A coalfield three-dimensional seismic fault identification method, device, medium and equipment

Abstract

This invention discloses a method, apparatus, medium, and equipment for three-dimensional seismic fault identification in coalfields, relating to the field of fault identification. The method includes: acquiring raw three-dimensional seismic data of the area to be measured; denoising the raw three-dimensional seismic data using a deep learning model to determine the denoised three-dimensional seismic data; constructing a ResU-Net model for coalfield seismic fault identification; training the ResU-Net model; using the trained ResU-Net model to perform fault identification on the denoised three-dimensional seismic data, generating a fault probability volume; retraining the trained ResU-Net model using a training set to iteratively optimize the fault probability volume, determining the optimized fault probability volume; and performing fault identification based on the optimized fault probability volume and geological interpretation requirements to obtain the identification results.

Description

Technical Field

[0001] This invention relates to the field of fault identification, and in particular to a method, apparatus, medium, and equipment for three-dimensional seismic fault identification in coalfields. Background Technology

[0002] As coal mining extends to deeper areas and geological conditions become increasingly complex, accurate identification of small-scale faults has become a key geological challenge for ensuring safe and efficient coal mining and supporting the construction of smart mines.

[0003] However, in the field of coalfield exploration, the requirements for fault interpretation accuracy are increasingly stringent. Currently, many mining areas have listed faults with a displacement of less than 5 meters as a mandatory geological task for detection (Zou Guangui et al., 2025). The seismic phase axis distortion caused by these small-scale faults is very weak and highly concealed. In addition, the lack of real small-scale fault samples available for training often leads to unsatisfactory practical application results of existing prediction methods. Although artificially synthesized fault data can assist in methodological research to some extent, there are still differences between them and actual coalfield fault data and geological characteristics, making it difficult to meet the requirements for accurate identification of different types of small-scale faults. Summary of the Invention

[0004] This invention provides a method, apparatus, medium, and device for three-dimensional seismic fault identification in coalfields, to solve the aforementioned problems in the prior art, namely, how to improve the accuracy of small-scale seismic fault identification in the prior art. This invention provides a method for three-dimensional seismic fault identification in coalfields, which includes: Acquire the original 3D seismic data of the area to be measured; use a deep learning model to denoise the original 3D seismic data, and determine the denoised 3D seismic data. A ResU-Net model for coalfield seismic fault identification is constructed; wherein, the convolutional blocks in the original U-Net model are replaced with residual modules. The ResU-Net model was trained using the FaultSeg3D synthetic dataset with known fault features generated by simulation and historical 3D seismic data of the area to be measured. The trained ResU-Net model was then used to identify faults in the denoised 3D seismic data and generate a fault probability volume. Positive samples are generated by manually labeling faults in 3D seismic data that are not identified by the ResU-Net model; negative samples are generated by manually labeling non-fault areas in 3D seismic data that are misjudged by the ResU-Net model; positive and negative samples are used as training sets, and the trained ResU-Net model is retrained using the training sets to iteratively optimize the fault probability volume and determine the optimized fault probability volume. Based on the optimized fault probability volume and combined with geological interpretation requirements, fault identification is performed, and the identification results are obtained.

[0005] Optionally, the ResU-Net model includes an encoder, a skip connection layer, and a decoder connected in sequence; wherein, the encoder extracts multi-scale cross-sectional features from the original 3D seismic data and downsamples the multi-scale cross-sectional features; the decoder's 3D deconvolution upsamples the multi-scale cross-sectional features to determine the feature map after size restoration.

[0006] Optionally, the encoder specifically includes a 3D convolution module (conv3D), a ReLU activation function module, and a downsampling module connected in sequence.

[0007] Optionally, the geological interpretation requirements specifically include: Requirements for fault geometry, requirements for coal seam mineability, and requirements for data self-consistency.

[0008] Optionally, before denoising the original 3D seismic data, the original 3D seismic data may be processed sequentially using dip-guided filtering and fault-focused imaging techniques.

[0009] Optionally, the dip-guided filtering constructs a spatially constrained guide body by extracting the local dip and azimuth of the seismic phase axis, so as to adaptively process and filter the original three-dimensional seismic data along the stratigraphic trend.

[0010] Optionally, the fault-focusing imaging performs fault-focusing imaging processing on the faults of the original three-dimensional seismic data by setting the main frequency to 40Hz, the low frequency to 15Hz, and the high frequency to 60Hz.

[0011] This invention provides a three-dimensional seismic fault identification device for coalfields, comprising: The acquisition module is used to acquire the original 3D seismic data of the area to be measured; a deep learning model is used to denoise the original 3D seismic data to determine the denoised 3D seismic data. A building module is used to construct a ResU-Net model for coalfield seismic fault identification; wherein, the ResU-Net model replaces the convolutional blocks in the original U-Net model with residual modules; The fault probability volume generation module is used to train the ResU-Net model using the FaultSeg3D synthetic dataset with known fault features generated by simulation and the historical 3D seismic data of the area to be measured. The trained ResU-Net model is then used to identify faults in the denoised 3D seismic data and generate the fault probability volume. The optimization module is used to generate positive samples by manually marking faults in 3D seismic data that are not recognized by the ResU-Net model; and to generate negative samples by manually marking non-fault areas in 3D seismic data that are misjudged by the ResU-Net model. The positive and negative samples are used as a training set to retrain the trained ResU-Net model to iteratively optimize the fault probability volume and determine the optimized fault probability volume. The fault identification module is used to identify faults based on the optimized fault probability volume and geological interpretation requirements, and to obtain the identification results.

[0012] The present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for identifying three-dimensional seismic faults in coalfields.

[0013] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the above-mentioned three-dimensional seismic fault identification method for coalfields.

[0014] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention provides a three-dimensional seismic fault identification method for coalfields. This method enhances the model's ability to identify weak fault features and its cross-data domain adaptability by training a constructed ResU-Net model, thereby improving the accuracy of subsequent small-scale fault identification. Then, using the trained ResU-Net model, fault identification is performed on the seismic data to be tested, generating a fault probability volume. The fault probability volume reflects the probability of faults existing at various locations in the seismic data, providing a foundation for subsequent fault identification. By iteratively optimizing the fault probability volume using a feedback optimization model based on transfer learning, the main faults can be clearly reproduced, revealing a large number of continuous and reasonably shaped secondary small faults, providing detailed and highly reliable fault identification images, thereby improving the accuracy of small-scale fault identification. Attached Figure Description

[0015] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0016] Figure 1 A flowchart of a three-dimensional seismic fault identification method for coalfields provided in an embodiment of the present invention; Figure 2 This is a comparison of the typical seismic profile of the test area before and after Inline487 processing provided in an embodiment of the present invention; Figure 3This is a comparison of the typical seismic profile of the area to be tested before and after Inline656 processing provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the Res-Unet model architecture provided in an embodiment of the present invention; Figure 5 This invention provides an intelligent identification process for small-scale faults in three-dimensional seismic data in coalfields. Figure 6 Comparison of the effects of different fault identification methods on a typical seismic profile (inline534) of the test area provided in this embodiment of the invention; Figure 7 Comparison of fault plane identification effects of different methods in the No. 2 coal seam of the test area provided in the embodiments of the present invention; Figure 8 A comparison of the fault identification effect of the working face 1022101 in the test area provided in this embodiment of the invention with the fault profile revealed by the roadway exploration and mining; Figure 9 A comparison of the fault identification effect of the working face 1012006 in the test area provided in this embodiment of the invention with the fault profile revealed by the roadway exploration and mining; Figure 10 A comparison of the fault identification effect of the working face 1012007 in the test area provided in this embodiment of the invention with the fault profile revealed by the roadway exploration and mining; Figure 11 A schematic diagram of a computer device for a three-dimensional seismic fault identification method for coalfields provided in an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions 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. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0018] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0019] Figure 1 This is a flowchart of a three-dimensional seismic fault identification method for coalfields provided in an embodiment of the present invention, as shown below. Figure 1 As shown in the figure, this embodiment illustrates a method for identifying three-dimensional seismic faults in coalfields, including: S1: Obtain the original 3D seismic data of the area to be measured; use a deep learning model to denoise the original 3D seismic data and determine the denoised 3D seismic data.

[0020] For example, by comparing different seismic profiles Inline487 and Inline656 before and after fault enhancement processing, it can be seen that the original profile has a low signal-to-noise ratio, insufficient vertical resolution, and most faults are not clearly identified (as shown in the figures below). Figure 2 (a) and Figure 3 As shown in (a). After structurally guided filtering, profile noise is suppressed, the quality of low signal-to-noise ratio seismic data is significantly improved, the continuity of phase axes is enhanced, and the accuracy of fault boundary identification is improved (e.g., Figure 2 (b) and Figure 3 As shown in (b). After further processing with fault-focused imaging technology, the signal-to-noise ratio of the seismic data was slightly improved, but the vertical resolution was significantly improved, the fault convergence was clearer, and the fault features were more prominent (as shown in (b)). Figure 2 (c) and Figure 3 (as shown in (c)).

[0021] For example, the present invention can use a deep learning model, such as a convolutional neural network (CNN), to denoise the original 3D seismic data. A comparison of different seismic profiles Inline487 and Inline656 before and after deep learning denoising can be seen ( Figure 2 (d) and Figure 3 The signal-to-noise ratio (S / N) improved from 67.3 and 46.6 to 107.2 and 142.1, respectively, resulting in a significant improvement in data quality. Specifically, this is manifested in enhanced effective reflection energy on seismic profiles, improved continuity of seismic phase axes, and clearer and more accurate display of fault location, strike, dip, and displacement, providing a more reliable data foundation for the accuracy of fine structural interpretation.

[0022] S2: Construct a ResU-Net model for coalfield seismic fault identification; wherein, the ResU-Net model replaces the convolutional blocks in the original U-Net model with residual modules.

[0023] For example, ResU-Net is a convolutional neural network that deeply integrates the U-Net encoder-decoder architecture with the ResNet residual learning mechanism. It inherits the fine-grained segmentation capabilities of U-Net and the stable training characteristics of ResNet, enabling more accurate and continuous identification of faults in seismic data. Figure 4As shown, the processing flow of this network is as follows: First, the input 3D seismic data is processed by the encoder, then sequentially through 3D convolution (Conv3D), batch normalization (BatchNorm), and ReLU activation function modules, and then downsampled through a max pooling (MaxPool) layer. This process reduces the spatial size while increasing the number of channels (64->128->...->512), gradually extracting the abstracted multi-scale cross-sectional features and completing the extraction of high-level pixel features. Then, the data enters the right-side decoder, where it is upsampled through 3D deconvolution (such as "DeConv3D k4s2n1024") to gradually restore the spatial size of the feature map. The encoder and decoder fuse multi-scale features through skip connections, enabling it to capture large-scale fault contours while preserving local details such as fault points and fault displacements, significantly improving the accuracy of fault boundary characterization. Meanwhile, the residual modules introduced into the network (such as "Residual Block-128") can effectively alleviate the gradient vanishing problem in deep networks through quick connections, enabling the network to be built deeper and more stable, thereby enhancing its ability to learn more complex fault features. Overall, the ResU-Net network, based on the strategy of "pre-training + iterative transfer learning on real data," enables the model to gradually adapt to the distribution differences between the model and real data, thereby continuously enhancing its ability to characterize the fault features of specific work areas.

[0024] For example, the encoder part of ResU-Net adopts a multi-level feature extraction architecture. Each encoding stage includes 3D convolution (Conv3D), batch normalization (BatchNorm), ReLU activation function, and downsampling module, and gradient propagation is enhanced through residual connections. Based on the network structure shown in the attached figure, the specific data processing flow of the encoder is as follows:

[0025] (1) Input layer: The encoder input is the preprocessed three-dimensional seismic data volume with a size of 128×128×128 (depth×height×width) and 1 channel (i.e. the original seismic amplitude data).

[0026] (2) First-level coding stage: (a) 3D Convolutional Layer (Conv3D k4s1n64): A 3D convolutional layer with a kernel size of 4×4×4 (k4), a stride of 1 (s1), and 64 output channels (n64) is used to perform convolution operations on the input data. This layer expands the number of feature channels to 64 while maintaining the same spatial size (128×128×128), extracting low-level features such as local edges and textures from the seismic data.

[0027] (b) Batch Normalization: Batch normalization is performed on the 64-channel feature map output by the convolutional layer, adjusting the feature values ​​of each channel to a distribution with a mean of 0 and a variance of 1, thereby accelerating network convergence and improving training stability.

[0028] (c) ReLU activation function: Apply the linear rectified function (ReLU) to the normalized feature map to introduce non-linear expressive power, set negative values ​​to zero and retain positive values.

[0029] (d) Downsampling Module (MaxPool 2×2×2): The feature map is downsampled using 3D max pooling with a pooling window of 2×2×2 and a stride of 2. After pooling, the spatial size of the feature map is halved, decreasing from 128×128×128 to 64×64×64, while the number of channels remains unchanged at 64. The downsampling operation expands the receptive field of subsequent convolutions while reducing computational complexity.

[0030] (3) Second-level coding stage (Residual Block-128) After the first stage of downsampling, the feature map size is 64×64×64 with 64 channels, and it is input into the residual block Residual Block-128. The structural details of this residual block are shown in the figure, and it includes the following operations:

[0031] (a) First convolutional layer (Conv3D k1s1n128): A 3D convolution with a kernel size of 1×1×1 (k1), a stride of 1 (s1), and 128 output channels (n128) is used. The 1×1×1 convolution adjusts the number of channels, increasing it from 64 to 128 while maintaining the spatial dimensions (still 64×64×64). The output feature map size of this layer becomes 64×64×64×128.

[0032] (b) Residual connection design: Residual blocks typically contain multiple convolutional layers ( Figure 4 The details are not fully elaborated here, but the standard structure of a Residual Block includes two convolutional layers and skip connections. Its core idea is to directly add the input of the current block to its output, forming an identity mapping. Mathematically, the output of the residual block... ,in For the residual mapping that needs to be learned, This is the input. This design allows gradients to flow directly to the lower layers, effectively alleviating the vanishing gradient problem in deep network training, while also helping to preserve detailed information.

[0033] (c) Output: The feature map size after residual block processing is 64×64×64×128.

[0034] (4) Third-level coding stage (Residual Block-256).

[0035] (a) The feature map (64×64×64×128) output from the second stage is input into Residual Block-256: Convolutional layer (Conv3D k1s1n256): Also uses 1×1×1 convolution with a stride of 1, increasing the number of output channels to 256. At this time, the feature map spatial size remains 64×64×64, but the number of channels is expanded to 256, i.e., 64×64×64×256.

[0036] (b) Residual Block Processing: Inside Residual Block-256, higher-level abstract features are further extracted through multiple convolutional layers, and the input (after channel adjustment) is added to the output through residual connections.

[0037] (c) Encoder Output: After the above multi-level encoding, the final output feature map size is 64×64×64×256. This feature map also has the following characteristics:

[0038] (d) Spatial dimension compression: The original input was compressed from 128×128×128 to 64×64×64, which expanded the receptive field and allowed each voxel to contain a wider range of contextual information.

[0039] (e) Channel dimension expansion: from a single channel to 256 channels, each channel represents an abstract feature that can characterize high-level semantic information of complex geological bodies such as faults and stratigraphic horizons.

[0040] The feature map output by the encoder will serve as a bridge, passing it to the corresponding layer of the decoder through skip connections to help recover detailed information. On the other hand, it will serve as the input to the bottleneck layer, allowing the decoder to gradually upsample and recover to the original resolution, ultimately outputting the tomographic probability volume.

[0041] For example, the encoder, connected to the decoder's deconvolutional layers (DeConv3D k4s2n128 and DeConv3Dk4s2n256), progressively restores the low-resolution, high-dimensional features output by the encoder to high resolution, ultimately outputting a tomographic probability volume with the same size as the input. The deconvolutional layers use a 4×4×4 kernel size and a stride of 2, doubling the feature map space size while correspondingly reducing the number of channels, forming a symmetrical structure with the encoder.

[0042] S3: The ResU-Net model is trained using the FaultSeg3D synthetic dataset with known fault features generated by simulation and historical 3D seismic data of the area to be measured. The trained ResU-Net model is then used to identify faults in the denoised 3D seismic data and generate a fault probability volume.

[0043] For example, the original 3D seismic data can be processed sequentially with dip-guided filtering and fault-focused imaging techniques before denoising.

[0044] Generally, fault-focused imaging is a fault enhancement technique based on frequency band optimization. Its core principle is to highlight the spatial distribution characteristics of faults by strengthening dominant frequency bands and suppressing noise. This method is based on the differences in the tuning response of different frequency seismic data to faults and strata thickness: low-frequency components are generally suitable for characterizing the overall outline of large faults, while high-frequency components are more sensitive to the identification of small faults. Fault-focused imaging technology mainly involves three parameters: low frequency, dominant frequency, and high frequency. Based on the spectral analysis of coal seam segments in seismic data, the main minable coal seam segments in the area to be measured have a frequency distribution range of approximately 10-100Hz, with a dominant frequency of approximately 40Hz. Therefore, based on dip-guided filtering, the dominant frequency is set to 40Hz, the low frequency to 15Hz, and the high frequency to 60Hz for fault-focused imaging fault processing to further enhance the subsequent fault identification effect.

[0045] For example, based on the characteristics of 3D seismic data from coal mines, the following formula was developed: Figure 5 The flowchart illustrates the small fault identification and prediction process. First, the denoised 3D seismic data volume is processed sequentially using dip-guided filtering and fault-focusing imaging techniques. Further, a deep learning model is employed for denoising to enhance fault response and suppress background noise. Based on this, the FaultSeg3D synthetic dataset is fused with historical 3D seismic data (confirmed by faults exposed in mine shafts) from multiple mining areas. A ResU-Net model is then jointly trained to generate a fault probability volume.

[0046] S4: Positive samples are generated by manually labeling faults in the 3D seismic data that are not identified by the ResU-Net model; negative samples are generated by manually labeling non-fault areas in the 3D seismic data that are misjudged by the ResU-Net model; positive and negative samples are used as training sets, and the trained ResU-Net model is retrained using the training sets to iteratively optimize the fault probability volume and determine the optimized fault probability volume.

[0047] For example, to adapt to the structural differences in different mining areas, especially small faults, a cross-validation sample labeling strategy can be adopted to continuously update the labeled sample library during iterative training and recalculate the fault probability volume. Specifically, this can include: Geological interpreters combine traditional attribute volumes with their professional knowledge to examine the initial fault probability volume. Judgment criteria include: fault continuity, reasonable attitude, and excessive false positives / noise. If the results align with geological understanding, the process proceeds directly to the final effect verification stage; otherwise, an optimization cycle is initiated.

[0048] Geological interpretation experts manually corrected the shortcomings of the current probabilistic volume on the 3D interpretation platform: positive samples were added to real faults that the model missed, and negative samples were added to noise points that were misclassified. These interactive labels constituted a high-quality feedback dataset.

[0049] The newly labeled samples are added to the training set to perform incremental training (or online learning) on ​​the current model. Since the model already has basic capabilities, only a few iterations are needed to update the weights at this stage, allowing the model to correct previous errors in the next round of predictions.

[0050] The updated model is used to recalculate the entire data volume, generating a new generation of optimized fault probability volumes. A conformity assessment is then performed again. This process is repeated, with each iteration making the model more closely reflect the actual geological characteristics of the work area.

[0051] The iterative process continues until any of the following stopping criteria are met: (1) Model performance convergence: the accuracy and recall on the validation set no longer improve; (2) Geological interpretation personnel confirmation: geological experts believe that the current fault probability volume has accurately characterized the main faults and minor faults without significant errors; (3) Quantitative indicators meet the standards: for example, the fault continuity index, signal-to-noise ratio, etc. reach the preset threshold.

[0052] Once the conditions are met, the fault probability volume generated in the last iteration is determined as the final optimized fault probability volume. This data volume then enters the effect comparison and analysis stage, where it is compared and verified with the coherent attribute volume and actual exploration and mining fault data in the roadway. This confirms its advantages in fault identification accuracy and reliability, and it is ultimately used for characterizing and analyzing faults at different scales to guide coalfield mining design.

[0053] Through multiple iterations, the model gradually adapts to the differences in the distribution of real data, enhances its ability to characterize the fault features of specific work areas, and finally obtains an optimized fault probability volume that meets the requirements of geological interpretation, enabling reliable identification and prediction of small-scale faults.

[0054] S5: Based on the optimized fault probability volume and combined with geological interpretation requirements, fault identification is performed, and the identification results are obtained.

[0055] For example, in coalfield geological exploration, the accuracy requirement for small fault identification is particularly prominent. To evaluate the effectiveness and superiority of ResU-Net deep learning technology in coalfield fault detection methods, this invention also selects commonly used attributes such as coherence, time-frequency continuous wavelet transform (TFCWT) frequency division, root mean square amplitude, and maximum likelihood volume for fine fault identification in the 3D seismic data of the area to be tested, and compares and analyzes their application effects.

[0056] like Figure 6 As shown, taking the Inline534 profile as an example, although the eigenvalues ​​are coherent, the TFCWT wavelet frequency division of 40Hz and the maximum likelihood volume ( Figure 6 (a) to Figure 6 As shown in (c), it can reliably delineate the outlines of major faults, but its results are not sensitive enough to minor faults, and the fault network presented is relatively coarse, making it difficult to meet the needs of fine interpretation. Furthermore, the maximum likelihood attribute image contains a large amount of potential subtle fault information, but its fatal flaw is its low signal-to-noise ratio, generating too much meaningless noise and artifacts, which seriously affects the efficiency and reliability of geological interpretation. In contrast, the ResU-Net deep learning technology stands out, achieving an excellent balance between sensitivity and signal-to-noise ratio. Figure 6 As shown in (d), it not only clearly reproduces all the major faults, but also reveals a large number of continuous, rationally shaped secondary faults. These details are either blurry or obscured by noise in images from other methods. This profoundly demonstrates the core advantage of deep learning: it can intelligently suppress background interference by learning complex data features, while enhancing effective structural signals, ultimately generating a fault identification image that is rich in detail and highly reliable, providing stronger technical support for the fine structural interpretation of coalfields and safe production.

[0057] For example, "geological interpretation requirements" are the core basis for manual judgment and model optimization, specifically including the following three aspects: (1) Fault geometry requirements: The strike, dip and dip angle of the identified faults must be consistent with the regional tectonic stress field background of the target work area. The intersecting relationship between faults must conform to the tectonic evolution logic and there should be no abnormal occurrences.

[0058] (2) Requirements for coal seam mineability: The fault must clearly cut off the reflected wave of the coal seam, the fault displacement accuracy must meet the requirements of the reserve calculation error, and it must be able to identify small faults with a drop greater than the mining design threshold, so as to ensure that it has practical guiding significance for the layout of the fully mechanized mining face.

[0059] (3) Data self-consistency requirement: The fault interpretation results are closed in three-dimensional space (i.e., there is no contradiction in the stratigraphic tracking), and the fault data list is completely consistent with the map representation, so as to ensure the reliability and verifiability of the results.

[0060] Figure 7 The comparison of the fault plane identification performance of different methods is shown. Figure 7 (a) is a planar view of the eigenvalue coherence tomography identification effect; Figure 7 (b) Planar view of the maximum likelihood attribute tomography identification effect; Figure 7 (c) Root mean square amplitude fault identification plan view; Figure 7 (d) The tomographic diagram of the ResU-Net model proposed in this invention. The comparison shows that the tomographic recognition effect of the ResU-Net model proposed in this invention is better than other methods. Figure 8 This invention provides a comparison of fault identification results at different scales in the 1022101 working face of the test area with faults revealed during roadway exploration and mining, as provided in this embodiment. Figure 8 (a) is a cross-sectional view of the tomographic identification of the ResU-Net model; Figure 8 (b) is a 40Hz fault identification profile of TFCWT wavelet frequency division; Figure 8 (c) is a cross-sectional view of the fault revealed during tunnel exploration and mining; Figure 9 This invention provides a comparison of fault identification results at different scales in the 1012006 working face of the test area with faults revealed during roadway exploration and mining, as provided in this embodiment. Figure 9 (a) is a cross-sectional view of the tomographic identification of the ResU-Net model; Figure 9 (b) is a 40Hz fault identification profile of TFCWT wavelet frequency division; Figure 9 (c) is a cross-sectional view of the fault revealed during tunnel exploration and mining; Figure 10 This invention provides a comparison between the fault identification effect at different scales in the 1022101 working face of the test area and the faults revealed during actual exploration and mining; wherein, Figure 10 (a) is a cross-sectional view of the tomographic identification of the ResU-Net model; Figure 10 (b) is a 40Hz fault identification profile of TFCWT wavelet frequency division; Figure 10 (c) is a cross-sectional view of the fault revealed during tunnel exploration and mining; by Figures 8 to 10 It can be seen that the ResU-Net model proposed in this invention has better performance in fault profile recognition than other methods.

[0061] For example, the effectiveness of the proposed method is verified by comparing the identification results of the ResU-Net model with those of traditional methods and fault information revealed in actual tunnels. Comparing the fault profile identification performance of four methods—eigenvalue coherence, time-frequency continuous wavelet transform (TFCWT) frequency division, root mean square amplitude, maximum likelihood attribute, and deep learning—a clear conclusion can be drawn: the overall performance of the proposed ResU-Net model is significantly superior to the other three traditional attribute methods.

[0062] The above are one or more embodiments of the coalfield three-dimensional seismic fault identification method provided in this specification. Based on the same idea, this specification also provides a corresponding coalfield three-dimensional seismic fault identification device, including: The acquisition module is used to acquire the original 3D seismic data of the area to be measured; a deep learning model is used to denoise the original 3D seismic data to determine the denoised 3D seismic data. A building module is used to construct a ResU-Net model for coalfield seismic fault identification; wherein, the ResU-Net model replaces the convolutional blocks in the original U-Net model with residual modules; The fault probability volume generation module is used to train the ResU-Net model using the FaultSeg3D synthetic dataset with known fault features generated by simulation and the historical 3D seismic data of the area to be measured. The trained ResU-Net model is then used to identify faults in the denoised 3D seismic data and generate the fault probability volume. The optimization module is used to generate positive samples by manually marking faults in 3D seismic data that are not recognized by the ResU-Net model; and to generate negative samples by manually marking non-fault areas in 3D seismic data that are misjudged by the ResU-Net model. The positive and negative samples are used as a training set to retrain the trained ResU-Net model to iteratively optimize the fault probability volume and determine the optimized fault probability volume. The fault identification module is used to identify faults based on the optimized fault probability volume and geological interpretation requirements, and to obtain the identification results.

[0063] Specific limitations regarding the coalfield 3D seismic fault identification device can be found in the limitations of the coalfield 3D seismic fault identification method described above, and will not be repeated here. Each module in the aforementioned coalfield 3D seismic fault identification device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0064] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described coalfield three-dimensional seismic fault identification method.

[0065] The present invention also provides Figure 11 The schematic diagram of the computer device shown is as follows: Figure 11As shown, at the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to implement the coalfield three-dimensional seismic fault identification method provided in the above embodiment.

[0066] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0067] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

Claims

1. A method for identifying three-dimensional seismic faults in coalfields, characterized in that, include: Acquire the original 3D seismic data of the area to be measured; use a deep learning model to denoise the original 3D seismic data, and determine the denoised 3D seismic data. A ResU-Net model for coalfield seismic fault identification is constructed; wherein, the convolutional blocks in the original U-Net model are replaced with residual modules. The ResU-Net model was trained using the FaultSeg3D synthetic dataset with known fault features generated by simulation and historical 3D seismic data of the area to be measured. The trained ResU-Net model was then used to identify faults in the denoised 3D seismic data and generate a fault probability volume. Positive samples are generated by manually labeling faults in 3D seismic data that are not identified by the ResU-Net model; negative samples are generated by manually labeling non-fault areas in 3D seismic data that are misjudged by the ResU-Net model; positive and negative samples are used as training sets, and the trained ResU-Net model is retrained using the training sets to iteratively optimize the fault probability volume and determine the optimized fault probability volume. Based on the optimized fault probability volume and combined with geological interpretation requirements, fault identification is performed, and the identification results are obtained.

2. The method for identifying three-dimensional seismic faults in coalfields as described in claim 1, characterized in that, The ResU-Net model comprises an encoder, skip connection layers, and a decoder connected in sequence. The encoder extracts multi-scale cross-sectional features from the original 3D seismic data and downsamples these features. The decoder uses 3D deconvolution to upsample the multi-scale cross-sectional features and determine the size-restored feature map.

3. The method for identifying three-dimensional seismic faults in coalfields as described in claim 2, characterized in that, The encoder specifically includes a 3D convolution module (conv3D), a ReLU activation function module, and a downsampling module connected in sequence.

4. The method for identifying three-dimensional seismic faults in coalfields as described in claim 1, characterized in that, The geological interpretation requirements specifically include: Requirements for fault geometry, requirements for coal seam mineability, and requirements for data self-consistency.

5. The method for identifying three-dimensional seismic faults in coalfields as described in claim 1, characterized in that, Before denoising the raw 3D seismic data, the raw 3D seismic data is sequentially processed using dip-guided filtering and fault-focused imaging techniques.

6. The method for identifying three-dimensional seismic faults in coalfields as described in claim 5, characterized in that, The dip-guided filtering extracts the local dip and azimuth of the seismic phase axis to construct a spatially constrained guide body, thereby adaptively processing and filtering the original 3D seismic data along the stratigraphic trend.

7. The method for identifying three-dimensional seismic faults in coalfields as described in claim 5, characterized in that, The fault-focusing imaging process performs fault-focusing imaging on the faults of the original three-dimensional seismic data by setting the main frequency to 40Hz, the low frequency to 15Hz, and the high frequency to 60Hz.

8. A three-dimensional seismic fault identification device for coalfields, characterized in that, include: The acquisition module is used to acquire the original 3D seismic data of the area to be measured; a deep learning model is used to denoise the original 3D seismic data to determine the denoised 3D seismic data. A building module is used to construct a ResU-Net model for coalfield seismic fault identification; wherein, the ResU-Net model replaces the convolutional blocks in the original U-Net model with residual modules; The fault probability volume generation module is used to train the ResU-Net model using the FaultSeg3D synthetic dataset with known fault features generated by simulation and the historical 3D seismic data of the area to be measured. The trained ResU-Net model is then used to identify faults in the denoised 3D seismic data and generate the fault probability volume. The optimization module is used to generate positive samples by manually marking faults in 3D seismic data that are not recognized by the ResU-Net model; and to generate negative samples by manually marking non-fault areas in 3D seismic data that are misjudged by the ResU-Net model. The positive and negative samples are used as a training set to retrain the trained ResU-Net model to iteratively optimize the fault probability volume and determine the optimized fault probability volume. The fault identification module is used to identify faults based on the optimized fault probability volume and in combination with geological interpretation requirements, and to obtain the identification results.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the coalfield three-dimensional seismic fault identification method according to any one of claims 1-7.

10. A computer device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the coalfield three-dimensional seismic fault identification method according to any one of claims 1-7.

Images