Seismic data abnormal trace identification method and device based on hybrid neural network

Abstract

The invention belongs to the technical field of seismic exploration data processing, and relates to a seismic data abnormal channel identification method and device based on a hybrid neural network, and the method comprises the steps: obtaining original single-shot seismic data; obtaining standardized single-shot seismic data according to the original single-shot seismic data; according to the standardized single-shot seismic data, obtaining first feature data, the first feature data including low-frequency trend feature data and spectral distribution feature data, or the first feature data including wavelet coefficient feature data and curvelet coefficient feature data; obtaining a first matrix according to the standardized single-shot seismic data and the first feature data; dividing the first matrix into a plurality of first data sub-blocks; obtaining an identification model according to the plurality of first data sub-blocks and a hybrid neural network model; obtaining to-be-identified seismic data; and according to the to-be-identified seismic data and the identification model, identifying an abnormal channel in the to-be-identified seismic data.

Description

Technical Field

[0001] This invention belongs to the field of seismic exploration data processing technology, and specifically relates to a method and apparatus for identifying seismic data anomalies based on a hybrid neural network. Background Technology

[0002] In related technologies, seismic exploration is a key link in the development of oil and gas resources. The quality of seismic data directly determines the reliability of subsurface structural interpretation and reservoir prediction, while anomaly identification is the core prerequisite for ensuring data quality.

[0003] As exploration targets extend into deeper and more complex geological structures, anomaly identification faces significant challenges. Traditional identification methods suffer from substantial deficiencies in feature extraction and analysis: statistical methods based on sliding windows can only capture local waveform abrupt changes but cannot identify global anomalies across traces; while global analysis methods based on inter-trace correlations can detect inter-trace relationships but are extremely insensitive to subtle local distortions. This disconnect between local and global features directly leads to a significant decrease in identification accuracy in complex scenarios, severely restricting the improvement of exploration accuracy. Against this backdrop, developing anomaly identification methods that can take into account both local features and global correlations has become an urgent need to improve the quality of seismic data processing. Summary of the Invention

[0004] In a first aspect, embodiments of the present invention provide a method for identifying seismic data anomalies based on a hybrid neural network, comprising: acquiring raw single-shot seismic data; acquiring standardized single-shot seismic data based on the raw single-shot seismic data; acquiring first feature data based on the standardized single-shot seismic data, wherein the first feature data includes low-frequency trend feature data and spectral distribution feature data, or the first feature data includes wavelet coefficient feature data and curvelet coefficient feature data; acquiring a first matrix based on the standardized single-shot seismic data and the first feature data; dividing the first matrix into multiple first data sub-blocks; acquiring an identification model based on the multiple first data sub-blocks and a hybrid neural network model; acquiring seismic data to be identified; and identifying anomalies in the seismic data to be identified based on the seismic data to be identified and the identification model.

[0005] In a second aspect, embodiments of the present invention provide a device for identifying seismic data anomalies based on a hybrid neural network, comprising: a first acquisition unit for acquiring raw single-shot seismic data; a first processing unit for acquiring standardized single-shot seismic data based on the raw single-shot seismic data; a second processing unit for acquiring first feature data based on the standardized single-shot seismic data, wherein the first feature data includes low-frequency trend feature data and spectral distribution feature data, or the first feature data includes wavelet coefficient feature data and curvelet coefficient feature data; a third processing unit for acquiring a first matrix based on the standardized single-shot seismic data and the first feature data; a fourth processing unit for dividing the first matrix into multiple first data sub-blocks; a fifth processing unit for acquiring an identification model based on the multiple first data sub-blocks and a hybrid neural network model; a second acquisition unit for acquiring seismic data to be identified; and a sixth processing unit for identifying anomalies in the seismic data to be identified based on the seismic data to be identified and the identification model.

[0006] Thirdly, embodiments of the present invention provide a storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of the method of the first aspect.

[0007] Fourthly, embodiments of the present invention provide an electronic device, including 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 the steps of the method of the first aspect.

[0008] The beneficial effects of this invention are as follows: The proposed method for identifying seismic anomalies based on hybrid neural networks first acquires raw single-shot seismic data, using this raw data as the starting point for processing. After acquiring the raw single-shot seismic data, this invention generates standardized single-shot seismic data based on it. The purpose of this process is to eliminate energy magnitude differences between seismic traces caused by variations in acquisition conditions, thereby establishing a unified benchmark for subsequent feature extraction and model analysis.

[0009] After obtaining standardized single-shot seismic data, this invention acquires first feature data based on the standardized data. The first feature data refers to mathematical results extracted from the standardized data that characterize the essential attributes of the signal. Specifically, the first feature data includes low-frequency trend feature data and spectral distribution feature data, or wavelet coefficient feature data and curvelet coefficient feature data. The core function of this process is to perform multi-dimensional feature fusion, that is, to enrich the characterization of the seismic trace from multiple attribute levels of the signal, thereby enhancing the feature difference between anomalous and normal traces.

[0010] Then, based on the standardized single-shot seismic data and the first feature data, the first matrix is ​​obtained. The purpose of this step is to concatenate the original standardized data with the extracted multiple feature data along the channel dimension to construct a more comprehensive and discriminative fusion matrix.

[0011] Dividing the first matrix into multiple first data sub-blocks not only adapts to the input size requirements of neural network models, but more importantly, preserves the spatial relationships between channels within each sub-block, creating conditions for the model to learn local features and global correlations within the sub-blocks simultaneously.

[0012] Subsequently, a recognition model is obtained based on multiple first data sub-blocks and a hybrid neural network model. The hybrid neural network model refers to a network structure capable of simultaneously processing local features and global correlations. The model training process utilizes labeled abnormal and normal channel sample data to teach the hybrid neural network model how to identify abnormal channels from these data sub-blocks. This invention, through the design of the recognition model, can simultaneously capture local detailed features such as waveform abrupt changes within channels as well as global correlation anomalies between channels, thereby achieving collaborative analysis and accurate identification of multi-scale features of abnormal channels.

[0013] Finally, in the application phase, the seismic data to be identified is acquired and input into the pre-trained identification model. The model outputs the identification result of whether each trace in the data to be identified is an anomalous trace, thus completing the automated identification.

[0014] In summary, the seismic data anomaly identification method based on hybrid neural networks proposed in this invention constructs a rich feature representation system by simultaneously utilizing information such as original standardized data, low-frequency trend features, and spectral distribution features. This multi-channel input structure ensures that even in high-noise environments, the distortion features of anomalies can be reflected across multiple feature dimensions, reducing the risk of missed detections due to single features being overwhelmed by noise. Compared to traditional energy thresholding methods, this invention significantly improves the robustness and reliability of feature extraction in low signal-to-noise ratio environments.

[0015] This invention significantly enhances the distinguishability of anomalies in the feature space through rich input features constructed by feature fusion, laying a solid foundation for subsequent accurate identification. This invention employs a hybrid neural network model to process block data, effectively solving the core problem of the separation between local and global features in traditional methods. It can synergistically utilize two key types of information—waveform detail abrupt changes and inter-trace spatial correlations—for analysis. This invention achieves high-precision identification of complex anomalies, effectively overcoming the problems of low identification accuracy, serious missed detections and misjudgments caused by insufficient feature utilization and a single analysis mode in existing technologies, providing a reliable guarantee for seismic data quality control and subsequent processing. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a method for identifying seismic data anomalies based on a hybrid neural network according to an embodiment of the present invention; Figure 2 A schematic block diagram illustrating the structure of a seismic data anomaly identification device based on a hybrid neural network according to an embodiment of the present invention; Figure 3 A schematic block diagram illustrating the structure of an electronic device according to an embodiment of the present invention; Figure 4 This describes an observation system designed according to the geological requirements of the exploration area, as described in this embodiment of the invention. Figure 5 This represents the actual single-shot data for a certain work area according to an embodiment of the present invention; Figure 6 This represents the single-shot data after abnormal channel processing in this embodiment of the invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0018] like Figure 1 As shown, in an embodiment of the present invention, a method for identifying seismic data anomalies based on a hybrid neural network is proposed, comprising: S102: Acquire raw single-shot seismic data; S104: Obtain standardized single-shot seismic data based on the original single-shot seismic data; S106: Based on standardized single-shot seismic data, obtain the first characteristic data, wherein the first characteristic data includes low-frequency trend characteristic data and spectral distribution characteristic data, or the first characteristic data includes wavelet coefficient characteristic data and curve wave coefficient characteristic data; S108: Obtain the first matrix based on standardized single-shot seismic data and the first feature data; S110: Divide the first matrix into multiple first data sub-blocks; S112: Obtain the recognition model based on multiple first data sub-blocks and a hybrid neural network model; S114: Acquire earthquake data to be identified; S116: Identify anomalous traces in the seismic data to be identified based on the seismic data to be identified and the identification model.

[0019] In this embodiment, the present invention first acquires raw single-shot seismic data, using this raw single-shot seismic data as the starting point for processing. Specifically, raw single-shot seismic data refers to the original dataset collected in the field that contains records of seismic waves generated by a single excitation.

[0020] Next, standardized single-shot seismic data is obtained from the original single-shot seismic data. The purpose of this step is to normalize the amplitude value of each seismic data to a uniform range through linear scaling and other methods, eliminate the energy magnitude differences between traces caused by differences in acquisition conditions, and establish a unified benchmark for subsequent feature extraction and model analysis.

[0021] Next, the first feature data is obtained based on the standardized data. Specifically, the first feature data refers to the mathematical transformation results extracted from the standardized data that characterize the essential attributes of the signal. This includes low-frequency trend feature data and spectral distribution feature data, or wavelet coefficient feature data and curvelet coefficient feature data. The core function of this process is to perform multi-dimensional feature fusion, that is, to enrich the characterization of the seismic trace from multiple attribute levels of the signal, thereby enhancing the feature differences between anomalous and normal traces.

[0022] Then, based on the standardized single-shot seismic data and the first feature data, a first matrix is ​​obtained. This step involves concatenating the original standardized data with the extracted multiple feature data along the channel dimension to construct a more comprehensive and discriminative multi-channel fusion matrix. This multi-channel input structure provides rich and complementary feature information for the subsequent hybrid neural network.

[0023] Dividing the first matrix into multiple first data sub-blocks not only adapts to the input size requirements of neural network models, but more importantly, preserves the spatial relationships between channels within each sub-block, creating conditions for the model to learn local features and global correlations within the sub-blocks simultaneously.

[0024] Subsequently, a recognition model is obtained based on multiple first data sub-blocks and a hybrid neural network model. A hybrid neural network model refers to a network structure capable of simultaneously processing local features and global correlations. The model training process utilizes labeled abnormal and normal channel sample data to teach the hybrid neural network model how to identify abnormal patterns from these data sub-blocks.

[0025] This invention, through the design of an identification model, can simultaneously capture local detailed features such as waveform abrupt changes within a channel as well as global correlation anomalies between channels, thereby achieving collaborative analysis and accurate identification of multi-scale features of abnormal channels.

[0026] Finally, in the application phase, the seismic data to be identified is acquired and input into the pre-trained identification model. The model outputs the identification result of whether each trace in the data to be identified is an anomalous trace, thereby completing the automated identification. Specifically, in this invention, the seismic data to be identified is also single-shot seismic data.

[0027] The proposed method for identifying seismic data anomalies based on hybrid neural networks constructs a rich feature representation system by simultaneously utilizing multi-source information, including original standardized data, low-frequency trend features, and spectral distribution features. The multi-channel input structure ensures that even in high-noise environments, the distortion features of anomalies are reflected across multiple feature dimensions, reducing the risk of missed detections due to single features being overwhelmed by noise. Compared to traditional energy thresholding methods, this invention significantly improves the robustness and reliability of feature extraction in low signal-to-noise ratio environments.

[0028] This invention significantly enhances the distinguishability of anomaly traces in the feature space through rich input features constructed by feature fusion, laying a solid foundation for subsequent accurate identification. This invention employs a hybrid neural network model to process block data, effectively solving the core problem of the separation between local and global features in traditional methods, and can synergistically utilize two key types of information: waveform detail abrupt changes and inter-trace spatial correlations. This invention achieves high-precision identification of complex anomaly traces, effectively overcoming the problems of low identification accuracy, serious missed detections and misjudgments caused by insufficient feature utilization and a single analysis mode in existing technologies, providing a reliable guarantee for seismic data quality control and subsequent processing.

[0029] In some embodiments of the present invention, optionally, standardized single-shot seismic data is obtained based on the original single-shot seismic data, specifically including: normalizing the original single-shot seismic data to obtain standardized single-shot seismic data.

[0030] In this embodiment, the step of obtaining standardized single-shot seismic data from the raw single-shot seismic data specifically involves normalizing the raw single-shot seismic data to obtain standardized single-shot seismic data. This invention performs normalization processing on the acquired raw single-shot seismic data, also commonly referred to as standardization processing, to eliminate the differences in dimensions and scales inherent in the data itself.

[0031] Specifically, normalization is a data preprocessing technique that aims to scale raw data to a uniform, standard numerical range through mathematical transformations.

[0032] Raw single-shot seismic data refers to the collection of seismic wave records received from a single excitation, acquired in the field without any processing. During actual acquisition, due to factors such as the acquisition environment and instrumentation, the signals received from different seismic traces may differ by orders of magnitude in absolute energy. This difference is not geologically significant but can severely interfere with the stability and effectiveness of subsequent feature extraction and model training. This invention uses normalization processing to unify the amplitude values ​​of all seismic traces to the same scale, for example, by linearly scaling to the [0, 1] interval, effectively eliminating the energy magnitude differences between traces.

[0033] Specifically, the normalization or standardization process can be achieved by calculating the maximum and minimum amplitude values ​​of all sampling points in the entire single-shot data, and then using a formula to calculate the value of each sampling point in each seismic trace, thereby linearly mapping each data trace to the interval [0, 1], and finally obtaining standardized single-shot seismic data with consistent energy levels.

[0034] In some embodiments of the present invention, optionally, when the first feature data includes low-frequency trend feature data and spectral distribution feature data, the first feature data is obtained based on standardized single-shot seismic data, specifically including: performing median filtering calculation on the standardized single-shot seismic data to obtain low-frequency trend feature data; and performing Fourier transform on the standardized single-shot seismic data to obtain spectral distribution feature data.

[0035] In this embodiment, the present invention extracts low-frequency trend features and spectral distribution features through median filtering and Fourier transform. Specifically, median filtering is a nonlinear signal processing technique that smooths a signal by taking the median value within a sliding window, effectively suppressing impulse noise and preserving signal edges. Fourier transform is a mathematical tool for transforming a signal from the time domain to the frequency domain.

[0036] Median filtering is performed on standardized single-shot seismic data to obtain low-frequency trend feature data; Fourier transform is then performed on the standardized single-shot seismic data to obtain spectral distribution feature data. This invention extracts deep features effective for identifying anomalies from the standardized data. Median filtering effectively separates the low-frequency trend components in the signal, while the spectral distribution features obtained from the Fourier transform reflect the energy distribution of the signal along the frequency axis. These two features provide complementary information from the time and frequency domains, enriching the feature dimensions used for anomaly identification and improving the feature representation capability.

[0037] Specifically, median filtering of the standardized single-shot data can be performed by setting a sliding window with a window size of 5, traversing all sampling points of each channel, and replacing the value of the current point with the median value of the sampling points within the window, ultimately obtaining low-frequency trend feature data that reflects the low-frequency change trend of the signal.

[0038] In some embodiments of the present invention, optionally, when performing median filtering calculation on standardized single-shot seismic data, the window size for median filtering is 3 to 9.

[0039] In this embodiment, the range of the window size used for median filtering in median filtering calculation of standardized single-shot seismic data is defined. Specifically, the window size refers to the number of adjacent sampling points considered in calculating the median during median filtering. This invention limits the window size ω to the range of 3 to 9, achieving an optimal balance between effectively extracting low-frequency trends and preserving necessary signal details, ensuring the reliability and effectiveness of the extracted features.

[0040] Specifically, values ​​of 3, 5, 7, or 9 can be selected as the window length. When the data noise scale is large, it can be dynamically adjusted to 7 or 9 to enhance the smoothing effect. When more details need to be retained, a smaller size of 3 can be selected.

[0041] Specifically, the default window size for median filtering is 5.

[0042] In some embodiments of the present invention, optionally, when the first feature data includes wavelet coefficient feature data and curvelet coefficient feature data, the first feature data is obtained based on standardized single-shot seismic data, specifically including: performing wavelet transform on the standardized single-shot seismic data to obtain wavelet coefficient feature data; and performing curvelet transform on the standardized single-shot seismic data to obtain curvelet coefficient feature data.

[0043] In this embodiment, another feature extraction scheme is provided, namely, obtaining wavelet coefficient features and curvelet coefficient features through wavelet transform and curvelet transform. Specifically, wavelet transform is a time-frequency analysis method that can provide localized information of the signal in both time and frequency. Curvelet transform is a multi-scale geometric analysis tool that can represent high-dimensional signals with line singularities or surface singularities.

[0044] When the noise in the exploration area is complex and it is difficult to effectively distinguish anomalies based solely on time-domain or frequency-domain features, wavelet and curve transform can extract more discriminative time-frequency domain and geometric structure features, providing the model with richer and more adaptable analytical basis for complex noise scenarios, thus enhancing the adaptability and flexibility of the method.

[0045] In some embodiments of the present invention, optionally, obtaining a first matrix based on standardized single-shot seismic data and first feature data specifically includes: obtaining first matrix data based on standardized single-shot seismic data; obtaining second matrix data and third matrix data based on the first feature data; and obtaining the first matrix based on the first matrix data, second matrix data, and third matrix data.

[0046] In this embodiment, the first matrix is ​​specifically a three-dimensional data structure, which can be compared to a multi-channel image, where each pixel position contains values ​​from multiple channels.

[0047] In acquiring the first matrix, this invention obtains the first matrix data based on standardized single-shot seismic data. This process can be viewed as using the original standardized data as the first feature matrix. Based on the first feature data, the second and third matrix data are then acquired. This process can be viewed as using the low-frequency trend feature data extracted by median filtering as the second feature matrix, and the spectral distribution feature data extracted by Fourier transform as the third feature matrix.

[0048] The first matrix is ​​obtained based on the first, second, and third matrix data. The data structures of the first, second, and third matrix data are identical; each matrix is ​​a two-dimensional matrix, specifically the number of traces multiplied by the number of sampling points. Each matrix represents a feature image describing the same batch of seismic traces from different perspectives. Obtaining the first matrix involves stacking these multiple feature matrices along the channel dimension to form a three-dimensional data volume.

[0049] For example, if each channel of the given data has 128 samples, after median filtering and Fourier transform, the feature data for each channel still consists of 128 samples. This can be transformed into a feature matrix with 128 samples vertically, and the same number of seismic channels horizontally. By stacking the first, second, and third feature matrices along the channel dimension, three-dimensional input data that can be used to train a hybrid neural network model is formed. The trained network model requires that all input feature matrices be of the same size, thus enabling normal training of the hybrid neural network model.

[0050] This invention integrates standardized raw signals and their features in different transform domains into a single three-dimensional data structure. This multi-feature matrix stacking approach enables subsequent hybrid neural network models to learn and analyze data from multiple perspectives and dimensions. Each feature matrix provides a unique information perspective: the first matrix retains the original waveform features, the second matrix highlights the macroscopic trend changes of the signal, and the third matrix reflects the frequency distribution characteristics of the signal. This multi-feature fusion method enhances the richness and complementarity of the input information, enabling the model to understand the data more comprehensively and deeply, thereby more accurately capturing the subtle differences between abnormal and normal channels.

[0051] In some embodiments of the present invention, optionally, the first matrix is ​​divided into a plurality of first data sub-blocks, specifically including: dividing the first matrix into a plurality of first data sub-blocks according to a preset size, wherein the preset size is 128×128 to 1024×1024.

[0052] This embodiment defines a method for dividing a first matrix into smaller data sub-blocks and its parameter range. Specifically, dividing the first matrix into multiple first data sub-blocks in this invention includes dividing the first matrix into multiple first data sub-blocks according to a preset size. Specifically, the preset size refers to the size of each data sub-block in terms of both channel count and sampling point count. The first data sub-block is the small data block obtained after division. This invention divides the first matrix into smaller sub-blocks, providing flexibility for data of different scales and computational resources. Specifically, the preset sub-block size is generally 128×128~1024×1024, with a default of 256×256, balancing local feature integrity and computational efficiency.

[0053] In some embodiments of the present invention, optionally, after dividing the first matrix into multiple first data sub-blocks, the identification method further includes: storing the correspondence between each first data sub-block and the seismic trace in the original single-shot seismic data.

[0054] In this embodiment, after dividing the first matrix into multiple smaller first data sub-blocks, the present invention synchronously stores the correspondence between each first data sub-block and each seismic trace in the original single-shot seismic data. This correspondence records the specific location index of the seismic trace contained in each data sub-block in the original single-shot data.

[0055] Because of the block processing strategy, a single continuous seismic signal in the original single-shot seismic data may be divided into multiple spatially or temporally adjacent data blocks. If, after the model completes its predictions, the identification results of these blocks cannot be accurately mapped back to the trace location coordinates of the original data, then the entire anomaly trace identification process will lose its practical application value. By pre-storing detailed trace index correspondences, it is ensured that the subsequent model prediction results for each block can be losslessly and accurately restored and located in the coordinate system of the original single-shot seismic data.

[0056] In some embodiments of the present invention, identifying anomalous traces in the seismic data to be identified based on the seismic data to be identified and the identification model specifically includes: obtaining standardized seismic data to be identified based on the seismic data to be identified; obtaining second feature data based on the standardized seismic data to be identified; obtaining a second matrix based on the standardized seismic data to be identified and the second feature data; dividing the second matrix into multiple second data sub-blocks; obtaining an anomaly probability map for each second data sub-block based on the multiple second data sub-blocks and the identification model; obtaining a comprehensive anomaly probability for each seismic trace based on the anomaly probability map; and identifying anomalous traces in the seismic data to be identified based on the comprehensive anomaly probability and a judgment threshold.

[0057] In this embodiment, the present invention utilizes a trained recognition model to identify abnormal channels in the data to be identified.

[0058] Specifically, the anomaly probability map is the output of the model after predicting the input data sub-blocks. The comprehensive anomaly probability is a unified anomaly probability value obtained by fusing the probability information of each seismic trace across all relevant sub-blocks. Seismic traces with an anomaly probability exceeding the determination threshold are ultimately identified as anomalous traces.

[0059] The above process represents the core steps of the identification method proposed in this invention during application. During identification, the same preprocessing, feature extraction, dataset formation, and data segmentation steps are performed to ensure that the data to be identified undergoes a completely consistent processing flow with the training data, avoiding biases introduced by inconsistent data processing. By generating an anomaly probability map and fusing the probabilities of overlapping channels, the problem of multiple judgments for a single channel caused by block segmentation is effectively solved, and multi-contextual information is utilized to improve the robustness of the judgment. Finally, by setting a scientific judgment threshold, the final identification result is output, improving the accuracy of anomaly channel identification.

[0060] Specifically, this invention loads the trained model and generates multiple second data sub-blocks from the test data using the same preprocessing method as the original data. These second data sub-blocks are then input into the recognition model, which outputs an anomaly probability map for each sub-block. Based on the anomaly probability map, the comprehensive anomaly probability of each seismic trace is obtained. Specifically, this process can be understood as employing a maximum probability fusion strategy for overlapping traces across sub-blocks to eliminate probability differences caused by sub-block segmentation. Based on the comprehensive anomaly probability and a judgment threshold, anomalous traces in the seismic data to be identified are obtained. Specifically, an anomaly judgment threshold is set to 0.5; traces with probabilities exceeding the threshold are identified as anomalous traces, generating preliminary identification results.

[0061] Specifically, the decision threshold can be dynamically adapted based on the signal-to-noise ratio of the actual seismic data, with a value ranging from 0.1 to 1.0. The default selection is 0.5, and this benchmark value was determined through iterative testing in the early stages of model building. By using the harmonic mean of precision and recall as the evaluation index, multiple candidate thresholds were tested and screened, and 0.5 was ultimately determined as the optimal benchmark threshold.

[0062] In practical applications, threshold settings follow a noise-adaptive principle: when processing low signal-to-noise ratio (SNR) data, the decision threshold is appropriately lowered to below 0.5, such as 0.4, to improve recognition sensitivity and reduce missed detections of abnormal channels; when processing high SNR data, the decision threshold is appropriately raised to above 0.5, such as 0.6, to enhance discrimination specificity and reduce false positives of normal channels. It is important to note that this SNR-based threshold adjustment mechanism is primarily applied during the initial parameter calibration phase of model building. In actual industrial applications, once a suitable threshold for the target work area is determined through initial calibration, subsequent processing will use this determined threshold to maintain consistency in processing standards.

[0063] In some embodiments of the present invention, optionally, after obtaining the anomalous traces in the seismic data to be identified based on the comprehensive anomaly probability and the judgment threshold, the identification method further includes: outputting a list of anomalous traces containing trace indexes and anomaly markers based on the anomalous traces.

[0064] In this embodiment, after identifying anomalous channels, the present invention outputs a structured list of results. Specifically, the channel index is the position number of the anomalous channel in the original data. The anomaly tag is a label used to identify the channel as anomalous. Outputting a structured list greatly facilitates subsequent data quality control work, allowing processing personnel to quickly locate and review anomalous channels based on this list. Specifically, the optimized anomalous channel identification results are saved according to the shot set number, and the output format is a text file containing the channel index and anomaly tag.

[0065] In some embodiments of the present invention, optionally, an abnormal channel list containing channel indexes and abnormal markers is output based on the abnormal channels. Specifically, this includes: performing spatial distribution smoothing processing on the abnormal channels and outputting an abnormal channel list containing channel indexes and abnormal markers; wherein the spatial distribution smoothing processing is implemented using median filtering or Gaussian filtering, and when median filtering is used, the window size of the median filtering is 3 to 9.

[0066] In this embodiment, the present invention performs spatial distribution smoothing processing on the initially identified abnormal trace results to eliminate isolated misjudgments. Specifically, spatial distribution smoothing processing is an image post-processing technique applied here to the spatial distribution map of abnormal traces on the shot gather record, aiming to make the distribution of abnormal regions more reasonable. Median filtering and Gaussian filtering are two commonly used smoothing filters. Through optimized design of the abnormal trace results, the present invention effectively eliminates sporadic misjudgments caused by random noise, improving the reliability and spatial consistency of the final results.

[0067] In some embodiments of the present invention, optionally, after dividing the second matrix into multiple second data sub-blocks, the identification method further includes: storing the correspondence between each second data sub-block and the seismic trace in the seismic data to be identified; obtaining the comprehensive anomaly probability of each seismic trace according to the anomaly probability map, specifically including: obtaining the correspondence between each second data sub-block and the seismic trace in the seismic data to be identified; mapping the probability value of each second data sub-block in the anomaly probability map back to the corresponding seismic trace according to the correspondence; and obtaining the comprehensive anomaly probability of each seismic trace according to all probability values ​​mapped to the same seismic trace.

[0068] In this embodiment, the present invention sets forth clear steps for the process of fusing the anomaly probability maps of sub-blocks into a comprehensive anomaly probability for each seismic trace. First, while dividing the second matrix formed from the seismic data to be identified into multiple second data sub-blocks, the correspondence between each second data sub-block and each seismic trace in the seismic data to be identified is established and stored. This crucial step records the position index of the seismic traces contained in each data sub-block within the original data to be identified, ensuring that the abstract results output by the model can be accurately located to specific seismic traces in the subsequent process.

[0069] Subsequently, after the model outputs the anomaly probability map for each second data sub-block, the probability value of each second data sub-block in its anomaly probability map is precisely mapped back to the corresponding seismic trace in the seismic data to be identified, based on the pre-stored correspondence. Due to the block segmentation strategy, the same seismic trace may appear in multiple adjacent or overlapping second data sub-blocks, thus resulting in multiple probability prediction values ​​from different contexts.

[0070] Finally, based on all probability values ​​mapped to the same seismic trace, a data fusion strategy is used to obtain the final comprehensive anomaly probability of that trace. In this invention, by establishing and utilizing the precise mapping relationship between sub-blocks and seismic traces, it is ensured that the probability prediction values ​​obtained from a single trace in all sub-blocks containing it can be effectively collected. A maximum probability fusion strategy is preferred for judgment, i.e., the highest anomaly probability obtained for the same trace in different sub-blocks is taken as the comprehensive anomaly probability of that trace. This strategy helps to significantly reduce the false negative rate and effectively eliminates the probability differences that may be caused solely by the data sub-block segmentation method.

[0071] In some embodiments of the present invention, optionally, before obtaining the recognition model based on multiple first data sub-blocks and a hybrid neural network model, the recognition method further includes: constructing a hybrid neural network model; wherein the hybrid neural network model includes a feature extraction module composed of convolutional layers and window multi-head self-attention modules alternately connected.

[0072] In this embodiment, before obtaining the recognition model using the hybrid neural network model, the present invention further includes a process of constructing the hybrid neural network model. Specifically, in the present invention, before obtaining the recognition model based on multiple first data sub-blocks and the hybrid neural network model, the recognition method further includes constructing a hybrid neural network model comprising a feature extraction module consisting of convolutional layers and window multi-head self-attention modules alternately connected.

[0073] Specifically, a convolutional layer is a type of layer in a neural network that extracts local spatial features by sliding convolutional kernels across the data. A multi-head self-attention module is a variant of the self-attention mechanism that divides the input data into multiple non-overlapping windows, computes attention only within each window to reduce computational complexity, and utilizes multiple heads to attend to different aspects of features in parallel.

[0074] By designing the feature extraction module in a hybrid neural network model, this invention enables the collaborative capture of both local and global features. By alternately connecting convolutional layers with multi-head self-attention modules in a window, efficient fusion of local and global features is achieved, fundamentally solving the problem of local-global feature fragmentation.

[0075] In some embodiments of the present invention, the convolutional layer may optionally employ a 3×3 or 5×5 convolutional kernel, and the window size of the multi-head self-attention module may be 4×4 to 16×16.

[0076] In this embodiment, the parameter ranges of the convolutional layers and self-attention modules in the hybrid neural network are defined. Specifically, the convolutional kernel is a filter used for feature extraction in the convolutional layer, and its size determines the size of the receptive field. The window size defines the size of the local region divided during the computation of the self-attention mechanism. In this invention, the convolutional layers use 3×3 or 5×5 kernels, achieving a good balance between receptive field and computational cost, and effectively capturing the local correlations between adjacent traces and adjacent sampling points in seismic data. The window size of the self-attention module is between 4×4 and 16×16, ensuring that there are enough elements in each window for global relationship modeling, while avoiding a sharp increase in computational complexity due to excessively large windows. The above-mentioned optimized parameter range of this invention ensures that the hybrid neural network can capture fine local details and efficiently understand medium-range global structures.

[0077] Specifically, convolutional layers typically use 3×3 or 5×5 convolutional kernels, and the window size of the self-attention module is generally 4×4 to 16×16, with a default of 8×8, to balance local details and global relationships.

[0078] In some embodiments of the present invention, the hybrid neural network model is optionally a structure with a connected encoder and decoder, wherein the encoder performs downsampling by max pooling or stride convolution, and the decoder performs upsampling by transposed convolution or interpolation.

[0079] In this embodiment, the encoder and decoder structures of the hybrid neural network model are designed. The encoder gradually compresses the spatial size and extracts high-level abstract features through layer-by-layer downsampling. The decoder gradually restores the spatial resolution through layer-by-layer upsampling, and finally outputs a result that matches the input size.

[0080] In some embodiments of the present invention, optionally, a recognition model is obtained based on a plurality of first data sub-blocks and a hybrid neural network model, specifically including: setting the recognition logic of the hybrid neural network model based on a plurality of first data sub-blocks to obtain a recognition model.

[0081] In this embodiment, the present invention trains a hybrid neural network model to learn recognition logic using training data, thereby obtaining a final recognition model. In this invention, the recognition model is not pre-programmed with rules, but rather trained on a large number of labeled samples, automatically learning the patterns within them. The training process involves continuously adjusting the model parameters to obtain the final recognition model. This invention sets the recognition logic of the hybrid neural network model based on multiple first data sub-blocks, avoiding the limitations of traditional methods that rely on manually designed features and rules.

[0082] In some embodiments of the present invention, optionally, in the process of setting the recognition logic of the hybrid neural network model according to multiple first data sub-blocks, a weighted binary classification loss function is used, wherein the formula for calculating the weight coefficients is: ; in, These are the weighting coefficients for the abnormal channel category. This is data from abnormal channels. This is normal channel sample data. This is the sensitivity adjustment factor, and its value ranges from 1.2 to 2.0.

[0083] In this embodiment, the present invention employs a weighted binary classification loss function during model training. The weight coefficients are dynamically calculated using a specific formula to address the imbalance in the number of abnormal and normal samples in the training data.

[0084] Specifically, the weighted binary classification loss function is an improved form of the standard cross-entropy loss function, assigning different weights to the prediction errors of samples from different classes. Outlier trace sample data refers to the number of seismic trace samples in the training data that are marked as anomalous. Normal trace sample data refers to the number of seismic trace samples in the training data that are marked as normal. The sensitivity adjustment factor is an adjustable hyperparameter used to control the degree of attention given to outlier trace samples.

[0085] In some embodiments of the present invention, optionally, an adaptive gradient optimizer with a regularization term is used in the process of setting the recognition logic of the hybrid neural network model according to a plurality of first data sub-blocks, wherein the strength of the regularization term is set to 1e-6 to 1e-4.

[0086] In this embodiment, the present invention uses an adaptive gradient optimizer with integrated regularization term during model training to suppress overfitting through weight decay techniques. Specifically, the adaptive gradient optimizer is an optimization algorithm that can automatically adjust the learning rate of each parameter. The regularization term is an additional term added to the end of the loss function, typically proportional to the norm of the model weights, used to penalize excessively large weight values, thereby constraining the complexity of the model.

[0087] In this embodiment, the present invention penalizes excessively large weight values ​​in the model by adding a regularization term to the optimizer, enabling the model to learn simpler and more generalized feature patterns. Setting the regularization term strength between 1e-6 and 1e-4 can significantly improve its generalization performance without excessively compromising the model's fitting ability, ensuring that the trained recognition model maintains stable and accurate recognition capabilities when faced with seismic data from different work areas.

[0088] Specifically, the strength of the regularization term, ranging from 1e-6 to 1e-4, represents the degree to which the regularization term is effective during the training of a hybrid neural network model. Strength values ​​from 1e-6 to 1e-4 correspond to 0.000001 to 0.0001. A relatively small value of 1e-6 indicates a weaker constraint on the model parameters, with the model focusing more on fitting the training data. A relatively large value of 1e-4 indicates a stronger constraint effect from the regularization.

[0089] In some embodiments of the present invention, optionally, in the process of setting the recognition logic of the hybrid neural network model according to multiple first data sub-blocks, a learning rate adaptive adjustment strategy is adopted, and when the verification loss stagnates, the learning rate is decayed at a ratio of 0.3 to 0.7.

[0090] In this embodiment, the present invention employs an adaptive learning rate adjustment strategy during model training. When the model's performance on the validation set stagnates, the learning rate is automatically reduced by a preset ratio. Specifically, the learning rate is a parameter in the optimization algorithm that determines the step size of each update of the model parameters. Learning rate decay refers to the strategy of gradually reducing the learning rate during training, which helps the model converge to the optimal solution more precisely in the later stages of training.

[0091] This invention improves the stability of the training process by adjusting the learning rate through monitoring and validation loss. Specifically, the learning rate decay ratio is 0.3 to 0.7, for example, the learning rate decay ratio can be 0.3, 0.4, 0.5, 0.6, or 0.7, where the default learning rate decay ratio is 0.5.

[0092] In some embodiments of the present invention, optionally, a gradient pruning strategy is adopted in the process of setting the recognition logic of the hybrid neural network model according to a plurality of first data sub-blocks, wherein the threshold of gradient pruning is set to 0.5 to 2.0.

[0093] In this embodiment, the present invention employs a gradient pruning strategy during model training to limit the norm of the calculated gradient vector to a preset threshold range in order to prevent gradient explosion.

[0094] This invention ensures that the step size of each parameter update is within a controllable and reasonable range by setting a gradient clipping threshold, greatly enhancing the robustness of the training process and avoiding training crashes caused by gradient instability. Specifically, the gradient clipping threshold can be set to 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4, 1.5, 1.6, 1.7, 1.8, 1.9, or 2.0; the default threshold is 1.0.

[0095] Optionally, in some embodiments of the present invention, a method for identifying seismic data anomalies based on a hybrid neural network is proposed, comprising: Step 1: Excite and collect single-shot seismic data. The receiver line spacing is L meters, the channel spacing is R meters, each receiver line has M channels, each channel has N sampling points, and the number of receiver lines is S.

[0096] In this embodiment, step one is the data preparation stage, which aims to acquire raw, unprocessed seismic observation data through actual field operations. Specifically, excitation refers to the process of generating seismic waves through a source at the surface or underground. Acquisition and recording refers to the process of using a geophone array deployed on the surface to receive seismic waves reflected or refracted from the underground medium and converting them into electrical signals for digital recording. Single-shot seismic data refers to the set of seismic wave signals synchronously recorded by all geophone channels after a single excitation, constituting a complete data unit, usually represented as a two-dimensional matrix. Receiver line spacing refers to the vertical distance between two adjacent receiver lines. A receiver line is a linear array of multiple geophones arranged according to a certain pattern. Channel spacing refers to the horizontal distance between two adjacent geophones within the same receiver line. M channels per receiver line refers to the total number of geophones deployed on a single receiver line. N sampling points per channel refers to the number of points discretized and recorded on the time axis for each seismic signal, determining the recording time length and temporal resolution. S receiver lines refer to the total number of receiver lines deployed throughout the entire observation area.

[0097] Step 2: Standardization preprocessing is applied to the raw seismic data, linearly scaling each trace to the [0, 1] interval to eliminate energy magnitude differences between traces. Median filtering and Fourier transform are then performed on the standardized data to extract low-frequency trend features, spectral distribution features, and amplitude statistical features. Based on these multi-dimensional attribute features, a multi-channel input framework is constructed, concatenating the raw standardized data, filtered features, and spectral features along the channel dimension to form a matrix that integrates multiple features, significantly improving the feature difference between anomalous and normal traces.

[0098] Specifically, in step two, the window size is generally set to 3-9, with a default of 5. The window size is dynamically adjusted to adapt to different noise scales. Furthermore, the median filtering and Fourier transform in step two can be replaced with wavelet transform, curvelet transform, etc., to extract wavelet coefficients, curvelet coefficients, and other features, flexibly adapting to complex noise scenarios.

[0099] Step 3: The multi-channel feature matrix obtained in Step 2 is divided into blocks, which are divided into two-dimensional blocks according to the preset sub-block size. The original trace index corresponding to each sub-block is recorded synchronously to ensure that the subsequent model prediction results can be accurately mapped to the trace position of the original seismic data. Based on the preset model training and testing parameters, an execution framework is provided for model learning.

[0100] In step three, the multi-channel feature matrix refers to the composite data volume generated in step two, which includes the original standardized data and its derived features. Block processing refers to the process of cutting the multi-channel feature matrix into several smaller, more easily processed local data blocks according to a preset size. This operation is mainly to adapt to the input size limitations of deep learning models and effectively control memory usage during computation. The preset sub-block size refers to the size of each data sub-block in terms of both trace count and sampling point count, predefined before performing block processing. The two-dimensional sub-block is the basic data processing unit obtained after block processing. It is a two-dimensional data slice in the trace and time directions, containing all feature channels, and can be directly input into the subsequent neural network model for computation. The original trace index refers to the specific location number of the seismic traces contained within each generated two-dimensional sub-block in the original single-shot data, recorded during the block processing.

[0101] Specifically, the preset sub-block size in step three is generally 128×128~1024×1024, with a default of 256×256, balancing local feature integrity and computational efficiency. Specifically, the preset parameters in step three cover data paths, training control parameters, hyperparameters, and result management parameters, achieving standardized configuration of the end-to-end process.

[0102] Step 4: Construct a hybrid neural network model that integrates convolutional and self-attention layers. This model includes a feature extraction module, which uses a hybrid feature extraction unit that alternates between convolutional layers and multi-head self-attention windows. Convolutional operations capture local spatial features of the seismic data, such as waveform abrupt changes within traces, while self-attention captures global dependencies, such as anomalies in trace correlations. Regarding the encoder-decoder structure, the encoder progressively compresses the feature dimension through downsampling to extract high-level abstract features, while the decoder restores the spatial resolution through upsampling to map back to the original data scale. Intermediate layers use skip connections to fuse feature information at different scales, enhancing the representational ability of anomalous trace features.

[0103] In step four, this invention constructs a hybrid neural network model, the core of which consists of a feature extraction module and an encoding / decoding structure. The feature extraction module employs an alternating design of convolutional layers and windowed multi-head self-attention modules. This invention comprehensively utilizes information at different scales, ultimately significantly enhancing the model's ability to represent and identify various abnormal channel features.

[0104] Specifically, the convolutional layers in step four typically use 3×3 or 5×5 convolutional kernels, and the window size of the self-attention module is generally 4×4 to 16×16, with 8×8 as the default, balancing local details and global relationships.

[0105] Specifically, in step four, downsampling uses max pooling or stride convolution, upsampling uses transposed convolution or interpolation, and skip connections are achieved through feature concatenation or element-wise addition, flexibly adapting to different data features.

[0106] Step 5: Model training and optimization. An improved binary classification loss function is adopted to solve the problem of imbalance between the number of samples in abnormal and normal channels by dynamically adjusting the class weight parameters; an adaptive gradient optimizer with regularization term is used to suppress model overfitting through weight decay; an adaptive learning rate adjustment mechanism is introduced to automatically reduce the learning rate proportionally when the validation loss stagnates continuously; a gradient constraint strategy is adopted during training (specifically, a maximum gradient threshold is set) to prevent gradient explosion, and the model weights with the lowest validation loss are saved periodically.

[0107] Specifically, the category weight parameters are adjusted using a dynamic formula: ; in, These are the weighting coefficients for the abnormal channel category. This is data from abnormal channels. This is normal channel sample data. This is a sensitivity adjustment factor, with a value ranging from 1.2 to 2.0. Compared to fixed weights, this dynamic adjustment mechanism can adaptively enhance the learning ability for sparse and outlier samples.

[0108] Specifically, the regularization term strength in step five is 1e-6 to 1e-4, the learning rate decay ratio is 0.3 to 0.7, and the default is 0.5. The maximum gradient threshold is 0.5 to 2.0, and the default is 1.0. The parameters are optimized collaboratively to balance training stability and convergence efficiency.

[0109] Step Six: Model Prediction and Preliminary Anomaly Channel Identification: Load the trained model and generate multi-channel sub-blocks from the test data using the preprocessing method in Step Six, maintaining consistency with the training data format. Input the sub-blocks into the model and output the anomaly probability map for each sub-block. For overlapping channels across sub-blocks, employ a maximum probability fusion strategy, i.e., take the highest probability of the same channel in different sub-blocks to eliminate probability differences caused by sub-block segmentation. Set an anomaly judgment threshold, and identify channels with probabilities exceeding the threshold as anomalies, generating preliminary identification results.

[0110] Specifically, the test data preprocessing in step six strictly matches the training stage, i.e., the sub-block size is 128×128~1024×1024 and the feature combination rules are consistent, so as to avoid the distribution deviation from affecting the recognition accuracy.

[0111] Specifically, in step six, the anomaly detection threshold is dynamically adapted to a signal-to-noise ratio of 0.4 to 0.6, with a default of 0.5. When the signal-to-noise ratio is low, it is lowered to 0.4 to reduce missed detections, and when the signal-to-noise ratio is high, it is raised to 0.6 to reduce false positives, thus achieving noise adaptive recognition.

[0112] Step 7: Dynamically adjust the model parameters based on the actual number of channels and noise level of the data. When the signal-to-noise ratio of the data is lower than the set value, increase the attention of the loss function to difficult-to-identify samples and increase the model protection factor to reduce the missed detection of abnormal channels. Iterate and test different decision thresholds, and use the harmonic mean of precision and recall as the evaluation index to select the optimal threshold. Post-process the identification results: smooth the spatial distribution of abnormal channels by median filtering to eliminate misjudgments caused by isolated noise points.

[0113] Specifically, the model protection factor in step seven ranges from 0.5 to 0.9, with a default value of 0.7. Its value increases as the signal-to-noise ratio decreases, thereby reducing missed detections by strengthening the focus on weak anomaly features.

[0114] Specifically, the test range for the judgment threshold in step seven is 0.4~0.6, with a default value of 0.5. It can be dynamically selected based on the sparsity of abnormal channels in the data. In sparse scenarios, the threshold can be appropriately reduced to capture more potential anomalies.

[0115] Specifically, the window size of the median filter in step seven is generally 3 to 9, with a default value of 3. The continuity correction of the spatial distribution of abnormal channels is achieved through neighborhood information fusion. Alternatively, smoothing methods such as Gaussian filtering can be used instead, and the choice can be made flexibly according to the type of data noise.

[0116] Step 8: Save the optimized abnormal track identification results by shot set number, and output them as a text file containing track indexes and abnormal markers; calculate the identification accuracy and recall by comparing with the manually labeled results to verify the algorithm's stability and ensure that the abnormal track identification accuracy is not lower than [previous value]. This meets the needs of exploration data processing.

[0117] Specifically, in step eight The preset minimum accuracy standard is generally 95%, which can be dynamically adjusted according to data quality and production needs. For example, it can be increased for high-precision exploration scenarios and appropriately reduced for rapid processing scenarios, taking into account both recognition reliability and adaptability to actual production.

[0118] In this embodiment, the invention first acquires single-shot seismic data, recording parameters such as receiver spacing and trace spacing. The raw data is standardized, and multi-dimensional features such as low-frequency trends and spectral distribution are extracted using median filtering and Fourier transform. These features are then concatenated to construct a multi-channel matrix. The feature extraction method can be replaced with wavelet transform or other methods adapted to complex noise. Next, the multi-channel matrix is ​​divided into two-dimensional sub-blocks, and trace indices are recorded. Training and testing parameters are configured. A hybrid neural network fusing convolutional and self-attention layers is constructed. The convolutional layer captures local waveform abrupt changes, while the self-attention module captures global inter-trace correlations. The encoding / decoding structure achieves feature mapping through downsampling compression and upsampling recovery. Subsequently, an improved binary classification loss function is used to dynamically adjust class weights to address sample imbalance. A regularized optimizer is used to train the model, and learning rate adjustment and gradient constraints ensure stability. Sub-blocks of the same format are generated for the test data and input into the model to obtain anomaly probability maps. Overlapping traces are fused with the highest probability, and the threshold is dynamically adjusted based on the signal-to-noise ratio to generate preliminary results. Finally, the model parameters are dynamically adjusted, the optimal threshold is selected, and the results are smoothed using median filtering. The results, including trace indices and labels, are saved to verify accuracy, providing a high-quality foundation for seismic data processing. This invention integrates features from multiple dimensions, uses a hybrid network that considers both local and global aspects, and dynamically adapts to noisy scenarios to provide high-quality data for subsequent processing.

[0119] In some embodiments of the present invention, specific data experiments are conducted on the seismic data anomaly trace identification method based on hybrid neural networks proposed in the present invention.

[0120] Step 1: Excite and acquire single-shot seismic data. This data is based on, for example... Figure 4 The observation system shown is designed for data acquisition and employs a regular grid layout: red dots represent excitation points, and blue dots represent receiver points. Due to limitations imposed by the actual terrain and features of the work area, receiver points could not be deployed in the gray area in the upper left corner due to obstacles. This design objectively reflects the real constraints of field data acquisition. The specific parameters of the observation system are as follows: receiver line spacing is 200 meters, excitation line spacing (shot row spacing) is 200 meters, track spacing is 40 meters, shot point spacing is 40 meters, each receiver line has 450 tracks, each track has 4000 sampling points, and the sampling interval is 1 ms.

[0121] Based on the shot collection records collected by the aforementioned observation system, as follows: Figure 5 As shown, the rapid urbanization and industrialization in the exploration area have resulted in a dense distribution of high-voltage power grids, heavy machinery operation areas, and transportation hubs, generating strong electromagnetic interference and continuous mechanical vibrations. These combined interference sources couple to form a high-energy noise field, causing phase distortion, amplitude anomalies, and waveform distortion in the seismic wave field, resulting in numerous anomalous traces with abrupt amplitude changes and broken phase axes in the records.

[0122] The second step involves standardizing the raw seismic data, linearly scaling each trace to the [0, 1] interval to eliminate energy magnitude differences between traces. Median filtering and Fourier transform are then applied to the standardized data to extract low-frequency trend features, spectral distribution features, and amplitude statistical features. Based on these multi-dimensional attributes, a multi-channel input framework is constructed, concatenating the original standardized data, filtered features, and spectral features along the channel dimension to form a matrix that integrates multiple features, significantly improving the feature difference between anomalous and normal traces.

[0123] The third step is to divide the multi-channel feature matrix obtained in the second step into blocks, dividing it into two-dimensional sub-blocks according to the preset sub-block size (256×256), and synchronously recording the original trace index corresponding to each sub-block to ensure that the subsequent model prediction results can be accurately mapped to the trace position of the original seismic data; based on the preset model training and testing parameters, an execution framework is provided for model learning.

[0124] Step 4: Construct a hybrid neural network model that integrates convolution and self-attention, including a feature extraction module: a hybrid feature extraction unit that alternately connects convolutional layers (3×3) and windowed multi-head self-attention modules (8×8). Convolutional operations capture local spatial features of waveform abrupt changes within seismic data traces, while self-attention mechanisms capture global dependencies. Encoding and decoding structure: the encoder progressively compresses feature dimensions through downsampling operations to extract high-level abstract features, while the decoder restores spatial resolution through upsampling operations to map back to the original data scale. Intermediate layers use skip connections to fuse feature information at different scales, enhancing the representation ability of anomalous trace features.

[0125] Step 5: Model training and optimization. An improved binary classification loss function is adopted, and the imbalance between the number of samples in abnormal and normal channels is solved by dynamically adjusting the class weight parameters. The sensitivity adjustment factor is set to 1.5. An adaptive gradient optimizer with regularization term is used to suppress model overfitting through weight decay. An adaptive learning rate adjustment mechanism is introduced, with a learning rate decay ratio of 0.5. When the validation loss stagnates continuously, the learning rate is automatically reduced proportionally. During training, a gradient constraint strategy is adopted (setting the maximum gradient threshold to 1.0) to prevent gradient explosion, and the model weights with the lowest validation loss are saved periodically.

[0126] Step 6: Model Prediction and Preliminary Identification of Abnormal Channels: Load the trained model, and generate multi-channel sub-blocks for the test data using the preprocessing method in Step 3 (keeping the format consistent with the training data); input the sub-blocks into the model and output the abnormal probability map for each sub-block; use the maximum probability fusion strategy (taking the highest probability of the same channel in different sub-blocks) for overlapping channels across sub-blocks to eliminate the probability differences caused by sub-block segmentation; set the abnormal judgment threshold to 0.5, and judge the channels with probabilities exceeding the threshold as abnormal channels to generate preliminary identification results.

[0127] Step 7: Dynamically adjust the model parameters based on the actual number of channels and noise level of the data. When the signal-to-noise ratio of the data is lower than the set value, increase the attention of the loss function to difficult-to-identify samples and increase the model protection factor to 0.6 to reduce the missed detection of abnormal channels. Iterate and test different decision thresholds, and use the harmonic mean of precision and recall as the evaluation index to select the optimal threshold, which is finally 0.5. Post-process the identification results: smooth the spatial distribution of abnormal channels by median filtering to eliminate misjudgments caused by isolated noise points.

[0128] Step 8: Save the optimized anomaly identification results by shot set number, and output them as a text file containing the track index and anomaly markers; calculate the identification accuracy and recall by comparing with the manual annotation results to verify the stability of the algorithm and ensure that the anomaly identification accuracy is not less than 95% to meet the needs of exploration data processing.

[0129] Figure 6 For single-shot data processed using this invention, and... Figure 5 In contrast, anomalies are accurately detected, and random noise energy is effectively suppressed. This invention captures local and global features through multi-dimensional feature fusion, a hybrid network of convolution and self-attention, dynamically adjusts parameters to adapt to noisy scenarios, solves the problem of imbalanced samples, and optimizes the identification results through post-processing to ensure that the accuracy is not lower than the set threshold. This provides high-quality basic data for subsequent seismic data processing and meets the needs of petroleum geophysical exploration.

[0130] like Figure 2 As shown, in some embodiments of the present invention, a seismic data anomaly identification device 200 based on a hybrid neural network is proposed, comprising: a first acquisition unit 210 for acquiring raw single-shot seismic data; a first processing unit 220 for acquiring standardized single-shot seismic data based on the raw single-shot seismic data; a second processing unit 230 for acquiring first feature data based on the standardized single-shot seismic data, wherein the first feature data includes low-frequency trend feature data and spectral distribution feature data, or the first feature data includes wavelet coefficient feature data and curvelet coefficient feature data; a third processing unit 240 for acquiring a first matrix based on the standardized single-shot seismic data and the first feature data; a fourth processing unit 250 for dividing the first matrix into multiple first data sub-blocks; a fifth processing unit 260 for acquiring an identification model based on the multiple first data sub-blocks and a hybrid neural network model; a second acquisition unit 270 for acquiring seismic data to be identified; and a sixth processing unit 280 for identifying anomalies in the seismic data to be identified based on the seismic data to be identified and the identification model.

[0131] In this embodiment, the seismic data anomaly identification device 200 based on a hybrid neural network proposed in this invention includes an acquisition unit, multiple processing units, and an identification unit. The first acquisition unit 210 acquires raw single-shot seismic data, which the identification device 200 uses as the starting point for processing. Subsequently, the first processing unit 220 acquires standardized single-shot seismic data based on the raw single-shot seismic data. The function of this processing unit is to eliminate energy magnitude differences between traces caused by variations in acquisition conditions, establishing a unified benchmark for subsequent feature extraction and model analysis.

[0132] The second processing unit 230 acquires first feature data based on the standardized data. This feature data specifically refers to mathematical results extracted from the standardized data that characterize the essential attributes of the signal. These include low-frequency trend feature data and spectral distribution feature data, or wavelet coefficient feature data and curvelet coefficient feature data. The core function of the second processing unit 230 is to perform multi-dimensional feature fusion, that is, to enrich the characterization of the seismic trace from multiple attribute levels of the signal, thereby enhancing the feature differences between anomalous and normal traces.

[0133] The third processing unit 240 obtains the first matrix based on the standardized single-shot seismic data and the aforementioned first feature data. The function of the third processing unit 240 is to splice the original standardized data with the extracted multiple feature data in the channel dimension to construct a multi-channel fusion matrix with more comprehensive information and greater discriminative power.

[0134] The fourth processing unit 250 divides the first matrix into multiple first data sub-blocks. This design not only adapts to the input size requirements of the neural network model, but more importantly, by preserving the spatial relationship between the channels within each sub-block, it creates conditions for the model to learn local features and global correlations within the sub-blocks at the same time.

[0135] The fifth processing unit 260 obtains a recognition model based on multiple first data sub-blocks and a hybrid neural network model. Here, the hybrid neural network model refers to a network structure capable of simultaneously processing local features and global correlations. The model training process utilizes labeled abnormal and normal channel sample data to teach the hybrid neural network model how to identify abnormal patterns from these data sub-blocks. Through the design of the recognition model, this invention can simultaneously capture local detailed features such as waveform abrupt changes within channels as well as global correlation anomalies between channels, thereby achieving collaborative analysis and accurate recognition of multi-scale features of abnormal channels.

[0136] Finally, in the application phase, the second acquisition unit 270 acquires the seismic data to be identified, and the sixth processing unit 280 inputs the seismic data to be identified into the pre-trained identification model. The model outputs the identification result of whether each trace in the data to be identified is an anomalous trace, thereby completing the automated identification.

[0137] The seismic data anomaly identification device 200 based on a hybrid neural network proposed in this invention constructs a rich feature representation system by simultaneously utilizing information such as original standardized data, low-frequency trend features, and spectral distribution features. This multi-channel input structure ensures that even in high-noise environments, the distortion features of anomalies can be reflected in multiple feature dimensions, greatly reducing the risk of missed detections caused by a single feature being overwhelmed by noise. Compared with traditional energy thresholding methods, it significantly improves the robustness and reliability of feature extraction in low signal-to-noise ratio environments.

[0138] This invention significantly enhances the distinguishability of anomalies in the feature space through rich input features constructed by feature fusion, laying a solid foundation for subsequent accurate identification. Secondly, by employing a hybrid neural network model to process block data, it effectively solves the core problem of the separation between local and global features in traditional methods, enabling the synergistic use of two key types of information—waveform detail abrupt changes and inter-trace spatial correlations—for analysis. This achieves high-precision identification of complex anomalies, effectively overcoming the problems of low identification accuracy, serious missed detections and misjudgments caused by insufficient feature utilization and a single analysis mode in existing technologies, providing a reliable guarantee for seismic data quality control and subsequent processing.

[0139] In this embodiment, during the process of obtaining standardized single-shot seismic data based on the original single-shot seismic data, the first processing unit 220 is specifically used to: normalize the original single-shot seismic data to obtain standardized single-shot seismic data.

[0140] In this embodiment, when the first feature data includes low-frequency trend feature data and spectral distribution feature data, during the process of obtaining the first feature data based on standardized single-shot seismic data, the second processing unit 230 is specifically used to: perform median filtering calculation on the standardized single-shot seismic data to obtain low-frequency trend feature data; and perform Fourier transform on the standardized single-shot seismic data to obtain spectral distribution feature data.

[0141] Specifically, the window size for median filtering is 3 to 9.

[0142] In this embodiment, when the first feature data includes wavelet coefficient feature data and curvelet coefficient feature data, during the process of obtaining the first feature data based on standardized single-shot seismic data, the second processing unit 230 is specifically used to perform wavelet transform on the standardized single-shot seismic data to obtain wavelet coefficient feature data; and to perform curvelet transform on the standardized single-shot seismic data to obtain curvelet coefficient feature data.

[0143] In this embodiment, during the process of obtaining the first matrix based on standardized single-shot seismic data and first feature data, the third processing unit 240 is specifically used to: obtain first matrix data based on standardized single-shot seismic data; obtain second matrix data and third matrix data based on the first feature data; and obtain the first matrix based on the first matrix data, second matrix data, and third matrix data.

[0144] In this embodiment, after dividing the first matrix into multiple first data sub-blocks, the fourth processing unit 250 is further configured to: store the correspondence between each first data sub-block and the seismic traces in the original single-shot seismic data.

[0145] In the process of obtaining anomalous traces in the seismic data to be identified based on the seismic data to be identified and the identification model, the sixth processing unit 280 is specifically used to obtain standardized seismic data to be identified based on the seismic data to be identified; obtain second feature data based on the standardized seismic data to be identified; obtain a second matrix based on the standardized seismic data to be identified and the second feature data; divide the second matrix into multiple second data sub-blocks; obtain anomaly probability maps for each second data sub-block based on the multiple second data sub-blocks and the identification model; obtain the comprehensive anomaly probability of each seismic trace based on the anomaly probability maps; and obtain anomalous traces in the seismic data to be identified based on the comprehensive anomaly probability and the judgment threshold.

[0146] In this embodiment, the sixth processing unit 280 is further configured to output an abnormal channel list containing channel indices and abnormal markers based on the abnormal channels. Specifically, outputting the abnormal channel list includes: performing spatial distribution smoothing processing on the abnormal channels and outputting the abnormal channel list containing channel indices and abnormal markers; wherein the spatial distribution smoothing processing is implemented using median filtering or Gaussian filtering, and when using median filtering, the window size of the median filter is 3 to 9.

[0147] In this embodiment, during the process of obtaining the comprehensive anomaly probability of each seismic trace based on the anomaly probability map, the sixth processing unit 280 is specifically used to: obtain the correspondence between each second data sub-block and the seismic trace in the seismic data to be identified; map the probability value of each second data sub-block in the anomaly probability map back to the corresponding seismic trace based on the correspondence; and obtain the comprehensive anomaly probability of each seismic trace based on all the probability values ​​mapped to the same seismic trace.

[0148] In this embodiment, the fifth processing unit 260 is further configured to construct a hybrid neural network model; wherein the hybrid neural network model includes a feature extraction module consisting of alternating connections of convolutional layers and windowed multi-head self-attention modules. The convolutional layers use 3×3 or 5×5 convolutional kernels, and the window size of the windowed multi-head self-attention modules is 4×4 to 16×16.

[0149] Specifically, the hybrid neural network model is a structure with a connected encoder and decoder, wherein the encoder performs downsampling through max pooling or stride convolution, and the decoder performs upsampling through transposed convolution or interpolation.

[0150] In an embodiment of the present invention, a storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the method for identifying seismic data anomalies based on a hybrid neural network in any of the above embodiments.

[0151] In this embodiment, the storage medium proposed by the present invention implements the steps of the method for identifying seismic data anomalies based on hybrid neural networks in any of the above embodiments when the computer program is executed by the processor. Therefore, it has all the beneficial effects of the method for identifying seismic data anomalies based on hybrid neural networks in any of the above embodiments, and will not be repeated here.

[0152] like Figure 3 As shown, in some embodiments of the present invention, an electronic device 300 is provided, including a memory 310, a processor 320, and a computer program stored in the memory 310 and executable on the processor 320. When the processor 320 executes the computer program, it implements the steps of the method for identifying seismic data anomalies based on hybrid neural networks in any of the above embodiments.

[0153] In this embodiment, the electronic device 300 and processor 320 of the present invention implement the steps of the method for identifying seismic data anomalies based on hybrid neural networks in any of the above embodiments when executing computer programs. Therefore, it has all the beneficial effects of the method for identifying seismic data anomalies based on hybrid neural networks in any of the above embodiments, which will not be repeated here.

[0154] The above are preferred embodiments of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying seismic data anomalies based on hybrid neural networks, characterized in that, include: Acquire raw single-shot seismic data; Based on the original single-shot seismic data, standardized single-shot seismic data are obtained; Based on the standardized single-shot seismic data, first feature data is obtained, wherein the first feature data includes low-frequency trend feature data and spectral distribution feature data, or the first feature data includes wavelet coefficient feature data and curve wave coefficient feature data. Based on the standardized single-shot seismic data and the first feature data, a first matrix is ​​obtained; The first matrix is ​​divided into multiple first data sub-blocks; Based on multiple first data sub-blocks and a hybrid neural network model, an identification model is obtained; Acquire earthquake data to be identified; Based on the earthquake data to be identified and the identification model, identify the anomalous traces in the earthquake data to be identified.

2. The method for identifying seismic data anomalies based on a hybrid neural network according to claim 1, characterized in that, The step of obtaining standardized single-shot seismic data based on the original single-shot seismic data specifically includes: The original single-shot seismic data is normalized to obtain the standardized single-shot seismic data.

3. The method for identifying seismic data anomalies based on a hybrid neural network according to claim 1, characterized in that, When the first feature data includes low-frequency trend feature data and spectral distribution feature data, obtaining the first feature data based on the standardized single-shot seismic data specifically includes: The low-frequency trend feature data are obtained by performing median filtering on the standardized single-shot seismic data. The standardized single-shot seismic data is subjected to Fourier transform to obtain the spectral distribution characteristic data.

4. The method for identifying seismic data anomalies based on a hybrid neural network according to claim 3, characterized in that, When performing median filtering calculations on the standardized single-shot seismic data, the window size for median filtering is between 3 and 9.

5. The method for identifying seismic data anomalies based on a hybrid neural network according to claim 1, characterized in that, When the first feature data includes wavelet coefficient feature data and curvelet coefficient feature data, the step of obtaining the first feature data based on the standardized single-shot seismic data specifically includes: Wavelet transform is performed on the standardized single-shot seismic data to obtain the wavelet coefficient feature data; The standardized single-shot seismic data are subjected to curvelet transform to obtain the curvelet coefficient characteristic data.

6. The method for identifying seismic data anomalies based on a hybrid neural network according to claim 1, characterized in that, The step of obtaining the first matrix based on the standardized single-shot seismic data and the first feature data specifically includes: Based on the standardized single-shot seismic data, obtain the first matrix data; Based on the first feature data, obtain the second matrix data and the third matrix data; The first matrix is ​​obtained based on the first matrix data, the second matrix data, and the third matrix data.

7. The method for identifying seismic data anomalies based on a hybrid neural network according to claim 1, characterized in that, After dividing the first matrix into multiple first data sub-blocks, the identification method further includes: Store the correspondence between each of the first data sub-blocks and the seismic traces in the original single-shot seismic data.

8. The method for identifying seismic data anomalies based on a hybrid neural network according to claim 1, characterized in that, The step of identifying anomalous traces in the seismic data to be identified based on the seismic data to be identified and the identification model specifically includes: Based on the earthquake data to be identified, obtain standardized earthquake data to be identified; Based on the standardized earthquake data to be identified, the second feature data is obtained; Based on the standardized seismic data to be identified and the second feature data, a second matrix is ​​obtained; The second matrix is ​​divided into multiple second data sub-blocks; Based on multiple second data sub-blocks and the recognition model, obtain an anomaly probability map for each second data sub-block; Based on the aforementioned anomaly probability map, the comprehensive anomaly probability of each seismic trace is obtained; Based on the comprehensive anomaly probability and the judgment threshold, the anomalous traces in the seismic data to be identified are determined.

9. The method for identifying seismic data anomalies based on a hybrid neural network according to claim 8, characterized in that, After identifying anomalous traces in the seismic data to be identified based on the comprehensive anomaly probability and the judgment threshold, the identification method further includes: Based on the abnormal path, output a list of abnormal paths containing path indexes and abnormal tags.

10. The method for identifying seismic data anomalies based on a hybrid neural network according to claim 9, characterized in that, The step of outputting an abnormal channel list containing channel indexes and abnormal tags based on the abnormal channels specifically includes: The abnormal channels are spatially smoothed, and a list of abnormal channels containing channel indexes and abnormal tags is output. The spatial distribution smoothing process is implemented using median filtering or Gaussian filtering. When using median filtering, the window size of the median filter is 3 to 9.

11. The method for identifying seismic data anomalies based on a hybrid neural network according to claim 8, characterized in that, After dividing the second matrix into multiple second data sub-blocks, the identification method further includes: Store the correspondence between each of the second data sub-blocks and the seismic traces in the seismic data to be identified; The step of obtaining the comprehensive anomaly probability of each seismic trace based on the anomaly probability map specifically includes: Obtain the correspondence between each of the second data sub-blocks and the seismic traces in the seismic data to be identified; Based on the aforementioned correspondence, the probability value of each second data sub-block in the anomaly probability map is mapped back to the corresponding seismic trace; The comprehensive anomaly probability of each seismic trace is obtained by mapping all probability values ​​to the same seismic trace.

12. The method for identifying seismic data anomalies based on hybrid neural networks according to any one of claims 1 to 11, characterized in that, Before obtaining the recognition model based on multiple first data sub-blocks and the hybrid neural network model, the recognition method further includes: Construct the hybrid neural network model; The hybrid neural network model includes a feature extraction module consisting of alternating connections of convolutional layers and window multi-head self-attention modules.

13. The method for identifying seismic data anomalies based on a hybrid neural network according to claim 12, characterized in that, The convolutional layer uses a 3×3 or 5×5 convolutional kernel, and the window size of the multi-head self-attention module is 4×4 to 16×16.

14. The method for identifying seismic data anomalies based on a hybrid neural network according to claim 12, characterized in that, The hybrid neural network model has a structure with a connected encoder and decoder, wherein the encoder performs downsampling through max pooling or stride convolution, and the decoder performs upsampling through transposed convolution or interpolation.

15. A device for identifying seismic data anomalies based on a hybrid neural network, characterized in that, The identification device includes: The first acquisition unit is used to acquire raw single-shot seismic data; The first processing unit is used to obtain standardized single-shot seismic data based on the original single-shot seismic data. The second processing unit is used to obtain first feature data based on the standardized single-shot seismic data, wherein the first feature data includes low-frequency trend feature data and spectral distribution feature data, or the first feature data includes wavelet coefficient feature data and curve wave coefficient feature data. The third processing unit is used to obtain the first matrix based on the standardized single-shot seismic data and the first feature data; The fourth processing unit is used to divide the first matrix into multiple first data sub-blocks; The fifth processing unit is used to obtain a recognition model based on multiple first data sub-blocks and a hybrid neural network model; The second acquisition unit is used to acquire earthquake data to be identified; The sixth processing unit is used to identify anomalous traces in the seismic data to be identified based on the seismic data to be identified and the identification model.

16. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for identifying seismic data anomalies based on a hybrid neural network as described in any one of claims 1 to 14.

17. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for identifying seismic data anomalies based on a hybrid neural network as described in any one of claims 1 to 14.

Images