiments.

[0042] In an exemplary embodiment, a method for automatically extracting dispersion curves based on deep learning is provided. This method is executed by a computer device, specifically, it can be executed independently by a computer device such as a terminal or a server, or jointly executed by a terminal and a server. In the embodiments of this application, this method is described by taking its application to an example, including the following steps 1 to step 5. Among them:

[0043] Step 1: Obtain station pair data; the station pair data is seismic waveform data recorded by at least two seismic stations.

[0044] Step 2: Perform cross-correlation processing on the station pair data to obtain a number of cross-correlation data.

[0045] Step 3: Perform data processing on each of the cross-correlation data to obtain processed cross-correlation data.

[0046] Step 4: According to the processed cross-correlation data, use the τ-p transformation technique to obtain a dispersion energy map; the dispersion energy map is used to show the variation law of the dispersion characteristics with the change speed of the period.

[0047] Step 5: Perform image recognition on the dispersion energy map based on a deep learning model to obtain a dispersion curve; the dispersion curve is used to show the variation law of the dispersion characteristics with the change speed of the period; the deep learning model adopts an encoder-decoder architecture; in the deep learning model, convolution with a stride of two is used to replace the pooling layer, and transposed convolution is used for upsampling.

[0048] Among them, in some embodiments, when performing Step 2, it can be specifically as follows:

[0049] Cross-correlation processing is a signal processing technique used to analyze the similarity between two signals. In this process, one signal is used as a reference, and the other signal is delayed by different times, and then the two are multiplied and integrated. By changing the delay time, the best alignment between the two signals can be found, thereby obtaining their similarity. In the fields of seismology, acoustics, electronics, etc., cross-correlation processing is often used for signal detection, time delay estimation, noise cancellation, etc. For example, in seismic monitoring, multiple stations record seismic wave signals. By cross-correlation processing, the signal-to-noise ratio of the signal can be improved, and clearer seismic waveform features can be extracted, so as to be used for seismic location, study of earthquake source mechanisms, etc.

[0050] Among them, in some embodiments, when performing Step 3, it can be specifically as follows:

[0051] The data processing includes the superposition operation of each cross-correlation data.

[0052] In this embodiment, the superposition operation can be linear superposition or weighted superposition. Linear superposition is to directly add multiple cross-correlation data, while weighted superposition assigns different weights according to the importance of each data and then performs superposition. Weighted superposition can improve the prominence of specific signal features, so as to more accurately extract key information in the seismic waveform.

[0053] Among them, in some embodiments, when performing Step 4, it can be specifically as follows:

[0054] By processing the cross-correlation data, this embodiment adopts the τ-p transform technology and successfully obtains a dispersion energy map. It can clearly depict the energy distribution of the signal at different period points and speeds, thus providing intuitive visual support for analyzing and understanding the dispersion behavior of the signal.

[0055] Among them, in some embodiments, when performing Step 5, it can be specifically as follows:

[0056] Based on the deep learning model, perform image recognition on the dispersion energy map to obtain a dispersion probability map output by the deep learning.

[0057] Based on the dispersion probability map output by the deep learning.

[0058] Convert the dispersion probability map output by deep learning into a curve form to obtain a number of initial dispersion curves.

[0059] Eliminate the abnormal curves in each of the initial dispersion curves to obtain the dispersion curves.

[0060] Among them, the deep learning model includes an encoder part and a decoder part. The encoder part includes a number of encoding layers, and a channel attention module is set after each encoding layer. The decoder part includes a number of decoding layers, and a channel attention module is set after each decoding layer.

[0061] Specifically, the model is divided into an encoder and a decoder part. The encoder extracts increasingly abstract features, and the decoder part restores the size of the image and performs feature fusion. The input image size is 1501*151*1. After downsampling by the convolutional kernel, the size of the input image is halved and the number of channels is doubled. There are a total of five layers in the encoder part. As the number of layers gets deeper, the number of channels increases, and the extracted features become more abstract. Finally, in the decoder part, the size of the original image is restored, and through the fusion between features, the extracted abstract coarse-grained features and the shallow specific fine-grained features are fused, thereby improving the model's recognition ability for images. A channel attention module is added after each layer of the encoder and decoder parts to refine the channel data of the model to further improve the model's recognition ability. After continuous restoration by the decoder part, the final output size of the model is 1501*151*2, and the model passes through the sigmoid function to obtain the final output probability map. The model uses a simple RELU function for non-linear activation.

[0062] Specifically, Figure 3 Shows the architecture of the automatic extraction model of dispersion curves. This model is based on the encoder-decoder architecture. Different from U-Net, in this application, convolutional layers with a stride of two are used to replace the pooling layers, and upsampling is achieved through transposed convolutions. The method of feature fusion is basically the same, but this application uses a convolution with a residual connection as the basic convolution operation, and a channel attention mechanism is introduced after each layer to refine the features extracted by the model. Finally, feature fusion is achieved through cross-layer connections, which enables the full combination of coarse-grained abstract features and fine-grained features, thereby improving the recognition effect of the model.

[0063] Specifically, in the model provided in this embodiment, Figure 4The residual connection and channel attention module of the model are shown. The residual connection is adopted to prevent problems that may occur when the model has too many layers, and it enhances the performance of the model. The channel attention module uses an adaptive attention mechanism to refine the data of each channel, further improving the recognition ability of the model. In addition, global pooling reduces the number of module parameters.

[0064] Specifically, as Figure 1 shown, steps 1-4 are the preprocessing stage in the method for automatically extracting dispersion curves based on deep learning. The model construction and model training in step 5 are the model processing stage, and the dispersion curve extraction part in step 5 is the subsequent processing stage.

[0065] This application further expands and introduces a specific application scenario that makes full use of the method for automatically extracting dispersion curves based on deep learning technology proposed in this application. Specifically, the method for automatically extracting dispersion curves based on deep learning involved in this embodiment is not limited to theoretical applications, but can be widely applied to actual engineering and scientific research fields. Through this method, the dispersion curves can be effectively and automatically extracted, thereby improving the efficiency and accuracy of data processing, reducing the need for manual intervention, and ultimately enhancing the automation level of the entire workflow.

[0066] Specifically, in practical applications, the model provided in this embodiment mainly includes three parts. The first part is to superimpose the cross-correlation data and perform subsequent signal processing and τ-p transformation to obtain the dispersion energy map. The second part is to input the group velocity dispersion energy map into the model through a trained machine learning model to obtain an output probability map, and input the phase velocity dispersion energy map into the model to obtain an output probability map. The third part is to obtain the target data through subsequent processing. For example, the larger value in the probability map is obtained. In the phase velocity probability map, if the probabilities of two dispersion curves are both relatively large, then these two dispersion curves are searched out through a search algorithm (the starting and ending positions of the two dispersion curves are obtained through subsequent operations on the output results of the machine learning method), and then the cumulative probability values of the two dispersion curves are compared, and the curve with the larger probability is selected as the final output dispersion curve. Then, the output dispersion curves that are too short in length are removed, and outliers are processed. Finally, the output final dispersion curve is obtained and output in the txt format as the required extracted dispersion curve file.

[0067] 1) Comparison of extraction effects:

[0068] As Figure 5 shown, Figure 5(a) in it is a schematic diagram of manually picked group velocity, (b) is a probability map of automatically picked group velocity, and (c) is a probability map of manually picked group velocity. According to the method proposed in this application, by processing cross-correlation data and applying the τ-p transformation, a dispersion energy map can be obtained. The role of the machine learning model is to receive these dispersion energy maps as inputs and output corresponding probability maps. The next step is to convert the probability maps into the required dispersion curves. In some cases, two curves may be output for the phase velocity, and their probability values are both relatively high (because in the manual extraction process, these two dispersion curves are equally applicable). In this case, a specific method will be used to select the curve with the highest cumulative probability and combine subsequent processing steps such as removing outliers to obtain the final output of the dispersion curve. As can be seen from Figure 6 it, the automatically extracted dispersion curve is very similar to the manually extracted one, but the automatic extraction method significantly reduces the manual requirements and improves the processing speed.

[0069] In this embodiment, data in area A was specifically collected, as Figure 6 shown. By comparing Figure 6 in it, the (a) group velocity dispersion curve of area A manually extracted, (b) group velocity dispersion curve of area A automatically extracted, (c) phase velocity dispersion curve of area A manually extracted, and (d) phase velocity dispersion curve of area A automatically extracted, it can be found that the number of automatically extracted dispersion curves is comparable to that of manually extracted ones, but the effect is even better than that of manually extracted ones. Maybe because there may be a situation of slacking off when the number of manually extracted curves is too large, and at the same time, limited by the functions of the human brain and eyes, the extraction in some areas is not precise enough.

[0070] As Figure 7 shown. By comparing Figure 7 in it, the (a) group velocity dispersion curve of Namcha Barwa of this model and (b) group velocity dispersion curve of Namcha Barwa of Disperpicker, it can be known that a total of 37,472 group velocity dispersion curves and 32,612 phase velocity dispersion curves were extracted from the closely spaced array data in the Namcha Barwa area using the model designed in this application, and the original cross-correlation data had 66,175. Corresponding to this, Disperpicker extracted 7,100 group velocity and phase velocity dispersion curves. The number of dispersion curves extracted in this application is about five times that of Disperpicker, and all the dispersion curves were extracted in only two days. While Disperpicker took more than a week even when ten programs were running simultaneously. Moreover, it can be found that some of the dispersion curves extracted by Disperpicker are too steep and curved.

[0071] As Figure 8 shown. By comparing Figure 8In (a) the phase velocity dispersion curve of the present model for Namcha Barwa and (b) the phase velocity dispersion curve of Disperpicker for Namcha Barwa, it can be seen that the extraction is faster. On the one hand, it is because a good graphics card is used, and on the other hand, it is also because a lot of optimizations have been made in the post-processing of the code compared to Disperpicker. Moreover, it can be found by comparison that some of the dispersion curves extracted by Disperpicker are highly curved, while the dispersion curves extracted by the present application are relatively smoother. The reason for the different velocity ranges is that different velocity ranges are set when initially outputting the dispersion energy diagram. The setting of the model of the present application is 1.5 - 4.5, while the setting of Disperpicker is 0.5 - 4.

[0072] 2) Comparison of extraction errors:

[0073] As Figure 9 shown, by comparing Figure 9 in (a) the group velocity prediction error of the present model and (b) the group velocity prediction error of Disperpicker, it can be seen that the automatically extracted points of the group velocity and the manually extracted points are almost the same, and only a small number of points deviate slightly. The reason is that the characteristics of the corresponding dispersion energy diagram are not common, resulting in difficult judgment for manual extraction, and naturally there will be some differences in automatic extraction. In comparison, the group velocity prediction of Disperpicker is not very good. The automatic extraction has a certain fitting ability, but compared with the automatic extraction results of the present model, the accuracy is lower.

[0074] As Figure 10 shown, by comparing Figure 10 in (a) the statistical chart of the group velocity prediction error of the present model and (b) the statistical chart of the group velocity prediction error of Disperpicker, it can be seen that the error of the model of the present application is basically concentrated around 0, and there is almost no error. The goodness of fit for the automatic extraction of the group velocity is 0.99879, and the root mean square error rmse is 0.0045. Compared with the error of Disperpicker, it is several times worse. The goodness of fit of Disperpicker is 0.97797, and the root mean square error rmse is 0.02056.

[0075] From the statistical chart, the error distribution of the model of the present application is almost concentrated around zero, while Disperpicker has a relatively discrete normal distribution with certain errors.

[0076] As Figure 11 shown, by comparing Figure 11Regarding (a) the phase velocity prediction error of this model and (b) the phase velocity prediction error of Disperpicker in [reference], it can be seen that for the phase velocity, the model of this application also has a good fit. There is only a little jitter near the central axis, and there is almost no significant difference from the manually extracted result. Sometimes, there may be a situation where two phase velocities can be extracted, which will lead to differences between manual extraction and automatic extraction. The current extraction effect of the model of this application is not much different from manual extraction, which is already very good. In contrast, the automatic extraction effect of Disperpicker for the phase velocity is not very good. The points in the figure are relatively discrete, and the general range has been extracted. Generally speaking, the error is within the controllable range.

[0077] As Figure 12 shown, by comparing Figure 12 the (a) statistical chart of the phase velocity prediction error of this model and (b) the statistical chart of the phase velocity prediction error of Disperpicker in [reference], it can be seen that the prediction error of the phase velocity is not as concentrated as that of the group velocity. After all, the phase velocity is not as easy to predict as the group velocity. The goodness of fit of the phase velocity is 0.99915, and the root mean square error rms is 0.0023. The prediction error of the phase velocity shows a certain degree of normal distribution trend, but the degree of dispersion is not too high. In comparison, the prediction goodness of fit of Disperpicker is 0.73457, and the root mean square error rms is 0.03903. The prediction error of this application is about one-tenth of that of Disperpicker. From the statistical chart, the prediction error of Disperpicker is more discrete, and the prediction error of this application is more concentrated. Basically, the prediction errors all tend to 0.

[0078] Based on the same inventive concept, the embodiments of this application also provide a data processing system for implementing the above-mentioned automatic extraction method of dispersion curves based on deep learning. The implementation solutions provided by this system to solve problems are similar to those recorded in the above method. Therefore, the specific limitations in one or more data processing system embodiments provided below can refer to the limitations on the automatic extraction method of dispersion curves based on deep learning in the above text, and will not be repeated here.

[0079] In an exemplary embodiment, a system for an automatic extraction method of dispersion curves based on deep learning is provided, including:

[0080] A data acquisition module for acquiring station pair data; the station pair data is seismic waveform data recorded by at least two seismic stations.

[0081] A cross-correlation processing module for performing cross-correlation processing on the station pair data to obtain a plurality of cross-correlation data.

[0082] A data processing module, configured to perform data processing on each of the cross-correlation data to obtain processed cross-correlation data.

[0083] A dispersion energy map construction module, configured to obtain a dispersion energy map according to the processed cross-correlation data by using the τ-p transformation technique; the dispersion energy map is used to show the variation law of the dispersion characteristics with the change speed of the period.

[0084] A dispersion curve extraction module, configured to perform image recognition on the dispersion energy map based on a deep learning model to obtain a dispersion curve; the dispersion curve is used to show the variation law of the dispersion characteristics with the change speed of the period; the deep learning model adopts an encoder-decoder architecture; in the deep learning model, a convolution with a stride of two is used to replace the pooling layer, and a transposed convolution is used for upsampling.

[0085] In an exemplary embodiment, a computer device is provided. The computer device can be a server or a terminal, and its internal structure diagram can be as Figure 13 shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data processing data. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements a method for automatically extracting a dispersion curve based on deep learning.

[0086] Those skilled in the art can understand that Figure 13 the structure shown in

[0087] is only a block diagram of some structures related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0088] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, which when executed by a processor implements the steps in the above method embodiments.

[0089] In an exemplary embodiment, a computer program product is provided, including a computer program, which when executed by a processor implements the steps in the above method embodiments.

[0090] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.

[0091] In summary, this application has the following technical effects:

[0092] 1) This application uses an encoder-decoder architecture neural network based on the channel attention mechanism to automatically extract dispersion curves. The model is divided into two parts: an encoder and a decoder. The original pooling layer is replaced with a 3*3 convolutional kernel with a stride of 2. Feature fusion is performed through cross-layer connections, and the final feature output is performed through a convolutional kernel with a larger size. Deep learning is applied to surface wave processing, achieving a faster speed and even higher accuracy compared to traditional manual extraction, and also making further improvements compared to the previous deep learning dispersion curve extraction models (such as Dispernet, Disperpicker, etc.).

[0093] 2) This application uses residual links to prevent possible problems when the model is too deep and at the same time improves the model's capabilities. The channel attention module uses an adaptive attention mechanism to refine the data of each channel, further improving the model's recognition ability, and uses global pooling to save model parameters.

[0094] 3) Using background noise to extract dispersion curves and invert the underground structure is important for both exploring the shallow underground structure in earth science and the underground situation before engineering construction. To explore the underground structure of a place, at least four to fifty thousand dispersion curves need to be manually extracted, which undoubtedly consumes a lot of manpower. Automatically extracting dispersion curves through deep learning can greatly save manpower. Entering the big data era, this work becomes even more important in the face of a huge amount of station signals.

[0095] 4) The model designed in this application for automatically extracting dispersion curves has not only made many improvements in the model architecture. Three regions with significantly different dispersion energy maps are adopted for the training data, and thousands of carefully selected data are manually extracted respectively to ensure the generalization performance of the model. Moreover, the subsequent post-processing process is optimized, greatly improving the speed of automatic extraction. The model automatically extracted 37,472 group velocity and phase velocity dispersion curves in Namcha Barwa in the experimental area, compared with 7,100 extracted by Disperpicker. The number of extracted curves is more than five times that of the latter. The time required is more than two days, with an average of more than 20,000 curves extracted per day. Compared with Disperpicker, which extracts two or three thousand curves per day, the extraction speed is ten times that of the latter. In addition, the extraction accuracy is indistinguishable from that of manual extraction, and large-scale popularization of automatic extraction of dispersion curves can already be achieved.

[0096] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0097] Specific examples are used in this article to elaborate on the principle and implementation mode of this application. The description of the above embodiments is only used to help understand the method and its core idea of this application; at the same time, for those of ordinary skill in the art, based on the idea of this application, there will be changes in the specific implementation mode and application scope. In summary, the content of this specification should not be construed as a limitation to this application.

Claims

1. A method for automatic extraction of dispersion curves based on deep learning, characterized in that: The method for automatically extracting dispersion curves based on deep learning includes: Acquire station pair data; the station pair data is seismic waveform data recorded by at least two seismic stations; Performing cross-correlation processing on the data of the stations to obtain a plurality of cross-correlation data; Performing data processing on each of the cross-correlation data to obtain processed cross-correlation data; According to the processed cross-correlation data, a dispersion energy diagram is obtained by using a τ-p transformation technique; the dispersion energy diagram is used to show the variation law of the dispersion characteristics with the speed of periodic variation; Based on the deep learning model, image recognition is performed on the dispersion energy map to obtain a dispersion curve; the dispersion curve is used to show the change law of the dispersion characteristics with the speed of periodic change; the deep learning model adopts an encoder-decoder architecture; in the deep learning model, a convolution with a step size of two replaces the pooling layer, and a transposed convolution is used for upsampling.

2. The method for automatic extraction of dispersion curves based on deep learning according to claim 1, characterized in that: Performing data processing on each of the cross-correlation data to obtain processed cross-correlation data specifically includes: The data processing includes a superposition operation of each cross-correlation data.

3. The method for automatic extraction of dispersion curves based on deep learning according to claim 1, characterized in that: Based on the deep learning model, the dispersion energy map is image recognized to obtain the dispersion curve, which includes: Based on the deep learning model, the dispersion energy map is image recognized to obtain the dispersion probability map output by deep learning; Dispersion probability map based on deep learning output; Convert the dispersion probability map output by deep learning into a curve form to obtain several initial dispersion curves; Abnormal curves in each of the initial dispersion curves are eliminated to obtain a dispersion curve.

4. The method for automatic extraction of dispersion curves based on deep learning according to claim 1, characterized in that: The deep learning model includes an encoder part and a decoder part.

5. The method for automatic extraction of dispersion curves based on deep learning according to claim 4, characterized in that: The encoder part includes several encoding layers, and a channel attention module is set after each encoding layer.

6. The method for automatic extraction of dispersion curves based on deep learning according to claim 5, characterized in that: The decoder part includes several decoding layers, and a channel attention module is set after each decoding layer.

7. A dispersion curve automatic extraction system based on deep learning, characterized in that: include: A data acquisition module, used to acquire station pair data; the station pair data is seismic waveform data recorded by at least two seismic stations; A cross-correlation processing module, used for performing cross-correlation processing on the station pair data to obtain a plurality of cross-correlation data; A data processing module, used for performing data processing on each of the cross-correlation data to obtain processed cross-correlation data; A dispersion energy map construction module is used to obtain a dispersion energy map using τ-p transformation technology based on the processed cross-correlation data; the dispersion energy map is used to show the variation law of the dispersion characteristics with the speed of periodic variation; The dispersion curve extraction module is used to perform image recognition on the dispersion energy map based on the deep learning model to obtain the dispersion curve; the dispersion curve is used to show the change law of the dispersion characteristics with the periodic change speed; the deep learning model adopts the encoder-decoder architecture; in the deep learning model, the convolution with a step size of two replaces the pooling layer, and the transposed convolution is used for upsampling.

8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a method for automatically extracting dispersion curves based on deep learning as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for automatically extracting dispersion curves based on deep learning described in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method for automatically extracting dispersion curves based on deep learning described in any one of claims 1 to 6 is implemented.

Images