Seismic body wave detection method and device based on dual-path deep learning, and medium

Abstract

The invention discloses a seismic body wave detection method and device based on dual-path deep learning, and a medium. An initial feature projection module is arranged to perform preliminary feature conversion and dimension reduction on an input waveform image and an F-K spectrum image; a dual-path feature extraction module is arranged, and deep features are synchronously extracted from two different forms of input data of waveforms and frequency spectrums; a multi-level feature fusion module is arranged and is responsible for effectively integrating features extracted by double paths; setting a classification output module to receive the fully fused advanced semantic features, firstly compressing a feature map into a feature vector through a global average pooling layer, then performing calculation through a full connection layer, and finally outputting probability distribution of the seismic slice belonging to the categories of body waves, mixed noise or surface waves. And automatic identification and classification of the signals are completed. According to the invention, high-precision and high-efficiency body wave signal detection is realized through a dual-path deep learning architecture and by utilizing complementary characteristics of waveforms and F-K frequency spectrums.

Description

Technical Field

[0001] This invention relates to the field of seismic body wave detection technology, specifically to a seismic body wave detection method, device, and storage medium based on dual-path deep learning. Background Technology

[0002] Seismic imaging is a key technology for studying subsurface structures, advancing seismological research, and improving earthquake prediction capabilities. Traditional seismic data acquisition methods mainly include active source seismic exploration (ASSE) and passive source seismic exploration (PSSE). Compared with ASSE, PSSE has advantages such as lower cost, less environmental impact, higher safety, and the ability to generate large amounts of data, thus becoming an important direction for future algorithm and technological innovation.

[0003] In passive source seismic exploration, body waves are crucial for deep Earth imaging due to their ability to penetrate deep layers and provide higher resolution and sensitivity to discontinuities. However, passive source data is often dominated by noise such as surface waves, and body wave signals are weak and difficult to extract. Traditional body wave detection methods mainly rely on seismic interferometry techniques (such as cross-correlation, cross-coherence, and multidimensional deconvolution) to reconstruct virtual sources, but these methods have the following drawbacks: 1. The virtual shot gather has poor quality, limited by uneven source distribution, false reflections caused by artificial noise, and interference from surface wave energy; 2. Traditional methods (such as lighting diagnostics, FK filtering, and manual visual inspection) rely on manual parameter adjustment and subjective judgment, which are inefficient and produce ambiguous results; 3. Existing deep learning methods are mostly limited to array data, which cannot be applied to short-profile volume wave tasks. They also rely on manual frequency range adjustment, resulting in low computational efficiency and insufficient data utilization.

[0004] Therefore, how to achieve automated, high-precision, and scalable volume wave signal detection to improve the quality of passive source depth reflection imaging is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] The present invention proposes a seismic body wave detection method, device and storage medium based on dual-path deep learning, which can at least solve one of the technical problems in the background art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A seismic body wave detection method based on dual-path deep learning executes the following steps via computer equipment: By setting up an initial feature projection module, a dual-path feature extraction module, a multi-level feature fusion module, and a classification output module; The initial feature projection module is used to perform preliminary feature transformation and dimensionality reduction on the input waveform image and FK spectrum image, laying the foundation for subsequent deep feature extraction. The dual-path feature extraction module simultaneously extracts complementary and discriminative deep features from input data of two different forms: waveform and spectrum. The multi-level feature fusion module is responsible for effectively integrating the features extracted from the two paths to make full use of the advantages of the two types of features. The classification output module receives fully fused high-level semantic features. First, it compresses the feature map into a feature vector through a global average pooling layer. Then, it performs calculations through a fully connected layer and finally outputs the probability distribution of the seismic slice belonging to categories such as body waves, mixed noise, or surface waves, thus completing the automatic identification and classification of the signal.

[0007] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0008] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0009] As can be seen from the above technical solution, the seismic body wave detection method based on dual-path deep learning of the present invention solves the technical problems of low automation, poor accuracy, reliance on manual intervention, and insufficient generalization ability of existing deep learning models in passive seismic body wave detection. By utilizing a dual-path deep learning architecture and simultaneously leveraging the complementary features of waveform and FK spectrum, high-precision and high-efficiency body wave signal detection is achieved.

[0010] The beneficial effects of this invention are: This invention automates body wave detection, significantly improving the efficiency of processing large-scale passive seismic data; it improves detection accuracy (body wave detection precision reaches 87%, recall reaches 90%) through dual-path input and attention feature fusion; and it enhances generalization ability by training the model using a physically guided hybrid dataset (real data and simulated data). This invention can extract high-quality body wave signals from low signal-to-noise ratio data, support reflection wave field reconstruction up to 16-second two-way travel time (TWT), clearly reveal deep structures (such as the double Moho), and provide reliable technical support for deep geological research. Attached Figure Description

[0011] Figure 1 This is a flowchart of the method of the present invention; Figure 2This refers to a low signal-to-noise ratio data segment in an embodiment of the present invention; Figure 3 This is a high bulk signal-to-noise ratio data segment according to an embodiment of the present invention; Figure 4 This is a high surface wave signal-to-noise ratio data segment according to an embodiment of the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0013] like Figure 1 As shown, this invention provides a seismic body wave detection model and method based on dual-path deep learning. Addressing the technical challenges of weak body wave signals in passive seismic data, which are masked by noise and difficult to detect automatically and accurately, the core of this solution lies in constructing a dedicated dual-path deep learning model (system) and defining its complete process (method) for processing seismic data to detect body waves.

[0014] The seismic body wave detection system based on dual-path deep learning described in this invention is a specially designed and trained dual-path deep neural network. For example... Figure 1 As shown, the system mainly includes an initial feature projection module, a dual-path feature extraction module, a multi-level feature fusion module, and a classification output module.

[0015] The initial feature projection module is mainly used to perform preliminary feature transformation and dimensionality reduction on the input waveform image and FK spectrum image, laying the foundation for subsequent deep feature extraction. This module performs the same operation on both input paths: first, it uses a 7x7 convolution kernel to perform convolution operation, initially extracting local features and mapping the input channels to a 64-dimensional feature space; then, it uses a 3x3 max pooling operation to downsample, halving the spatial size of the feature map while retaining key features.

[0016] The dual-path feature extraction module is the core of the entire system, designed to simultaneously extract complementary and discriminative deep features from input data of two different forms: waveform and spectrum. This module consists of parallel DenseNet and ResNet paths, containing four sequentially connected feature extraction stages. Each stage contains the two parallel convolutional blocks mentioned above, each block being composed of repeatedly stacked base convolutional layers. The number of repetitions in each stage is configured as [3, 6, 12, 8], and the number of output channels is configured as [64, 128, 256, 512], thereby achieving channel augmentation and spatial dimensionality reduction of the feature map.

[0017] The DenseNet path employs a dense connection approach, concatenating the output of each layer with the features of all preceding layers along the channel dimension to achieve high feature reuse and smooth information flow. Its basic units consist of 1x1 convolutions, 3x3 convolutions, and 1x1 convolutions, respectively.

[0018] The ResNet path employs a skip connection approach, directly adding the block's output to the input to effectively mitigate the vanishing gradient problem during deep network training, stabilizing the training process and ensuring the model can learn effectively. Its basic units sequentially include a 1x1 convolution for dimensionality reduction and feature fusion, a 3x3 convolution for core spatial feature extraction, and a 1x1 convolution for channel expansion.

[0019] The multi-level feature fusion module is primarily responsible for effectively integrating the features extracted from both paths to comprehensively utilize the advantages of both types of features. After the first three feature extraction stages, the system employs a standard transition layer for feature fusion. This transition layer consists of a 1x1 convolution and a 3x3 max pooling operation in sequence. The 1x1 convolution is used to fuse features and control the number of channels, while the max pooling is used to reduce the spatial size of the feature map to improve computational efficiency. After the final fourth feature extraction stage, the standard transition layer is replaced by an attention-based feature fusion module. This module utilizes a channel attention mechanism to dynamically calculate the importance weight of each channel in the dual-path output feature map and performs weighted fusion accordingly. This adaptively highlights the feature information that is more critical to the current classification task, achieving more refined and intelligent feature integration.

[0020] The classification output module receives fully fused high-level semantic features. First, it compresses the feature map into a feature vector through a global average pooling layer. Then, it performs calculations through a fully connected layer and finally outputs the probability distribution of the seismic slice belonging to categories such as body waves, mixed noise, or surface waves, thus completing the automatic identification and classification of the signal.

[0021] It should be noted that the system architecture and workflow described above are merely illustrative and do not limit the scope of protection of this invention. In practical applications, those skilled in the art can modify or replace the specific implementation methods, network depths, parameter configurations, and processing steps of the modules according to actual needs, and no restrictions are imposed here.

[0022] The present invention provides a seismic body wave detection method based on dual-path deep learning, comprising the following steps: S1. Data preprocessing; S2. Model classification; S3. Wavelength reconstruction.

[0023] In step S1, data preprocessing refers to preparing and transforming the original continuous passive seismic record. Specifically, this includes: cutting the continuous record into multiple fixed-length short-time-window slices; performing inter-trace normalization on the waveform data within each slice to compensate for amplitude differences between different receivers; applying a bandpass filter to suppress low-frequency environmental noise and high-frequency instrument noise while preserving effective body wave signal frequency bands; and simultaneously performing a spatiotemporal Fourier transform on each slice of data to generate a corresponding frequency-wavenumber spectrum (FK) image. Finally, paired waveform images and FK spectrum images are generated for each segment of seismic data.

[0024] In step S2, model classification refers to inputting the preprocessed paired waveform image and FK spectrum image into the aforementioned dual-path deep learning model for automatic identification and classification. Through a series of calculations, including initial feature projection, dual-path feature extraction, and multi-level feature fusion, the model ultimately outputs the probability that the seismic slice belongs to the category of body wave, surface wave, or noise.

[0025] In step S3, wavefield reconstruction refers to using the results output from the model classification stage to filter out all high signal-to-noise ratio data segments identified as "body waves". Based on these high-quality segments, a virtual shot gather is reconstructed using seismic interferometry techniques (such as cross-correlation). After conventional seismic imaging processing such as velocity analysis, normal time difference correction, and stacking, a high-resolution seismic reflection profile that can clearly reveal deep geological structures is finally generated.

[0026] The method of this invention was applied to actual data collected in the Songliao Basin, successfully extracting high signal-to-noise ratio data segments. The passive source data collection location was near the Songke-2 well in the Songliao Basin, China, with a survey line length of 2500m and a total of 50 geophones spaced 50m apart. Figure 2 This is a common mixed noise data segment in passively sourced data. Figure 3 4 represents data segments containing typical volume wave and surface wave characteristics, respectively. The data feature discrimination method is based on (jinzhongyuan et al., 2024). Volume wave characteristics are represented as a straight line on the waveform diagram, located in the middle of the spectrum diagram; surface wave characteristics are represented as a curve on the waveform diagram, located on both sides of the upper end of the spectrum diagram.

[0027] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0028] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0029] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the seismic body wave detection methods based on dual-path deep learning in the above embodiments.

[0030] It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

[0031] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0032] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0033] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0034] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A seismic body wave detection method based on dual-path deep learning, characterized in that, Perform the following steps using a computer device. By setting up an initial feature projection module, a dual-path feature extraction module, a multi-level feature fusion module, and a classification output module; The initial feature projection module is used to perform preliminary feature transformation and dimensionality reduction on the input waveform image and FK spectrum image, laying the foundation for subsequent deep feature extraction. The dual-path feature extraction module simultaneously extracts complementary and discriminative deep features from input data of two different forms: waveform and spectrum. The multi-level feature fusion module is responsible for effectively integrating the features extracted from the two paths to make full use of the advantages of the two types of features. The classification output module receives fully fused high-level semantic features. First, it compresses the feature map into a feature vector through a global average pooling layer. Then, it performs calculations through a fully connected layer and finally outputs the probability distribution of the seismic slice belonging to categories such as body waves, mixed noise, or surface waves, thus completing the automatic identification and classification of the signal.

2. The seismic body wave detection method based on dual-path deep learning according to claim 1, characterized in that: The initial feature projection module performs the same operation on both input paths: First, a 7x7 convolution kernel is used to perform convolution operations to initially extract local features and map the input channels to a 64-dimensional feature space. Then, a 3x3 max pooling operation is used for downsampling, which halves the spatial size of the feature map while retaining key features.

3. The seismic body wave detection method based on dual-path deep learning according to claim 2, characterized in that: The dual-path feature extraction module consists of parallel DenseNet and ResNet paths, and includes four sequentially connected feature extraction stages. Each stage contains the two parallel convolutional blocks mentioned above. Each convolutional block is formed by repeatedly stacking basic convolutional layers. The number of repetitions in each stage is configured as [3, 6, 12, 8], and the number of output channels is configured as [64, 128, 256, 512], thereby realizing channel augmentation and spatial dimensionality reduction of the feature map. The DenseNet path employs a dense connection approach, concatenating the output of each layer with the features of all preceding layers along the channel dimension to achieve high feature reuse and smooth information flow. Its basic units consist of 1x1 convolutions, 3x3 convolutions, and 1x1 convolutions, respectively. The ResNet path employs a skip connection approach, which effectively alleviates the gradient vanishing problem in deep network training by directly adding the output of the block to the input, stabilizing the training process and ensuring that the model can learn effectively. Its basic units include a 1x1 convolution for dimensionality reduction and feature fusion, a 3x3 convolution for core spatial feature extraction, and a 1x1 convolution for channel expansion.

4. The seismic body wave detection method based on dual-path deep learning according to claim 3, characterized in that: The multi-level feature fusion module uses a standard transition layer for feature fusion after the first three feature extraction stages. This transition layer consists of a one-to-one convolution and a three-to-three max pooling operation in sequence. The one-to-one convolution is used to fuse features and control the number of channels, while the max pooling is used to reduce the spatial size of the feature map to improve computational efficiency. After the final fourth feature extraction stage, the standard transition layer is replaced with an attention-based feature fusion module.

5. A seismic body wave detection method based on dual-path deep learning, characterized in that: Includes the following steps, S1, Data preprocessing; S2, Model classification; S3, Wavelength reconstruction; In step S1, data preprocessing refers to the preparation and transformation of the original continuous passive seismic records; In step S2, model classification refers to inputting the preprocessed paired waveform image and FK spectrum image into the dual-path deep learning model for automatic recognition and classification. In step S3, wavefield reconstruction refers to using the results output from the model classification stage to filter out all high signal-to-noise ratio data segments identified as "body waves"; based on these high-quality segments, a virtual shot gather is reconstructed using seismic interferometry technology, and then conventional seismic imaging processes such as velocity analysis, normal time difference correction, and overlay are performed to finally generate a high-resolution seismic reflection profile that can clearly reveal deep geological structures.

6. The seismic body wave detection method based on dual-path deep learning according to claim 5, characterized in that: S1 specifically includes, The continuous recording is cut into multiple fixed-length short time window slices; the waveform data in each slice is normalized between channels to compensate for amplitude differences between different receivers. A bandpass filter is applied to suppress low-frequency environmental noise and high-frequency instrument noise while preserving the effective volume wave signal band. Simultaneously, a spatiotemporal Fourier transform is performed on each slice of data to generate the corresponding frequency-wavenumber spectrum (FK) image. Finally, paired waveform images and FK spectrum images are generated for each segment of seismic data.

7. The seismic body wave detection method based on dual-path deep learning according to claim 5, characterized in that: In S2, the model performs a series of calculations, including initial feature projection, dual-path feature extraction, and multi-level feature fusion, to ultimately output the probability that the seismic slice belongs to the categories of body wave, surface wave, or noise.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the processor, it causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

Images