Free gas structure seismic data identification method and system based on large model

Abstract

The invention relates to the technical field of seismic data detection, in particular to a free gas structure seismic data identification method and system based on a large model. The method comprises the following steps: acquiring original seismic data; performing short-time Fourier transform on the original seismic data to obtain time-frequency domain seismic data; performing Hilbert-Huang transform on the time-frequency domain seismic data to obtain enhanced seismic feature data; performing feature extraction on the enhanced seismic feature data to obtain a multi-scale seismic feature map; performing semantic segmentation on the enhanced seismic feature data to obtain a free gas structure mask; performing uncertainty evaluation on the free gas structure mask to obtain a confidence coefficient thermodynamic diagram; and carrying out probability prediction on the confidence coefficient thermodynamic diagram to obtain a low-confidence coefficient region marking diagram. According to the method, the problem that effective signals are difficult to accurately extract due to the complexity of seismic data in a traditional method is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of seismic data detection, and particularly relates to a free gas structure seismic data identification method and system based on a large model. BACKGROUND

[0002] In the field of oil and gas exploration, seismic data interpretation is a key link in finding oil and gas reservoirs. Free gas structure is one of the important characteristics of oil and gas reservoirs, and its accurate identification is of great significance for the evaluation and development of oil and gas resources. However, in the actual process of seismic data processing and interpretation, the identification of free gas structure faces many challenges. Traditional seismic data processing methods mainly rely on human experience, and subjective judgment is made on the shape, amplitude, frequency and other characteristics of seismic waveforms to identify free gas structure. This method not only consumes time and effort, but also due to the complexity and diversity of seismic data, the experience difference of different interpreters will lead to inconsistency of interpretation results, thereby affecting the accuracy and efficiency of oil and gas exploration.

[0003] In actual seismic data interpretation work, seismic data is often disturbed by noise superposition and other factors, making the effective signal in the seismic data unclear, increasing the difficulty of free gas structure identification. In this case, the traditional seismic data processing method is difficult to accurately extract the effective signal related to the free gas structure, thereby affecting the accurate identification of the free gas structure. SUMMARY

[0004] Therefore, it is necessary to provide a free gas structure seismic data identification method and system based on a large model to solve at least one of the above technical problems.

[0005] To achieve the above purpose, a free gas structure seismic data identification method based on a large model comprises the following steps:

[0006] Step S1: obtaining original seismic data; performing short-time Fourier transform on the original seismic data to obtain time-frequency domain seismic data; performing Hilbert-Huang transform on the time-frequency domain seismic data to obtain enhanced seismic feature data;

[0007] Step S2: performing feature extraction on the enhanced seismic feature data to obtain a multi-scale seismic feature map; performing semantic segmentation on the enhanced seismic feature data to obtain a free gas structure mask;

[0008] Step S3: performing uncertainty evaluation on the free gas structure mask to obtain a confidence heat map; performing probability prediction on the confidence heat map to obtain a low confidence region label map; performing data enhancement on the low confidence region label map to obtain corrected seismic feature data; inputting the corrected seismic feature data into a pre-trained visual Transformer large model to extract features and obtain an optimized free gas structure mask;

[0009] Step S4: based on the optimized free gas structure mask, three-dimensional space reconstruction is performed to obtain an initial free gas structure model; line interpolation is performed on the optimized free gas structure mask to obtain three-dimensional free gas structure point cloud; spatial fitting is performed on the three-dimensional free gas structure point cloud to obtain continuous free gas structure body data;

[0010] Step S5: interactive interpretation is performed on the continuous free gas structure body data to obtain a free gas structure three-dimensional model; based on the free gas structure three-dimensional model, visual rendering is performed to obtain a free gas structure three-dimensional model with a probability label.

[0011] Preferably, the present application also provides a large model-based free gas structure seismic data identification system for executing the large model-based free gas structure seismic data identification method as described above, and the large model-based free gas structure seismic data identification system comprises:

[0012] A data preprocessing module is configured to acquire original seismic data; perform short-time Fourier transform on the original seismic data to obtain time-frequency domain seismic data; and perform Hilbert-Huang transform on the time-frequency domain seismic data to obtain enhanced seismic feature data.

[0013] A feature extraction module is configured to perform feature extraction on the enhanced seismic feature data to obtain multi-scale seismic feature maps; and perform semantic segmentation on the enhanced seismic feature data to obtain a free gas structure mask.

[0014] An optimization iteration module is configured to perform uncertainty evaluation on the free gas structure mask to obtain a confidence heat map; perform probability prediction on the confidence heat map to obtain a low confidence area label map; perform data enhancement on the low confidence area label map to obtain corrected seismic feature data; and input the corrected seismic feature data into a pre-trained visual Transformer large model to perform feature extraction to obtain an optimized free gas structure mask.

[0015] A three-dimensional modeling module is configured to perform three-dimensional space reconstruction based on the optimized free gas structure mask to obtain an initial free gas structure model; perform line interpolation on the optimized free gas structure mask to obtain three-dimensional free gas structure point cloud; and perform spatial fitting on the three-dimensional free gas structure point cloud to obtain continuous free gas structure body data.

[0016] A visualization module is configured to perform interactive interpretation on the continuous free gas structure body data to obtain a free gas structure three-dimensional model; and perform visual rendering based on the free gas structure three-dimensional model to obtain a free gas structure three-dimensional model with a probability label.

[0017] The present application has the following beneficial effects:

[0018] On the one hand, the quality of the seismic data can be significantly improved by preprocessing the original seismic data with specific signal processing techniques, making the effective signal originally unclear prominent, laying a solid foundation for subsequent free gas structure identification work, and overcoming the difficulty of traditional methods in accurately extracting effective signals due to the complexity of seismic data.

[0019] On the other hand, advanced pre-trained large models can be used to deeply mine and analyze the features in the seismic data, identify free gas structure masks in a more efficient and accurate manner, and effectively evaluate and correct the uncertainty of the identification results, thereby significantly improving the accuracy of free gas structure identification and overcoming the drawbacks of traditional manual experience judgment, such as large subjective differences and unstable results.

[0020] On the other hand, by performing a series of complex spatial processing and interactive interpretation operations on the optimized data, an intuitive and probability-labeled free gas structure three-dimensional model is finally constructed, realizing the visualization and quantitative analysis of the free gas structure. BRIEF DESCRIPTION OF DRAWINGS

[0021] Other features, objects and advantages of the present application will become more apparent from the following detailed description, made with reference to the accompanying drawings:

[0022] Fig. 1 FIG. 1 shows a step flow diagram of a large model-based free gas structure seismic data identification method according to an embodiment.

[0023] Fig. 2 FIG. 3 shows a detailed step flow diagram of step S23 according to an embodiment.

[0024] Fig. 3 FIG. 4 shows a seismic data main frequency distribution characteristic analysis diagram according to an embodiment. DETAILED DESCRIPTION

[0025] The technical method of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0026] Further, the accompanying drawings are included to provide a further understanding of the present application, and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments of the present application and, together with the description, serve to explain the principles of the present application. In the drawings:

[0027] It is to be understood that, although terms such as "first", "second", and so on can be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, without departing from the scope of the exemplary embodiments, a first element can be referred to as a second element, and similarly a second element can be referred to as a first element. The term "and / or" as used herein includes any and all combinations of one or more of the associated associated items.

[0028] To achieve the above object, there is provided Figs. 1 to 3 The present application provides a free gas structure seismic data identification method based on a large model, comprising the following steps:

[0029] Step S1: obtaining original seismic data; performing short-time Fourier transform on the original seismic data to obtain time-frequency domain seismic data; performing Hilbert-Huang transform on the time-frequency domain seismic data to obtain enhanced seismic feature data;

[0030] Step S2: performing feature extraction on the enhanced seismic feature data to obtain a multi-scale seismic feature map; performing semantic segmentation on the enhanced seismic feature data to obtain a free gas structure mask;

[0031] Step S3: performing uncertainty evaluation on the free gas structure mask to obtain a confidence heat map; performing probability prediction on the confidence heat map to obtain a low confidence area label map; performing data enhancement on the low confidence area label map to obtain modified seismic feature data; inputting the modified seismic feature data into a pre-trained visual Transformer large model to perform feature extraction, and obtaining an optimized free gas structure mask;

[0032] Step S4: based on the optimized free gas structure mask, performing three-dimensional space reconstruction to obtain an initial free gas structure model; performing line interpolation on the optimized free gas structure mask to obtain a three-dimensional free gas structure point cloud; performing spatial fitting on the three-dimensional free gas structure point cloud to obtain continuous free gas structure body data;

[0033] Step S5: interactive interpretation of the continuous free gas structure data to obtain a free gas structure three-dimensional model; visual rendering based on the free gas structure three-dimensional model to obtain a free gas structure three-dimensional model with a probability label.

[0034] Preferably, step S1 comprises the following steps:

[0035] Step S11: obtaining original seismic data by a seismic acquisition device;

[0036] In the embodiment of the present application, a marine seismic acquisition system is used, which is equipped with high-sensitivity geophones and precise positioning devices. In a specific exploration sea area, the original seismic data covering the target area are continuously collected by the seismic acquisition device towed at the stern of the ship at a set sampling rate (e.g., 250 Hz) and sampling interval (e.g., 4 ms). These data contain the propagation information of seismic waves in different strata.

[0037] Step S12: wavelet threshold denoising of the original seismic data to obtain denoised seismic data;

[0038] In the embodiment of the present application, a wavelet threshold denoising method is used for preprocessing. Specifically, Daubechies wavelet (db4) is selected as the wavelet basis. The original seismic data are decomposed into detail coefficients and approximation coefficients of different frequency components by multi-scale wavelet decomposition. According to the general threshold theorem of Donoho, a suitable threshold (e.g., based on the standard deviation of the data and the square root of the number of sampling points) is calculated to perform soft threshold processing on the detail coefficients to remove the noise components therein. Finally, the denoised seismic data after noise removal are obtained by wavelet reconstruction.

[0039] Step S13: time window division based on the denoised seismic data to obtain segmented seismic data;

[0040] Step S14: short-time Fourier transform of the segmented seismic data to obtain time-frequency domain seismic data;

[0041] In the embodiment of the present application, for the effective segmented seismic data selected and labeled, the short-time Fourier transform (STFT) technique is used to convert it to the time-frequency domain. The Hanning window is selected as the window function, and the window length is set to 128 sampling points to adapt to the signal characteristics of the seismic data. The signal is decomposed into time-frequency spectrograms of different frequency components by performing Fourier transform on the seismic data in each time window.

[0042] Step S15: Hilbert-Huang transform of the time-frequency domain seismic data to obtain instantaneous frequency feature data;

[0043] In the embodiment of the present application, after obtaining the time-frequency domain seismic data, the empirical mode decomposition (EMD) part in Hilbert-Huang transform (HHT) is further used to process the time-frequency domain seismic data. In actual operation, the EMD toolbox in MATLAB software is used to decompose the time-frequency domain seismic data point by point. Through multiple iterations and screening, the data is decomposed into multiple IMF components, and each component represents the characteristics of the seismic signal at different frequencies and time scales. Further, Hilbert transform is performed on each IMF component to obtain its instantaneous frequency characteristic data. These instantaneous frequency characteristic data can more accurately describe the propagation characteristics of seismic waves in different strata.

[0044] Step S16: normalizing the instantaneous frequency characteristic data to obtain enhanced seismic characteristic data.

[0045] In the embodiment of the present application, the maximum and minimum normalization method is used to adjust the values of all data points to the range of 0 to 1 by calculating the maximum and minimum values in the instantaneous frequency characteristic data. In actual operation, a corresponding normalization script is written in the seismic data processing software to process each instantaneous frequency characteristic data point point by point. After normalization, the obtained enhanced seismic characteristic data is more stable in numerical value, and the difference between different characteristics is more obvious.

[0046] Preferably, step S13 comprises the following steps:

[0047] Step S130: identifying the seismic wave velocity based on the denoised seismic data to obtain the seismic wave propagation velocity;

[0048] In the embodiment of the present application, the seismic wave velocity is identified based on the reflection travel time method. In actual operation, the semblance analysis tool in the seismic data processing software SeisSpace is used. A seismic trace with a clear reflection interface is selected as a reference, and two reflection wave arrival time points are selected on the seismic trace. The time difference between the two time points is calculated. According to the known stratum depth information, the actual distance between the two reflection interfaces is calculated. Through the ratio of distance to time, the average propagation velocity of the seismic wave in this path is calculated, for example, the calculation result is 2500 m / s.

[0049] Step S131: calculating the main frequency based on the seismic wave propagation velocity to obtain seismic wave main frequency characteristic data; setting the time window parameter according to the seismic wave main frequency characteristic data to obtain time window length configuration data;

[0050] In the embodiment of the present application, after determining that the propagation speed of the seismic wave is 2500 m / s, further main frequency calculation is performed. The fft (fast Fourier transform) function in MATLAB software is used to perform spectrum analysis on the denoised seismic data. In specific operation, a length of 1024 sampling points of seismic data is selected, and fast Fourier transform is performed on it to obtain its spectrum graph. By analyzing the peak position of the spectrum graph, it is determined that the main frequency of the seismic data is 35 Hz. According to this main frequency value, the time window parameter is set. In order to ensure that at least one complete main frequency period is contained in each time window, the time window length is calculated to be 1 / 35 s, that is, about 28.57 ms.

[0051] Step S132: performing overlap rate calculation on the denoised seismic data based on the time window length configuration data to obtain time window overlap parameters;

[0052] In the embodiment of the present application, after determining that the time window length is 28.57 ms, the overlap rate of the time window needs to be calculated. A 50% overlap rate is selected to ensure the continuity of the data and the integrity of the features. In specific operation, the numpy library in the Python programming language is used to process the seismic data. By writing a simple script, the start and end positions of each time window are calculated, and it is ensured that there is 50% data overlap between adjacent time windows. For example, the first time window starts from 0 ms and ends at 28.57 ms; the second time window starts from 14.285 ms and ends at 42.855 ms. In this way, each time window contains part of the data of the previous time window, thereby ensuring the continuity of the seismic signal.

[0053] Step S133: performing sliding time window segmentation on the denoised seismic data using the time window overlap parameters to obtain initial segmented seismic data;

[0054] In the embodiment of the present application, after determining that the time window length is 28.57 ms and the overlap rate is 50%, the window command in the SeismicUnix (SU) software is used to perform sliding time window segmentation on the denoised seismic data. In specific operation, the entire seismic data sequence is segmented according to the set time window length and overlap rate. For example, for a length of 10 seconds of seismic data, a time window data with a length of 28.57 ms is intercepted from 0 ms, and the time window is slid by 14.285 ms, and the next time window data is continuously intercepted, until the entire seismic data sequence is segmented into multiple time window data segments. Through the above steps, initial segmented seismic data is obtained, and each segmented data contains information of the seismic signal in a specific time interval.

[0055] Step S134: performing energy equalization processing on the initial segmented seismic data to obtain equalized segmented seismic data;

[0056] In the embodiment of the present application, after obtaining the initial segmented seismic data, the Gain function in the Open dTect software is used to perform energy equalization processing on each segmented data. In the specific operation, the automatic gain control (AGC) method is selected, which can automatically adjust the energy of the seismic signal in each time window to make the energy distribution more uniform. By setting the window length of AGC to 28.57 milliseconds (consistent with the time window length), the software will automatically calculate the energy of the signal in each time window and adjust it to a preset average energy level. For example, if the signal energy in a time window is low, the AGC will increase its amplitude; on the contrary, if the signal energy is high, the AGC will reduce its amplitude. After the energy equalization processing, the obtained equalized segmented seismic data is more stable in energy.

[0057] Step S135: performing quality screening on the equalized segmented seismic data based on a preset signal-to-noise ratio threshold to obtain effective segmented seismic data;

[0058] In the embodiment of the present application, a signal-to-noise ratio threshold of 3 is preset, that is, the energy of the signal must be at least 3 times the energy of the noise to be considered as an effective data segment. In the specific operation, the Signal-to-Noise Ratio (SNR) analysis tool in the Pro MAX software is used. By calculating the signal-to-noise ratio of each segmented seismic data, the segmented data with a signal-to-noise ratio lower than 3 is marked as invalid data and is excluded. For example, among 100 segmented data, the signal-to-noise ratio of 10 segmented data is lower than 3, and these data segments are excluded. Through the above steps, the effective segmented seismic data is obtained.

[0059] Step S136: performing time marking on the effective segmented seismic data to obtain segmented seismic data.

[0060] In the embodiment of the present application, after screening out the effective segmented seismic data, the TimeMark function in the SeisWorks software is used to mark the time of these data. In the specific operation, according to the start time and end time of each segmented data in the original seismic data, a time label is added. For example, the start time of the first effective segmented seismic data is 0 milliseconds, and the end time is 28.57 milliseconds, the start time of the second effective segmented seismic data is 14.285 milliseconds, and the end time is 42.855 milliseconds, and so on. Through the above steps, each segmented seismic data is given clear time information.

[0061] Preferably, step S2 comprises the following steps:

[0062] Step S21: performing channel dimension expansion on the enhanced seismic feature data to obtain three-dimensional seismic feature data;

[0063] In the embodiment of the present application, when processing enhanced seismic feature data, Python programming language and NumPy library are used to expand the channel dimension of the data. In specific operation, two-dimensional seismic feature data is loaded into memory, which contains the features of seismic signals at different times and frequencies. Through the array operation function of NumPy, the two-dimensional data is expanded to three-dimensional data. For example, the original two-dimensional data shape is (1000, 50), indicating that there are 1000 time steps and 50 frequency features. By adding a channel dimension, it is expanded to a three-dimensional array with a shape of (1, 1000, 50), where the first dimension represents the number of channels.

[0064] Step S22: using a pre-trained visual Transformer model to extract features from the three-dimensional seismic feature data to obtain a primary seismic feature map;

[0065] In the embodiment of the present application, after obtaining the three-dimensional seismic feature data, a pre-trained visual Transformer model is used for feature extraction. A PyTorch deep learning framework is selected, and a pre-trained Vision Transformer (ViT) model is loaded. In specific operation, the three-dimensional seismic feature data is input into the ViT model. The ViT model can effectively capture long-distance dependencies and local features in the data through its multi-head self-attention mechanism. The output of the model is a higher-level feature representation, i.e. a primary seismic feature map. For example, after the input three-dimensional seismic feature data is processed by the ViT model, the output primary seismic feature map has a shape of (1, 64, 25), indicating that there are 64 feature channels and 25 feature points. These primary seismic feature maps contain high-level features of seismic signals.

[0066] Step S23: feature enhancement is performed on the primary seismic feature map to obtain a multi-scale seismic feature map;

[0067] Step S24: the multi-scale seismic feature map is up-sampled and reconstructed to obtain a semantic segmentation probability map;

[0068] In the embodiment of the present application, the torch.nn.functional.interpolate function in PyTorch is used to up-sample and reconstruct the multi-scale seismic feature map. In specific operation, Bilinear Interpolation is selected as the up-sampling method. For example, the original multi-scale seismic feature map has a shape of (1, 64, 25), indicating that there are 64 feature channels and 25 feature points. Through Bilinear Interpolation, the resolution of the feature map is improved to the same resolution as the input data, for example, to (1, 64, 1000). In this way, the obtained semantic segmentation probability map can more accurately reflect the free gas structure information in the seismic signal.

[0069] Step S25: binarizing the semantic segmentation probability map based on a preset threshold to obtain a free gas structure mask.

[0070] In the embodiment of the present application, the cv2.threshold function in the Open CV library is used to binarize the semantic segmentation probability map. In the specific operation, a threshold is preset, for example, 0.5. By comparing each pixel value in the semantic segmentation probability map with the threshold, the pixel value greater than the threshold is set to 1 (indicating the free gas structure region), and the pixel value less than or equal to the threshold is set to 0 (indicating the non-free gas structure region). For example, a pixel value in the semantic segmentation probability map is 0.6, which is greater than the threshold 0.5, so the pixel is marked as 1 in the free gas structure mask. Through the above steps, a clear free gas structure mask is obtained.

[0071] Preferably, step S23 comprises the following steps:

[0072] Step S231: pyramid pooling processing is performed on the primary seismic feature map to obtain multi-scale context feature data;

[0073] In the embodiment of the present application, when processing the primary seismic feature map, the torch.nn.functional.adaptive_avg_pool2d function in the PyTorch framework is used to realize the pyramid pooling processing. In the specific operation, four different pooling scales are selected: 1x1, 2x2, 4x4 and 8x8. For example, the shape of the primary seismic feature map is (1, 64, 25, 25), indicating that there are 64 feature channels and 25x25 feature points. By applying adaptive average pooling operations of the four different scales, four feature maps of different scales are obtained. These feature maps capture different ranges of context information, thereby obtaining multi-scale context feature data.

[0074] Step S232: using a spatial attention mechanism to calculate feature weights of the multi-scale context feature data to obtain a spatial attention weight map;

[0075] In the embodiment of the present application, after obtaining the multi-scale context feature data, the spatial attention mechanism is used to calculate the feature weight of the feature data. The torch.nn.Conv2d and torch.nn.Sigmoid functions in the PyTorch framework are used to realize the spatial attention mechanism. In the specific operation, the multi-scale context feature data is passed through a convolutional layer, and the output channel number of the convolutional layer is 1, so as to generate a spatial attention map. The value of the spatial attention map is normalized to 0 to 1 through the Sigmoid activation function, and a spatial attention weight map is obtained. For example, after convolution and activation function processing, the shape of the obtained spatial attention weight map is (1, 1, 25, 25), and the value of each position represents the spatial importance weight of the position.

[0076] Step S233: recalibrating the primary seismic feature map according to the spatial attention weight map to obtain a spatial enhanced feature map;

[0077] In the embodiment of the present application, the torch.mul function in the PyTorch framework is used to recalibrate the primary seismic feature map. In the specific operation, the primary seismic feature map and the spatial attention weight map are multiplied element by element. For example, the shape of the primary seismic feature map is (1, 64, 25, 25), and the shape of the spatial attention weight map is (1, 1, 25, 25). Through element-by-element multiplication, the shape of the obtained spatial enhanced feature map is still (1, 64, 25, 25). This operation makes the feature points in each feature channel be reweighted according to their spatial importance, thereby obtaining the spatial enhanced feature map.

[0078] Step S234: performing channel correlation analysis on the spatial enhanced feature map through the channel attention mechanism to obtain a channel attention weight map;

[0079] In the embodiment of the present application, the torch.nn.AdaptiveAvgPool2d and torch.nn.Linear functions in the PyTorch framework are used to realize the channel attention mechanism. In the specific operation, the global average pooling operation is performed on the spatial enhanced feature map, and the feature map of each feature channel is compressed into a scalar value to obtain a channel descriptor. The channel descriptors are processed through a fully connected layer to obtain a channel attention weight. For example, the shape of the spatial enhanced feature map is (1, 64, 25, 25), and the shape of the channel descriptor obtained after the global average pooling is (1, 64). After processing through the fully connected layer, the shape of the obtained channel attention weight is also (1, 64). These channel attention weights represent the importance of each channel.

[0080] Step S235: performing channel feature enhancement on the spatial enhanced feature map based on the channel attention weight map to obtain a multi-scale seismic feature map.

[0081] In the embodiment of the present application, the torch.mul function in the PyTorch framework is used to perform channel feature enhancement on the spatial enhanced feature map. In the specific operation, each channel of the spatial enhanced feature map is multiplied by the corresponding channel attention weight element by element. For example, the shape of the spatial enhanced feature map is (1, 64, 25, 25), and the shape of the channel attention weight is (1, 64). Through element-by-element multiplication, the shape of the obtained multi-scale seismic feature map is still (1, 64, 25, 25). This operation causes the features of each channel to be reweighted according to their importance, thereby obtaining a multi-scale seismic feature map.

[0082] Preferably, step S3 comprises the following steps:

[0083] Step S31: performing multiple predictions on the free gas structure mask based on the Monte Carlo method to obtain prediction variance data;

[0084] Step S32: performing Gaussian smoothing on the prediction variance data to obtain a confidence heat map;

[0085] In the embodiment of the present application, the cv2.GaussianBlur function in the Open CV library is used to perform Gaussian smoothing on the prediction variance data. In the specific operation, the size of the Gaussian kernel is selected to be (5, 5), and the standard deviation is 1. For example, the shape of the prediction variance data is (1, 1, 256, 256), and after applying Gaussian blur processing, the smoothed prediction variance data, i.e., the confidence heat map, is obtained. This heat map can intuitively show the confidence levels of different regions.

[0086] Step S33: performing binary segmentation on the confidence heat map according to a preset confidence threshold to obtain a low-confidence region marker map;

[0087] In the embodiment of the present application, the cv2.threshold function in the Open CV library is used to perform binary segmentation on the confidence heat map. In the specific operation, a confidence threshold is preset, for example, 0.3. By comparing each pixel value in the confidence heat map with the threshold, the pixel values greater than the threshold are set to 0 (indicating a high-confidence region), and the pixel values less than or equal to the threshold are set to 1 (indicating a low-confidence region). For example, the shape of the confidence heat map is (1, 1, 256, 256), and after binary segmentation, the low-confidence region marker map is obtained, which also has the shape of (1, 1, 256, 256). This low-confidence region marker map clearly identifies the regions that need to be further processed.

[0088] Step S34: feature completion is performed on the low-confidence region label map by using a preset generative adversarial network to obtain enhanced feature data;

[0089] In the embodiment of the present application, after obtaining the low-confidence region label map, a preset generative adversarial network (GAN) is used to perform feature completion. The PyTorch framework is used to implement this process. In specific operation, a pre-trained GAN model is selected, which can generate missing feature data according to the low-confidence region label map. For example, the shape of the low-confidence region label map is (1, 1, 256, 256), and after inputting it into the GAN model, the model outputs enhanced feature data, which also has a shape of (1, 1, 256, 256). This enhanced feature data can fill in the missing information of the low-confidence region, making the feature map more complete and accurate.

[0090] Step S35: feature fusion is performed on the enhanced feature data and the original seismic feature data to obtain modified seismic feature data;

[0091] In the embodiment of the present application, the torch.cat function in the PyTorch framework is used to perform feature fusion on the enhanced feature data and the original seismic feature data. In specific operation, the enhanced feature data and the original seismic feature data are spliced along the channel dimension. For example, the shape of the enhanced feature data is (1, 1, 256, 256), and the shape of the original seismic feature data is (1, 64, 256, 256), and after splicing, the shape of the modified seismic feature data is (1, 65, 256, 256).

[0092] Step S36: input the modified seismic feature data into a pre-trained visual Transformer large model for iterative optimization to obtain an optimized free gas structure mask.

[0093] In the embodiment of the present application, the PyTorch framework is used to input the modified seismic feature data into a pre-trained visual Transformer large model for iterative optimization. In specific operation, a pre-trained Vision Transformer (ViT) model is selected and set to a trainable mode. The modified seismic feature data is input as input, and after being processed by the multi-head self-attention mechanism and the feedforward network of the ViT model, the model outputs an optimized free gas structure mask. For example, the shape of the modified seismic feature data is (1, 65, 256, 256), and after being processed by the ViT model, the shape of the optimized free gas structure mask is (1, 1, 256, 256).

[0094] Preferably, step S31 comprises the following steps:

[0095] Step S311: Set the Monte Carlo sampling number parameter to obtain sampling number configuration data;

[0096] In the embodiment of the application, the Python programming language and the PyTorch framework are used to set the Monte Carlo sampling number parameter. In the specific operation, a variable num_samples is defined to store the sampling number, for example, set to 100 times.

[0097] Step S312: Perform multiple forward propagation calculations on the free gas structure mask according to the sampling number configuration data to obtain multiple groups of prediction probability maps;

[0098] In the embodiment of the application, after setting the Monte Carlo sampling number, the pre-trained deep learning model is used to perform multiple forward propagation calculations on the free gas structure mask. A pre-trained convolutional neural network (CNN) model in the PyTorch framework is used, which is specifically used to output the probability map of the free gas structure. In the specific operation, the free gas structure mask is input into the model, and forward propagation is performed according to the previously set sampling number (100 times). Each forward propagation generates a prediction probability map, and the shape of the probability map is (1, 1, 256, 256). After 100 times of forward propagation, 100 groups of prediction probability maps are collected.

[0099] Step S313: Perform pixel-level variance calculation on the multiple groups of prediction probability maps to obtain initial prediction variance data;

[0100] In the embodiment of the application, after obtaining 100 groups of prediction probability maps, pixel-level variance calculation needs to be performed on these data to obtain initial prediction variance data. The statistical function in the PyTorch framework is used to complete this task. In the specific operation, the 100 groups of prediction probability maps are integrated into a four-dimensional tensor with a shape of (100, 1, 256, 256). The variance of each pixel point is calculated along the sampling number dimension (the first dimension). Through the above steps, a variance map with a shape of (1, 1, 256, 256) is obtained, that is, the initial prediction variance data. This variance map reflects the change of each pixel point in multiple predictions.

[0101] Step S314: Perform normalization processing on the initial prediction variance data based on the signal-to-noise ratio of the seismic data to obtain prediction variance data.

[0102] In the embodiment of the present application, the initial prediction variance data is normalized based on the signal-to-noise ratio (SNR) of the seismic data. In the specific operation, the signal-to-noise ratio of the seismic data is calculated, and an SNR value of 5 is obtained by measuring the ratio of the signal power to the noise power. The initial prediction variance data is normalized using this SNR value. The specific method is to divide the variance value of each pixel point by the maximum value in the variance map, and then multiply by the SNR value. Through the above steps, the normalized prediction variance data is obtained, and the value is adjusted to a reasonable range.

[0103] Preferably, step S4 comprises the following steps:

[0104] Step S41: performing three-dimensional coordinate conversion on the optimized free gas structure mask to obtain an initial free gas structure model;

[0105] In the embodiment of the present application, the meshgrid and surf functions in MATLAB software are used for three-dimensional coordinate conversion. In the specific operation, the two-dimensional free gas structure mask data (for example, a matrix with a shape of 256x256) is loaded into MATLAB. The meshgrid function is used to generate corresponding three-dimensional coordinate grids, which define the position of the mask in three-dimensional space. The surf function is used to combine these coordinates and mask data to generate a three-dimensional surface model, i.e. the initial free gas structure model. This model intuitively shows the shape of the free gas structure in three-dimensional space.

[0106] Step S42: performing line direction interpolation calculation on the initial free gas structure model to obtain interpolation parameters;

[0107] In the embodiment of the present application, the interp2d function in the Python programming language and SciPy library is used for line direction interpolation calculation. In the specific operation, the direction of the survey line is determined, for example, along the x-axis direction. According to the data points of the initial model, the interp2d function is used to perform interpolation calculation in the direction of the survey line. The interpolation parameters are set, for example, the interpolation resolution (for example, interpolation once every 0.1 unit length). Through the above steps, the interpolation parameters are obtained, which define the distribution of the model data points in the direction of the survey line.

[0108] Step S43: performing spatial sampling on the initial free gas structure model according to the interpolation parameters to obtain a three-dimensional free gas structure point cloud;

[0109] In the embodiment of the present application, Python programming language and NumPy library are used for spatial sampling. In specific operation, data points are uniformly sampled in the three-dimensional space of the model according to interpolation parameters. For example, the sampling interval is set to 0.5 unit length. Through the above steps, three-dimensional free gas structure point cloud is obtained, and these point cloud data contain detailed position information of the model in the three-dimensional space.

[0110] Step S44: density analysis is performed on the three-dimensional free gas structure point cloud to obtain point cloud distribution feature data;

[0111] In the embodiment of the present application, Python programming language and gaussian_kde function in SciPy library are used for density analysis. In specific operation, point cloud data is loaded into the program, and gaussian_kde function is used to estimate the density of the point cloud. The bandwidth parameter of the density estimation is set (for example, the bandwidth is 0.2). Through the above steps, point cloud distribution feature data is obtained, which reflects the density distribution of the point cloud in different regions.

[0112] Step S45: based on the point cloud distribution feature data, Kriging spatial interpolation is performed to obtain continuous free gas structure data.

[0113] In the embodiment of the present application, Python programming language and PyKrige library are used for Kriging spatial interpolation. In specific operation, point cloud distribution feature data is loaded into the program, and kriging function in PyKrige library is used for spatial interpolation. A suitable Kriging model (for example, a spherical model) is selected, and the parameters of the model (for example, the range parameter is 10, and the nugget effect is 0.1) are set. Through the above steps, continuous free gas structure data is obtained, which can more accurately reflect the continuous form of free gas structure in the three-dimensional space.

[0114] Preferably, step S5 comprises the following steps:

[0115] Step S51: based on the continuous free gas structure data, isosurface extraction is performed to obtain an initial three-dimensional surface model;

[0116] In the embodiment of the present application, ParaView software is used for isosurface extraction. In specific operation, continuous free gas structure data is imported into ParaView. Contour filter of ParaView is used to extract isosurfaces. The value of the isosurface is set to 0.5 (indicating the probability threshold of free gas structure), and a suitable number of isosurfaces (for example, 10 isosurfaces) is selected. Through the above steps, an initial three-dimensional surface model is obtained, which intuitively shows the form and distribution of free gas structure in the three-dimensional space.

[0117] Step S52: topological structure correction is made to the initial three-dimensional surface model to obtain topological correction parameters;

[0118] In the embodiment of the present application, Mesh Lab software is used for topological structure correction. In the specific operation, the initial three-dimensional surface model is imported into Mesh Lab. The topological structure of the model is checked and corrected by using the Topology Cleaning tool of Mesh Lab. The parameters of topological correction are set, for example, Max Edge Length is 0.1 unit length, to ensure that the geometric details of the model are retained. Through the above steps, topological correction parameters are obtained.

[0119] Step S53: geometric repair is made to the initial three-dimensional surface model according to the topological correction parameters to obtain an optimized three-dimensional surface model;

[0120] In the embodiment of the present application, Mesh Lab software is used for geometric repair. In the specific operation, the Remeshing tool of Mesh Lab is used to make geometric repair to the model according to the topological correction parameters. The parameters of remeshing are set, for example, Target Number of Faces is 10000. Through the above steps, an optimized three-dimensional surface model is obtained, which is more accurate and stable in geometry and topology.

[0121] Step S54: manual correction is made to the optimized three-dimensional surface model to obtain interpretation correction data;

[0122] In the embodiment of the present application, 3ds Max software is used for manual correction. In the specific operation, the optimized three-dimensional surface model is imported into 3ds Max. The Edit Poly tool of 3ds Max is used to manually edit and correct the model. According to the experience and knowledge of geologists, some parts of the model are adjusted, for example, some too sharp edges are smoothed, and some missing parts are filled. Through the above steps, interpretation correction data is obtained.

[0123] Step S55: the optimized three-dimensional surface model is reconstructed based on the interpretation correction data to obtain a free gas structure three-dimensional model;

[0124] In the embodiment of the present application, Blender software is used to reconstruct the optimized three-dimensional surface model. In the specific operation, the interpretation correction data is imported into Blender. The Mesh tool of Blender is used to reconstruct the model according to the interpretation correction data. The parameters of reconstruction are set, for example, Subdivision Level is 3, to ensure the smoothness and details of the model. Through the above steps, a free gas structure three-dimensional model is obtained, which is more accurate in geometry and topology and can better reflect the actual morphology of the free gas structure.

[0125] Step S56: probability label assignment is performed on the free gas structure three-dimensional model to obtain a free gas structure three-dimensional model with probability labels.

[0126] In the embodiments of the present application, Python programming language and NumPy library are used to perform probability label assignment on the free gas structure three-dimensional model. In the specific operation, the data of the free gas structure three-dimensional model is loaded into Python. Using NumPy array operation, the probability value (for example, in the range from 0 to 1) of each vertex is assigned to the model. The threshold of the probability label is set, for example, 0.8, and the area with the probability value higher than 0.8 is marked as a high confidence area, and the area with the probability value lower than 0.8 is marked as a low confidence area. Through the above steps, the free gas structure three-dimensional model with probability labels is obtained, which not only shows the morphology of the free gas structure, but also provides the probability information of each area.

[0127] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, the scope of the present application is defined by the appended claims rather than the above description, and therefore all changes falling within the meaning and scope of the equivalent elements of the application file are intended to be included in the present application.

[0128] The above description is only a specific implementation of the present application, enabling those skilled in the art to understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying free gas structure seismic data based on a large model, characterized in that, The method comprises the following steps: Step S1: obtaining original seismic data; performing short-time Fourier transform on the original seismic data to obtain time-frequency domain seismic data; performing Hilbert-Huang transform on the time-frequency domain seismic data to obtain enhanced seismic feature data; Step S2: performing feature extraction on the enhanced seismic feature data to obtain a multi-scale seismic feature map; performing semantic segmentation on the enhanced seismic feature data to obtain a free gas structure mask; Step S3: performing uncertainty evaluation on the free gas structure mask to obtain a confidence heat map; performing probability prediction on the confidence heat map to obtain a low confidence area label map; performing data enhancement on the low confidence area label map to obtain corrected seismic feature data; inputting the corrected seismic feature data into a pre-trained visual Transformer large model to perform feature extraction, and obtaining an optimized free gas structure mask; Step S4: based on the optimized free gas structure mask, performing three-dimensional space reconstruction to obtain an initial free gas structure model; performing line interpolation on the optimized free gas structure mask to obtain a three-dimensional free gas structure point cloud; performing spatial fitting on the three-dimensional free gas structure point cloud to obtain continuous free gas structure body data; Step S5: performing interactive interpretation on the continuous free gas structure body data to obtain a free gas structure three-dimensional model; based on the free gas structure three-dimensional model, performing visual rendering to obtain a free gas structure three-dimensional model with probability labels.

2. The large model-based free gas structure seismic data identification method of claim 1, wherein, Step S1 comprises the following steps: Step S11: obtaining original seismic data through a seismic acquisition device; Step S12: performing wavelet threshold denoising on the original seismic data to obtain denoised seismic data; Step S13: performing time window division based on the denoised seismic data to obtain segmented seismic data; Step S14: performing short-time Fourier transform on the segmented seismic data to obtain time-frequency domain seismic data; Step S15: performing Hilbert-Huang transform on the time-frequency domain seismic data to obtain instantaneous frequency feature data; Step S16: performing normalization processing on the instantaneous frequency feature data to obtain enhanced seismic feature data.

3. The large model based free gas structure identification method of claim 2, wherein, Step S13 comprises the following steps: Step S130: performing seismic wave speed identification on the denoised seismic data to obtain seismic wave propagation speed; Step S131: performing main frequency calculation based on the seismic wave propagation speed to obtain seismic wave main frequency characteristic data; setting time window parameters according to the seismic wave main frequency characteristic data to obtain time window length configuration data; Step S132: performing overlap rate calculation based on the time window length configuration data on the denoised seismic data to obtain time window overlap parameters; Step S133: performing sliding time window segmentation on the denoised seismic data by using the time window overlap parameters to obtain initial segmented seismic data; Step S134: performing energy equalization processing on the initial segmented seismic data to obtain equalized segmented seismic data; Step S135: performing quality screening on the equalized segmented seismic data based on a preset signal-to-noise ratio threshold to obtain effective segmented seismic data; Step S136: performing time marking on the effective segmented seismic data to obtain segmented seismic data.

4. The large model based free gas structure identification method of claim 1, wherein, Step S2 comprises the following steps: Step S21: performing channel dimension expansion on the enhanced seismic feature data to obtain three-dimensional seismic feature data; Step S22: Use a pre-trained visual Transformer model to extract features from the 3D seismic feature data to obtain a primary seismic feature map; Step S23: Enhance the primary seismic feature map to obtain a multi-scale seismic feature map; Step S24: Upsample and reconstruct the multi-scale seismic feature map to obtain a semantic segmentation probability map; Step S25: Binarize the semantic segmentation probability map based on a preset threshold to obtain a free gas structure mask.

5. The large model based free gas structure identification method of claim 4, wherein, Step S23 includes the following steps: Step S231: Perform pyramid pooling on the primary seismic feature map to obtain multi-scale contextual feature data; Step S232: Calculate feature weights for multi-scale contextual feature data using a spatial attention mechanism to obtain a spatial attention weight map; Step S233: Recalibrate the primary seismic feature map based on the spatial attention weight map to obtain the spatially enhanced feature map; Step S234: Perform channel correlation analysis on the spatial augmentation feature map using the channel attention mechanism to obtain the channel attention weight map; Step S235: Perform channel feature enhancement on the spatial enhancement feature map based on the channel attention weight map to obtain a multi-scale seismic feature map.

6. The large model based free gas structure identification method for seismic data according to claim 1, wherein, Step S3 includes the following steps: Step S31: Perform multiple predictions on the free gas structure mask based on the Monte Carlo method to obtain prediction variance data; Step S32: Perform Gaussian smoothing on the predicted variance data to obtain a confidence heatmap; Step S33: Perform binary segmentation on the confidence heatmap according to the preset confidence threshold to obtain a low confidence region marking map; Step S34: Use a pre-defined adversarial generative network to complete the features of the low-confidence region label map to obtain enhanced feature data; Step S35: Perform feature fusion between the enhanced feature data and the original seismic feature data to obtain the corrected seismic feature data; Step S36: Input the corrected seismic feature data into the pre-trained visual Transformer large model for iterative optimization to obtain the optimized free gas structure mask.

7. The large model based free gas structure identification method of claim 6, wherein, Step S31 includes the following steps: Step S311: Set the Monte Carlo sampling number parameter to obtain the sampling number configuration data; Step S312: Perform multiple forward propagation calculations on the free gas structure mask based on the data configured by the number of samplings to obtain multiple sets of predicted probability maps; Step S313: Calculate the pixel-level variance of multiple sets of prediction probability maps to obtain the initial prediction variance data; Step S314: Normalize the initial prediction variance data based on the signal-to-noise ratio of the seismic data to obtain the prediction variance data.

8. The large model based free gas structure identification method of claim 1, wherein, Step S4 includes the following steps: Step S41: Perform a three-dimensional coordinate transformation on the optimized free gas structure mask to obtain the initial free gas structure model; Step S42: Perform survey line direction interpolation calculations on the initial free gas structure model to obtain interpolation parameters; Step S43: Spatial sampling is performed on the initial free gas structure model according to the interpolation parameters to obtain a three-dimensional free gas structure point cloud; Step S44: Perform density analysis on the three-dimensional free gas structure point cloud to obtain point cloud distribution characteristic data; Step S45: Perform Kriging space interpolation based on point cloud distribution feature data to obtain continuous free gas structure data.

9. The large model based free gas structure identification method of claim 1, wherein, Step S5 includes the following steps: Step S51: Extract isosurfaces based on continuous free gas structure data to obtain an initial three-dimensional surface model; Step S52: Perform topology correction on the initial 3D surface model to obtain topology correction parameters; Step S53: Perform geometric repair on the initial 3D surface model according to the topology correction parameters to obtain an optimized 3D surface model; Step S54: Manually correct the optimized 3D surface model to obtain interpretation and correction data; Step S55: Reconstruct the optimized three-dimensional surface model based on the interpretation and correction data to obtain a three-dimensional model of free gas structure; Step S56: Assign probability labels to the three-dimensional model of free gas structure to obtain a three-dimensional model of free gas structure with probability labels.

10. A free gas structure seismic data identification system based on a large model, characterized in that, For executing the large-model-based free gas structure seismic data identification method as described in claim 1, the large-model-based free gas structure seismic data identification system comprises: The data preprocessing module is used to acquire raw seismic data; perform short-time Fourier transform on the raw seismic data to obtain time-frequency domain seismic data; and perform Hilbert-Huang transform on the time-frequency domain seismic data to obtain enhanced seismic feature data. The feature extraction module is used to extract features from the enhanced seismic feature data to obtain a multi-scale seismic feature map; and to perform semantic segmentation on the enhanced seismic feature data to obtain a free gas structure mask. The optimization iteration module is used to evaluate the uncertainty of the free gas structure mask and obtain a confidence heatmap; perform probability prediction on the confidence heatmap to obtain a low-confidence area marker map; perform data augmentation on the low-confidence area marker map to obtain corrected seismic feature data; input the corrected seismic feature data into a pre-trained visual Transformer large model for feature extraction to obtain the optimized free gas structure mask. The 3D modeling module is used to reconstruct the 3D space based on the optimized free gas structure mask to obtain the initial free gas structure model; perform survey line interpolation on the optimized free gas structure mask to obtain the 3D free gas structure point cloud; and perform spatial fitting on the 3D free gas structure point cloud to obtain continuous free gas structure data. The visualization module is used to interactively interpret continuous free gas structure data to obtain a three-dimensional model of the free gas structure; based on the three-dimensional model of the free gas structure, a visualization rendering is performed to obtain a three-dimensional model of the free gas structure with probability labels.

Images