Deep Learning Based Seismic Data Reconstruction Method Considering Data Bias for Improved Generalization Performance

Abstract

The objective of the present invention is to provide a deep learning-based seismic data recovery method that considers data bias for improved generalization performance, thereby solving the problem of limited generalization performance of existing technologies, by implementing an integrated data recovery system capable of restoring various forms of damage, such as regular, irregular, and continuous defects in seismic exploration data, with a single deep learning model. A deep learning-based seismic data recovery method considering data bias for improved generalization performance according to the present invention for solving the above-mentioned problem comprises: (a) a step of configuring training data that reflects a plurality of defect types and noise conditions by considering the bias of acquired field seismic data; (b) a step of training a deep learning model using the training data so that a plurality of defect types can be recovered as a single model; and (c) a step of recovering a defect area of ​​actual seismic data using the trained deep learning model.

Description

Technology Field

[0001] The present invention relates to a method for recovering seismic data, and more specifically, to a method for augmenting and generating training data that considers dataset bias in order to improve the generalization performance of deep learning-based interpolation technology. Background Technology

[0003] Generally, when acquiring field data for seismic exploration, problems of data quality degradation, such as data missing, occur in the acquired seismic data due to various causes, including topographical constraints of the exploration area, limitations in exploration costs, and limitations in the structural geometry of the exploration methodology.

[0004] Figure 1 is a graph showing typical data loss (damage) that occurs during conventional seismic data acquisition. Figure 1 (a) shows regular data loss, (b) shows irregular data loss, and (c) shows continuous data loss.

[0005] As shown in Figure 1, there are three types of typical examples of such data loss (damage): regular data loss (damage) (Figure 1 (a)), irregular data loss (damage) (Figure 1 (b)), and continuous data loss (damage) (Figure 1 (c)).

[0006] Conventional technologies to solve these problems include trace interpolation techniques, and recently, deep learning-based trace interpolation techniques are also being developed.

[0007] In the general field of deep learning-based image processing, there is a wealth of diverse training data available that can be used to train machine learning models, and such trained models demonstrate generalized performance.

[0008] Furthermore, with the introduction of image super-resolution technology, the technology for restoring low-resolution images to high resolution has advanced dramatically. However, in the field of seismic exploration, data acquisition is difficult and costly, so the data required for training models for data restoration is very scarce and limited.

[0009] General machine learning-based interpolation and restoration techniques, which incorporate advanced technologies from the field of image processing, first construct training data by randomly extracting patches from available seismic data sets to recover regularly or irregularly lost traces.

[0010] After that, a deep learning model is trained to recover seismic data based on image superresolution techniques, and the trained model is applied regularly to the lost seismic data to recover the data.

[0011] However, generating generalized models in the aforementioned field of image processing is considerably difficult with seismic field data, where data limitations are clear. Therefore, most existing deep learning-based seismic data interpolation techniques have been developed by focusing on one of the three representative types of data omissions mentioned above.

[0012] Figure 2 is an example drawing of an actual field seismic survey data line with conventional complex data omissions.

[0013] As shown in Figure 2, conventional technologies demonstrate reliable recovery performance for focused types of data damage, but they do not guarantee data recovery performance for complex data damage occurring in actual exploration sites as shown in Figure 2, that is, generalized data recovery performance.

[0014] For example, if trace interpolation is performed on data containing irregular or continuous omissions using a deep learning model trained on training data to recover regular omissions, the performance drops sharply.

[0015] Therefore, in order to apply deep learning-based seismic interpolation technology to actual field conditions such as those shown in Figure 2, it is necessary to configure training data that reflects complex types of slippage and to have a deep learning model and training method suitable for it. Prior art literature

[0017] Korean Patent Publication No. 10-2514531 (Publication Date: March 27, 2023) Korean Patent Publication No. 10-2849303 (Publication Date: August 25, 2025) The problem to be solved

[0018] The objective of the present invention to solve the aforementioned problem is to provide a deep learning-based seismic data recovery method that considers data bias for improved generalization performance, thereby solving the problem of limited generalization performance of existing technologies, by implementing an integrated data recovery system capable of restoring various forms of damage, such as regular, irregular, and continuous defects in seismic exploration data, with a single deep learning model.

[0019] In addition, the objective of the present invention is to provide a method for augmenting and generating training data considering dataset bias to improve the generalization performance of deep learning-based interpolation technology, a deep learning model for training the same, and a new interpolation technology that recovers all missing data with a single trained deep learning model. means of solving the problem

[0021] A deep learning-based seismic data recovery method considering data bias for improved generalization performance according to the present invention for solving the above-mentioned problem comprises: (a) a step of configuring training data that reflects a plurality of defect types and noise conditions by considering the bias of acquired field seismic data; (b) a step of training a deep learning model using the training data so that a plurality of defect types can be recovered as a single model; and (c) a step of recovering a defect area of ​​actual seismic data using the trained deep learning model.

[0022] In addition, in a deep learning-based seismic data recovery method considering data bias for improved generalization performance according to the present invention, the training data configuration step of step (a) is characterized by simulating a plurality of damage types including regular defects, irregular defects, and continuous defects, and integrating them into a single training dataset.

[0023] In addition, in a deep learning-based seismic data recovery method considering data bias for improved generalization performance according to the present invention, the step (a) is characterized by including a data augmentation step including flipping, stitching, or random noise addition to mitigate selection bias, capture bias, and negative set bias of the data.

[0024] In addition, in a deep learning-based seismic data recovery method considering data bias for improved generalization performance according to the present invention, the step (a) is characterized by including complete seismic data without defects as part of the training input data to minimize distortion of amplitude information and to learn amplitude preservation characteristics.

[0025] In addition, in the deep learning-based seismic data recovery method considering data bias for improved generalization performance according to the present invention, the deep learning model of step (b) is based on a U-Net3+ structure including multi-scale feature extraction and skip connection between encoder and decoder, and is configured to perform deep supervision.

[0026] In addition, in a deep learning-based seismic data recovery method considering data bias for improved generalization performance according to the present invention, the step (b) is characterized by including the step of repeatedly inputting training data while changing the intensity of Gaussian random noise to reflect a plurality of signal noise environments in the field.

[0027] In addition, in the deep learning-based seismic data recovery method considering data bias for improved generalization performance according to the present invention, the recovery process of step (c) is characterized by applying a hybrid loss function designed to minimize the residual between the output and input of a learned deep learning model to improve the recovery accuracy.

[0028] In addition, the deep learning-based seismic data recovery method considering data bias for improved generalization performance according to the present invention is characterized by further including a performance verification step that evaluates the quality of the recovered seismic data to verify the generalization performance of the model.

[0029] In addition, in the deep learning-based seismic data recovery method considering data bias for improved generalization performance according to the present invention, the evaluation process of the performance verification step is characterized by calculating the signal-to-noise ratio (SNR) and structural similarity (SSIM) between the recovered seismic data and the reference data to quantitatively verify the generalization performance.

[0031] In addition, the computer program stored on a storage medium is characterized by executing a deep learning-based seismic data recovery method that considers data bias for the improved generalization performance described above on a computer. Effects of the invention

[0033] According to the present invention, data can be effectively recovered for complex data damage that matches the conditions of actual seismic field data, and economic efficiency in exploration can be secured because there is no need to perform "infill" to re-examine missing data.

[0034] In addition, according to the present invention, the quality of data processing can be improved by effectively recovering damaged data.

[0035] In addition, according to the present invention, unlike conventional methods, only a single deep learning model is trained and can be applied to various data damage cases, making computational costs efficient and allowing for flexible application in actual field situations.

[0036] In addition, according to the present invention, the problem of poor generalization performance in existing deep learning-based seismic interpolation technology, which targets only one type of defect, is resolved, thereby providing a technical basis for integrating and training various types of damage data, such as regular, irregular, and continuous defects, into a single training dataset.

[0037] This enables stable and consistent restoration results even in complex defect environments similar to actual exploration sites.

[0038] In addition, according to the present invention, by systematically mitigating major bias factors such as selection bias, collection bias, and negative bias during the dataset construction process, the data imbalance problem and overfitting (especially creeping overfitting) of existing models can be minimized.

[0039] Furthermore, according to the present invention, by effectively simulating various signal-to-noise ratio (SNR) environments that may occur during actual seismic measurement processes through the addition of Gaussian random noise, the model is trained to exhibit stable reconstruction performance even under realistic exploration noise conditions. This significantly enhances the robustness of the model in withstanding the uncertainty of field data.

[0040] In addition, according to the present invention, by introducing data augmentation based on flipping and stitching, local biases such as data directionality or near-distance gap loss can be mitigated, and richer learning patterns can be obtained from the same data.

[0041] In addition, according to the present invention, by including a complete data set (negative set) in the learning process to enhance amplitude preservation performance, the restored elastic wave data maintains the amplitude characteristics of the original signal without distortion.

[0042] This provides highly reliable data for subsequent geological interpretation stages, such as AVO (Amplitude Variation with Offset) analysis.

[0043] In addition, according to the present invention, various missing patterns can be restored with a single U-Net3+ based deep learning model without requiring multiple models, thereby drastically reducing computational resources and training time compared to existing methods while maintaining high restoration quality.

[0044] In addition, the deep supervision structure reduces information loss between multi-resolution features, enhancing the detail of the restoration results. Brief explanation of the drawing

[0046] Figure 1 is a graph showing typical types of data loss (damage) that occur during conventional seismic data acquisition. Figure 2 is an example drawing of an actual field seismic survey data line with conventional complex data omissions. FIG. 3 is a diagram showing the detailed flow of a deep learning-based seismic data recovery method considering data bias for improved generalization performance according to an embodiment of the present invention. Figure 4 is a simplified schematic diagram of an algorithm to which a deep learning-based seismic data recovery method considering data bias for improved generalization performance according to an embodiment of the present invention is applied. FIG. 5 is a schematic diagram of a dataset bias mitigation strategy for deep learning (DL)-based earthquake data interpolation applied to a deep learning-based seismic data recovery method considering data bias for improved generalization performance according to an embodiment of the present invention. FIG. 6 is a schematic diagram of a modified U-Net3+ structure applied to a deep learning-based seismic data recovery method considering data bias for improved generalization performance according to an embodiment of the present invention. Figure 7 is a structure-corrected seismic cross-section of the Viking Graben dataset. The red arrow and the green arrow represent the target adjacent data and target distant data, respectively, from the training dataset. Figure 8 is a graph showing the interpolation results of adjacent target data. Figure 9 is a graph showing a comparison of interpolation results for target distant data. Top: Conventional technology, Bottom: Developed technology Figure 10 is a graph showing the results of comparing enlarged portions of the images in Figure 8. From left to right: original, prior art, and developed technology. Fig. 11 is a comparison graph of interpolation results for target distant data. Top: Conventional technology, Bottom: Developed technology Figure 12 is a graph showing the results of comparing enlarged portions of the images in Figure 10. From left to right: original, prior art, and developed technology. Fig. 13 is a graph comparing the interpolation results of target distant data. Top: Conventional technology, Bottom: Developed technology Figure 14 is a graph showing a detailed comparison of the interpolation results of the 20 consecutive defect traces presented in Figure 12 using a wiggle trace plot. Left: Conventional technology, Right: Developed technology Fig. 15 is a comparison graph of the interpolation results for target distant data. Top: Conventional technology, Bottom: Developed technology Figure 16 is a graph showing a detailed comparison of the interpolation results of the 30 consecutive defect traces presented in Figure 14 using a wiggle trace plot. Left: Conventional technology, Right: Developed technology Specific details for implementing the invention

[0047] Before describing the present invention in detail, it should be understood that the terms and words used in this specification should not be interpreted as being limited to their ordinary or dictionary meanings, and that the inventor of the present invention may appropriately define and use the concepts of various terms to best describe their invention, and furthermore, that these terms and words should be interpreted in a meaning and concept consistent with the technical spirit of the present invention.

[0048] In other words, it should be understood that the terms used in this specification are used merely to describe preferred embodiments of the present invention and are not intended to specifically limit the content of the present invention, and that these terms are defined in consideration of various possibilities of the present invention.

[0049] In addition, it should be noted that in this specification, singular expressions may include plural expressions unless the context clearly indicates a different meaning, and that even if they are expressed in a similarly plural form, they may include a singular meaning.

[0050] Throughout this specification, where it is stated that a component "includes" another component, unless specifically stated otherwise, this may mean that it does not exclude any other component but may include any other component.

[0051] Furthermore, it should be noted that in cases where it is stated that a component "exists inside or is installed in connection with" another component, this component may be installed in direct connection or contact with the other component, or it may be installed at a certain distance apart, and in the case where it is installed at a certain distance apart, there may be a third component or means for fixing or connecting the component to the other component, and a description of this third component or means may be omitted.

[0052] On the other hand, if it is stated that one component is "directly connected" or "directly connected" to another component, it should be understood that there is no third component or means.

[0053] Likewise, other expressions describing the relationship between each component, such as “between” and “right between”, or “adjacent to” and “directly adjacent to”, should be interpreted as having the same intent.

[0054] In addition, it should be understood that in this specification, terms such as “one side,” “other side,” “one side,” “other side,” “first,” “second,” etc., are used to clearly distinguish one component from another component, and that the meaning of the component is not restricted by such terms.

[0055] In addition, position-related terms such as "up," "down," "left," and "right" used in this specification should be understood as indicating the relative position of the corresponding component in the drawing, and unless an absolute position is specified, these position-related terms should not be understood as referring to an absolute position.

[0056] Furthermore, in specifying the reference numerals for each component of each drawing in this specification, the same component has the same reference numeral even if it is shown in different drawings; that is, the same reference numeral throughout the specification indicates the same component.

[0057] In the drawings attached to this specification, the size, position, connection relationships, etc., of each component constituting the present invention may be described in a partially exaggerated, reduced, or omitted manner for the convenience of explanation or to sufficiently clearly convey the concept of the present invention, and therefore, the proportions or scale may not be strictly accurate.

[0058] In addition, in describing the present invention below, detailed descriptions of components, such as prior art and known technology, that are deemed to unnecessarily obscure the essence of the invention may be omitted.

[0060] Preferred embodiments of the present invention will be described in detail below with reference to the drawings.

[0061] FIG. 3 is a diagram showing the detailed flow of a deep learning-based seismic data recovery method considering data bias for improved generalization performance according to an embodiment of the present invention, and FIG. 4 is a simple schematic diagram of an algorithm to which the deep learning-based seismic data recovery method considering data bias for improved generalization performance according to an embodiment of the present invention is applied.

[0062] As shown in FIGS. 3 and 4, a deep learning-based seismic data recovery method considering data bias for improved generalization performance according to an embodiment of the present invention may be configured to include (a) a training data configuration step (S100), (b) a deep learning model training step (S200), (c) a seismic data recovery step (S300), and (d) a model generalization performance verification step (S400) after evaluating the quality of the recovered data.

[0063] The deep learning-based seismic data recovery method considering data bias for improved generalization performance according to the embodiment of the present invention is intended to solve the problem in which generalization performance is degraded when existing deep learning-based recovery techniques are overfitted to specific types of defects or specific datasets when recovering irregular and continuous damage regions present in seismic data acquired from field seismic exploration.

[0064] In particular, the embodiment of the present invention takes into account that dataset bias degrades generalization performance during the deep learning process, and (1) constructs a bias-reducing dataset that reflects various types of defects and noise conditions in the training data construction stage, (2) uses this to train a single deep learning model for multiple types of damage, and (3) provides a general-purpose and efficient deep learning-based seismic wave recovery method that can recover various damage patterns of actual field data with just one training.

[0065] This allows users to respond to various types of damage with a single model without the need to design separate models or perform iterative training for each type of defect, thereby reducing re-exploration costs at the site and maximizing quality improvement and processing efficiency.

[0066] More specifically, (a) step (S100) may be a step of constructing training data that reflects multiple defect types (regular, irregular, continuous) and various noise conditions by taking into account the bias of the acquired field seismic data.

[0067] Step (a) (S100) can perform a data composition strategy to mitigate selection bias, capture bias, negative set bias, etc., so that the deep learning model does not overfit to various damage patterns that may occur in the actual field.

[0068] Here, selection bias mitigation reduces data bias by generating training data that simulates various loss ratios and loss types (e.g., continuous or irregular loss), allowing a single model to accumulate learning experience for multiple types of impairment.

[0069] In addition, capture bias mitigation removes directional bias in the data by applying data augmentation techniques such as flipping and stitching.

[0070] In addition, negative set bias mitigation involves training with a portion of intact elastic data to allow the model to preserve the energy of the original signal without amplitude distortion.

[0071] In addition, noise diversification reflects various noise conditions in the actual field by injecting Gaussian random noise at varying intensities.

[0072] Thus, through step (a) (S100), the constructed data can provide a generalizable learning base by integrating various damage and noise conditions, unlike existing datasets dedicated to specific patterns.

[0073] That is, in an embodiment of the present invention, (a) step (S100) includes a data augmentation strategy to mitigate selection bias, collection bias, and voice set bias occurring during the process of constructing training data.

[0074] In general, seismic data acquired during field surveys is obtained under limited spatial and temporal conditions, resulting in capture bias where the directionality of the data or the sampling interval is skewed.

[0075] In addition, when patches are extracted and trained only in certain sections, selection bias is formed, leading to a problem where the model overfits to specific structures or geological conditions.

[0076] To address this, the present invention eliminates directional bias in data through flipping and compensates for missing patterns between adjacent sections through stitching.

[0077] In addition, by adding Gaussian random noise at various intensities, data with different noise levels is incorporated into the training, enabling the model to robustly respond to changes in the signal-to-noise ratio that may occur in actual exploration environments.

[0078] These data augmentation procedures are a key technical means of improving the generalization performance of deep learning models by artificially expanding the diversity of training data.

[0079] And, (a) step (S100) may include a step of including complete elastic wave data without defects as part of the training input data to minimize distortion of amplitude information and to learn amplitude preservation characteristics.

[0080] As such, the deep learning-based seismic data recovery method considering data bias for improved generalization performance according to the embodiment of the present invention can be configured to include undamaged complete seismic data as part of the training input data through step (a) (S100) to minimize distortion of amplitude information and strengthen amplitude preservation characteristics.

[0081] Conventional deep learning-based restoration models were trained using only damaged data as input, which resulted in a problem where the amplitude of the restored trace was distorted and differed from the actual signal. In particular, in amplitude-based analysis (e.g., AVO, Amplitude Variation with Offset), this amplitude loss causes serious interpretation errors.

[0082] To solve this problem, in an embodiment of the present invention, a complete dataset (negative set) without defects is included as a partial training input, so that the model learns both the complete data and the damaged data simultaneously.

[0083] This applies the learning principles of an autoencoder, allowing the model to implicitly learn the ability to reconstruct input data into its original signal.

[0084] As a result, the trained model maintains the original amplitude characteristics during the reconstruction process, which can directly contribute to the preservation of the physical meaning of seismic waves and the improvement of subsequent interpretation accuracy.

[0085] More specifically, the training data configuration step of (a) step (S100) may be configured by simulating multiple types of damage, including regular loss, irregular loss, and continuous loss, and integrating them into a single training dataset.

[0086] As such, the deep learning-based seismic data recovery method considering data bias for improved generalization performance according to the embodiment of the present invention is intended to fundamentally resolve the problem of degraded generalization performance of conventional models dedicated to a single defect pattern by configuring various defect types occurring in field seismic data into a single integrated training dataset.

[0087] In conventional technology, models were trained by simulating only one type of defect, either regular or irregular, so there was a problem where restoration performance rapidly deteriorated when complex defects, such as those in actual exploration environments, were present.

[0088] Accordingly, in an embodiment of the present invention, all three representative types of defects (regular, irregular, and continuous) are simulated and included in the training data, and the ratio of each defect, the defect interval, and the number of continuous defect traces are set as variables so that various defect patterns are integratedly learned within a single training set.

[0089] This integrated training data configuration mitigates selection bias and enables a single model to handle multiple types of damage, thereby allowing for stable recovery even from irregular loss of actual field data.

[0090] And, as shown in FIGS. 3 and 4, step (b) (S200) is a deep learning model training step, which may be a step of training a deep learning model so as to be able to recover multiple types of defects into a single model using the training data configured in step (a) (S100).

[0091] Step (b) (S200) such as this may be a step of creating a general-purpose model capable of recovering various types of defects with a single network by learning all damage cases in an integrated manner, instead of learning the model separately for each type of defect as in the conventional method.

[0092] In addition, in step (b) (S200), the structure of the deep learning model is based on the U-Net3+ structure, and local detailed structures and global patterns are restored simultaneously through skip connections between encoders and decoders and connections between multi-scale feature maps.

[0093] The learning strategy of this model is to use a bias-relaxing dataset to perform regularization so that the deep learning network does not overfit to a specific pattern, and to learn common features for multiple types of defects.

[0094] Through this, various damaged data can be recovered using only a single model, which can lead to improvements in data and computational efficiency.

[0095] More specifically, the deep learning model of step (b) (S200) is based on a U-Net3+ structure including multi-scale feature extraction and skip connections between encoder and decoder, and can be configured to perform deep supervision.

[0096] That is, in step (b) (S200), a network based on the U-Net3+ structure is used as the deep learning model applied, and may be configured to perform deep supervision, including a multi-scale feature extraction and skip connection structure between the encoder and the decoder.

[0097] U-Net3+ is an advanced form of the basic U-Net or U-Net++ structure, and may be a network structure that strengthens horizontal and vertical connections (inter-decoder and encoder-decoder interconnection) to minimize information loss between multi-resolution feature maps and allow each level of decoder to simultaneously refer to the output of other decoders as well as the encoder.

[0098] The U-Net3+ network aggregates detailed features (fine structure) extracted from the encoder and global semantics (geological structure) reconstructed from the decoder at multiple scales to simultaneously restore the fine shape and location information of missing regions.

[0099] In addition, by performing deep supervision at each level, the decoders of all layers receive the learning signal directly, thereby minimizing information loss between low-resolution and high-resolution features.

[0100] This structure can more accurately capture the fine boundary and amplitude characteristics required for recovering missing areas in earthquake data than the single-skip connection-based learning of the existing U-Net.

[0101] Consequently, the U-Net3+ based model can accurately reconstruct location and amplitude information even in sections with high loss rates, and can improve generalization performance on actual exploration site data due to its small number of parameters and efficient learning capabilities.

[0102] And, (b) step (S200) may include a step of repeatedly inputting training data while varying the intensity of Gaussian random noise to reflect a plurality of signal noise environments in the field.

[0103] Thus, step (b) (S200) may include a step of repeatedly inputting training data while varying the intensity of Gaussian random noise in order to reflect various signal-to-noise ratio (SNR) conditions of the field during the learning process.

[0104] This may be a procedure to induce feature-robust learning so that the deep learning model can withstand various noise conditions occurring in the actual exploration environment.

[0105] Gaussian noise simulates the stochastic signal distortion naturally included in field data due to the sensitivity of exploration equipment, differences in stratum reflectance, and external interference.

[0106] In this invention, such noise is injected at various intensities to train the model so that it does not respond unstably to changes in the noise level.

[0107] In other words, during the process of iteratively training with data to which noise has been added, the model recognizes the statistical distribution of the input signal in a more generalized form and does not overfit to a specific SNR environment.

[0108] Consequently, this serves as a technical basis for maintaining restoration accuracy even in sections with low signal quality in actual field data and enhancing the generalization performance of the model.

[0109] And, as shown in FIGS. 3 and 4, step (c) (S300) is a seismic data recovery step, which may be a step of recovering missing areas of seismic data obtained at an actual exploration site using a learned deep learning model.

[0110] In this way, in step (c) (S300), a trained single model analyzes the input damage data to interpolate and reconstruct the missing trace.

[0111] This ensures the continuity of actual field data and enables the repair of defects without the need for infill surveys.

[0112] More specifically, through the mapping of input and output in step (c) (S300), when damaged data is input, the model can extract multi-scale features through the encoder and restore the signal of the missing region in the decoder.

[0113] In addition, the difference between the restored signal and the original can be minimized by applying a residual-based loss function (hybrid loss) in step (c) (S300).

[0114] This enables stable signal recovery even in areas with high or irregular loss rates, and reduces re-exploration costs.

[0115] More specifically, the recovery process of step (c) (S300) may improve the recovery accuracy by applying a hybrid loss function designed to minimize the residual between the output and input of the learned deep learning model.

[0116] Here, the hybrid loss function is a loss calculation method constructed by combining the traditional L1 loss (mean absolute error), L2 loss (mean squared error), and the structural similarity metric (SSIM loss), which can complement the limitations of learning that relies on a single metric.

[0117] Existing L2 loss-based reconstruction models reduce overall pixel-level differences by minimizing mean squared error, but they have limitations in that they cannot accurately reflect the fine structure or subtle changes in amplitude of the reconstructed elastic wave signal.

[0118] On the other hand, SSIM loss is effective in maintaining visual and physical similarity because it considers local brightness, contrast, and structural information.

[0119] In an embodiment of the present invention, a hybrid loss function combining L1 / L2 loss and SSIM loss is used to minimize residuals while simultaneously preserving the morphological similarity of the geological structure.

[0120] This hybrid loss is combined with the multi-scale learning architecture of U-Net3+ to restore detailed signal patterns as well as maintain overall reflection surface continuity, and as a result, both phase consistency and amplitude fidelity of the restored seismic data can be secured.

[0121] And, as shown in FIG. 3, a deep learning-based seismic data recovery method considering data bias for improved generalization performance according to another embodiment of the present invention may further include (d) a performance verification step that evaluates the quality of the recovered seismic data to verify the generalization performance of the model.

[0122] This is characterized by including an additional step for determining the actual performance of the model by evaluating the restoration results on field unseen data that was not included in the training, rather than simply the reproduction performance on the training data.

[0123] The generalization performance of a model is judged by the accuracy and consistency of the reconstruction results in test data with geological structures or exploration conditions different from the training data.

[0124] In the embodiments of the present invention, not only the signal quality of the restored data but also the continuity of the stratigraphic boundaries (layered structure) and the degree of preservation of the shape of the reflective surface are comprehensively verified.

[0125] This step serves as a post-processing stage to determine whether the model is overfitting and to empirically verify the effectiveness of data bias mitigation strategies (flipping, stitching, noise variation, etc.).

[0126] In addition, the evaluation results can be used as feedback for model retraining or hyperparameter optimization, ultimately completing a cyclic structure of bias minimization, generalization enhancement, and restoration precision improvement.

[0127] More specifically, in the evaluation process of the performance verification stage, it is desirable to quantitatively verify generalization performance by calculating the signal-to-noise ratio (SNR) and structural similarity (SSIM) between the reconstructed seismic data and the reference data.

[0128] SNR indicates the physical restoration quality of data by measuring the ratio of noise components to the energy of the restored signal, and SSIM can evaluate the similarity of local patterns, contrast, and brightness between the original and the restored signal.

[0129] In other words, SNR calculation quantitatively represents the noise suppression performance of reconstructed data, and SSIM can evaluate the continuity of stratigraphic structures and the morphological reconstruction quality that are difficult to distinguish with existing RMS error-based indicators.

[0130] In the embodiments of the present invention, by applying two indicators in parallel, an integrated performance evaluation that considers not only numerical error but also structural similarity is possible.

[0131] In particular, evaluations including SSIM are reconstruction assessments that reflect geological significance rather than simple signal matching, confirming that the reconstructed seismic data maintains visual consistency with the physical characteristics of the actual reflective surface.

[0132] This evaluation procedure objectively verifies the model's generalization performance and enables the quantification of the impact of data bias mitigation strategies presented during the training phase on actual restoration quality.

[0134] In addition, another embodiment of the present invention may feature a computer program stored in a storage medium to execute a deep learning-based seismic data recovery method that takes into account data bias for enhanced generalization performance on a computer.

[0135] In addition, a program applied to a deep learning-based seismic data recovery method considering data bias for enhanced generalization performance according to one embodiment of the present invention may be implemented as computer-readable code on a computer-readable recording medium. The codes and code segments implementing the above program can be easily inferred by a computer programmer in the field.

[0136] Here, a computer-readable recording medium may include any type of recording device in which data that can be read by a computer system is stored. Examples of computer-readable recording media may include ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical disk, etc. Additionally, computer-readable recording media may be distributed across networked computer systems and may be written and executed as computer-readable code in a distributed manner.

[0138] Hereinafter, the results of a study (hereinafter referred to as "this study") on various missing data scenario experiments targeting actual field data by applying a deep learning-based seismic data recovery method considering data bias for improved generalization performance according to an embodiment of the present invention will be described in detail with reference to the drawings.

[0139] summation( Abstract )

[0140] Since spatial undersampling of earthquake data causes significant problems during the signal processing stage, earthquake data interpolation techniques are crucial in the preprocessing process.

[0141] Recently, several deep learning-based interpolation techniques have been proposed to address various missing data scenarios, such as regular, irregular, or large gaps.

[0142] However, this standardized approach can cause the 'creeping overfitting' problem in various types of missing data, which severely degrades the generalization performance of trained deep learning models.

[0143] This study presents a new approach to address this generalization problem.

[0144] We emphasize that data bias in the training dataset degrades interpolation performance depending on the various characteristics of the target data, and propose guidelines to minimize such bias during the training data construction phase.

[0145] In addition, we propose a single deep learning model that utilizes U-Net3+ as a backbone and is applicable to all various missing data situations in actual field data.

[0146] As a result of experiments with various missing data scenarios using actual field data, this method demonstrated excellent interpolation performance even for target unseen data by using an unbiased dataset.

[0147] Ultimately, this study emphasizes the importance of systematically designed training datasets in deep learning-based interpolation and suggests the need for future research to develop a universally applicable deep learning-based earthquake data interpolation network.

[0149] METHODOLOGY

[0150] training dataset Bias (Training dataset Strategies for mitigating bias

[0151] Existing DL (Deep Learning)-based earthquake data interpolation methods contain the aforementioned various biases.

[0152] In this study, we proposed a training data construction strategy that minimizes such bias and experimentally verified its effectiveness.

[0153] Existing methods selected specific missing patterns (regular, irregular, consecutive) and constructed training data using only those patterns, but this leads to overfitting for specific types.

[0154] In this study, we constructed an integrated training dataset that includes three representative missing patterns and each variable (e.g., irregular missing rate, number of consecutive missing traces, etc.). Through this, we designed a single DL model capable of handling various missing situations.

[0155] FIG. 5 is a schematic diagram of a dataset bias mitigation strategy for deep learning (DL)-based earthquake data interpolation applied to a deep learning-based seismic data recovery method considering data bias for improved generalization performance according to an embodiment of the present invention.

[0156] Data bias refers to four types of bias that must be considered in training data construction strategies in the fields of machine learning and deep learning. It serves as a guideline for structuring training data in a way that mitigates each of these biases. Therefore, data augmentation methods are selected based on four criteria to suit the conditions of the field data.

[0157] As shown in Figure 5, a strategy was used to mitigate the remaining data biases, excluding label bias (label bias; which is not suitable for field conditions as it requires additional acquisition of correct answer data).

[0158] That is, as shown in Figure 5, to ensure diversity in the training dataset, selection bias was mitigated by including data with different noise levels and combining various defect patterns. Specifically, Gaussian random noise was added at different intensities to reflect various noise conditions in the actual field.

[0159] In addition, two simple yet effective augmentation techniques were used to reduce capture bias: (1) Flipping and (2) Data stitching.

[0160] Flipping allows for the reflection of various gradient directions in limited data, while stitching diversifies the shape of missing intervals by compensating for missing patterns in near-offset regions.

[0161] This process prevented learning that was biased in a specific direction.

[0162] Finally, a negative set was included in the training data. By adding complete sampled data as input without any loss, the DL model was enabled to learn amplitude recovery.

[0163] Various levels of Gaussian noise were added to this data as well to ensure learning diversity. The unbiased learning dataset constructed in this way (d unbias ) is designed to mitigate selection bias, gain bias, and negative set bias.

[0164] However, label bias is difficult to completely eliminate unless the ground truth of the actual field target is known.

[0165] Ultimately, this study presents a methodology for systematically constructing a structurally unbiased dataset, rather than "generating a large amount of data unconditionally."

[0166] This approach provides a practical foundation for improving generalization performance.

[0168] Modified U- Net3 + Network(Modified U- Net3 + network)

[0169] A representative architecture widely used in deep learning interpolation is U-Net.

[0170] U-Net is based on an encoder-decoder structure and connects the low-level detailed features of the encoder with the high-level semantic features of the decoder using skip connections.

[0171] Here, a skip connection is a structure that directly connects low-level feature maps extracted in the encoder stage with high-level semantic feature maps in the decoder stage.

[0172] In other words, it is a connection line that “skips” the weighting of information to allow detailed spatial information of the input image (e.g., edges, boundaries, detailed patterns, etc.) to be reused during the decoding process.

[0173] Thanks to this, it is very effective for restoring damaged data and has been actively used in various studies.

[0174] However, it is difficult to completely reduce information loss with simple skip connections alone.

[0175] To address this limitation, the U-Net++ architecture was proposed, which reduced the semantic gap between the encoder and decoder by adding dense skip connections.

[0176] However, information transfer between multiple feature scales is still not sufficiently efficient.

[0177] Accordingly, this study introduced an improved version based on the U-Net3+ structure.

[0178] FIG. 6 is a schematic diagram of a modified U-Net3+ structure applied to a deep learning-based seismic data recovery method considering data bias for improved generalization performance according to an embodiment of the present invention.

[0179] As shown in Figure 6, the modified U-Net3+ adds inter-decoder connections as well as connections between encoders and decoders, thereby capturing fine-grained detail and coarse-grained semantics simultaneously.

[0180] In addition, through the hybrid loss function and deep supervision

[0181] The learning of hierarchical representations at multiple scales was enhanced.

[0182] As a result, the modified U-Net3+ enables more accurate location and amplitude reconstruction in situations with high defects compared to the existing U-Net, and has fewer parameters and clearer boundary representation.

[0184] FIELD DATA APPLICATIONS

[0185] Experimental data composition

[0186] The Viking Graben Line 12 dataset was used for the experiment. This data consists of 1,001 blast shot records, each shot recorded at 120 channels, 6 seconds in length, and a 4 ms sampling interval.

[0187] The blast interval and receiver interval are each 25m. All traces were completely collected, so the data is completely free of any defects.

[0188] For the experiment, the data was divided into three sections.

[0189] Figure 7 is a structure-corrected seismic cross-section of the Viking Graben dataset. The red arrow and the green arrow represent the target adjacent data and target distant data, respectively, from the training dataset.

[0190] As shown in Figure 7, the middle section (blue box) was set as the complete sampling section (training data), and the two sections (green box) were set as the verification / test data with missing patterns.

[0191] In other words, the learning model is trained only on the middle interval data, and the intervals on both sides are evaluated as “unseen targets.”

[0192] In particular, two representative target regions were selected based on their distance from the training data.

[0193] The red arrows represent adjacent data close to the training interval, and the green arrows represent distant data with different geological characteristics.

[0195] Experimental scenario

[0196] In this study, two experiments were conducted to verify the effectiveness of the proposed unbiased learning dataset.

[0197] Scenario 1: Adjacent intervals similar to the training data (red arrow, see Fig. 7)

[0198] 1) Objective: Comparison of the stability of the proposed U-Net3+ and its performance compared to existing models

[0199] 2) Expectation: Since the data characteristics are similar, both models show good interpolation performance.

[0200] Scenario 2: Distant section different from the training data (green arrow)

[0201] 1) Objective: Verification of generalization performance

[0202] 2) Expectation: Existing method experiences a sharp decline in performance, while the proposed method maintains performance.

[0203] All experiments were conducted with the same network (U-Net3+) structure, the same hyperparameters (ADAM optimizer, hybrid loss), and the same number of patches (approx. 120,000).

[0204] The only difference is the method of configuring the training dataset. That is, the existing method uses individual models for each deficiency type (a total of 8), while the proposed method uses only a single model.

[0206] Learning stages and comparison

[0207] In the existing method, a separate model must be trained for each type of deficit. For example,

[0208] Regular missing 50%, 67%, irregular missing 30%, 50%, 70%, continuous missing 10, 20, 30 traces

[0209] A total of 8 models must be trained individually.

[0210] In contrast, the proposed method used a single integrated training dataset to train all these deficit types into a single U-Net3+ model.

[0211] In this process, selection bias was reduced by adding various noise levels, and acquisition and speech biases were mitigated through data augmentation and the addition of speech sets.

[0213] Comparison of results

[0214] Figure 8 is a graph showing the interpolation results of adjacent target data.

[0215] Figures 8(a) to 8(c) show the experimental results for 50% regular defects, 50% irregular defects, and 20 consecutive defect traces, respectively.

[0216] From left to right, the original CSG (label), the decimated CSG, the interpolation result of the previous framework, and the interpolation result of the method proposed in this study are shown.

[0217] As shown in Figure 8, looking at the results of adjacent intervals similar to the training data, both the existing method and the proposed method showed excellent interpolation performance for three representative defect types (ruled 50%, irregular 50%, continuous 20 traces).

[0218] This is because the target data has similar geological characteristics to the training data, demonstrating that in this case, both models can achieve sufficient performance even by 'memorizing' the input-output relationship.

[0219] Figure 9 is a graph showing a comparison of interpolation results for target distant data.

[0220] Figures 9 (a) and (b) show the results for the 50% irregular defect case of the previous framework and the proposed model, respectively. From left to right, they show the original CSG (labeled), the decimated CSG, the interpolation result, and the difference between the original common blast record (CSG) and the interpolation result.

[0221] Figure 10 is a graph showing the result of comparing enlarged portions of the images in Figure 8.

[0222] Figure 10 (a) shows the enlarged area of ​​the original CSG, (b) shows the enlarged area of ​​the previous framework, and (c) shows the enlarged area of ​​the proposed method.

[0223] As shown in Figures 9 and 10, the results differ for the distant section (green arrow). In the case of 50% irregular loss, the performance of the existing method deteriorates rapidly when the characteristics differ from the training data, whereas the proposed method maintained stable interpolation performance without additional training.

[0224] Figure 11 is a comparison graph of the interpolation results for target distant data.

[0225] Figure 11 (a) shows the results of the previous framework and (b) shows the results of the proposed model, respectively, and a comparison was performed for 70% irregular defect cases.

[0226] As shown in Figure 11, each column represents, from left to right, the original CSG (label), the decimated CSG, the interpolation result, and the difference between the original CSG and the interpolation result.

[0227] Figure 12 shows the result of comparing an enlarged portion of the image in Figure 10.

[0228] Figures 12 (a) to (c) respectively show the enlarged area of ​​the original CSG, the enlarged area of ​​the interpolation result of the previous framework, and the enlarged area of ​​the interpolation result of the proposed method.

[0229] Figure 13 is a graph comparing the interpolation results of target distant data.

[0230] Figures 13 (a) and (b) show the results for 20 consecutive missing trace cases of the previous framework and the proposed model, respectively.

[0231] Each column, from left to right, represents the original CSG (label), the attenuated CSG, the interpolation result, and the difference between the original CSG and the interpolation result.

[0232] Figure 14 is a graph showing the results of a detailed comparison of the interpolation results of 20 consecutive missing traces presented in Figure 12 using a wiggle trace plot.

[0233] The interpolated trace was displayed superimposed on the original trace (label).

[0234] Figure 14 (a) is a comparison graph between the interpolated trace of the previous framework and the original trace, and Figure 14 (b) shows the comparison between the interpolated trace of the proposed method and the original trace, respectively. The black line represents the original trace, and the red line represents the interpolated trace.

[0235] Figure 15 is a comparison graph of the interpolation results of target distant data.

[0236] Figures 15 (a) and (b) show the results for 30 consecutive missing trace cases of the previous framework and the proposed model, respectively.

[0237] Each column shows, from left to right, the original CSG (label), the attenuated CSG, the interpolation result, and the difference between the original CSG and the interpolation result.

[0238] Figure 16 is a graph showing the results of a detailed comparison of the interpolation results of 30 consecutive missing traces presented in Figure 14 using a wiggle trace plot.

[0239] The interpolated trace is superimposed on the original trace, and Figure 16 (a) shows a comparison graph between the interpolation result of the previous framework and the original trace, and Figure 16 (b) shows a comparison between the interpolation result of the proposed method and the original trace, respectively. The black line represents the original trace, and the red line represents the interpolated trace.

[0240] As such, when the defect was deepened to 70% (Figs. 11 and 12) or when the continuous defect reached 20 to 30 traces (Figs. 13 to 16), the proposed method was significantly superior in amplitude recovery capability.

[0241] While existing methods suffered from significant amplitude loss and increased noise during interpolation for new data, the proposed method produced amplitudes that closely matched the original with minimal structural distortion.

[0243] Quantitative evaluation

[0244] Quantitative evaluation was performed using SNR (Signal-to-Noise Ratio) and SSIM (Structural Similarity Index). The results are as follows:

[0245] 1) Adjacent intervals (Table 1): Both methods show similar high performance (SSIM 0.98~0.99)

[0246] 2) Long-distance section (Table 2): Existing method SSIM average 0.6–0.8 units, maintained at an average of 0.95 or higher for the proposed method SSIM.

[0247] This means that the proposed U-Net3+ performs stable learning without being biased toward specific data.

[0249] Discussion

[0250] While existing frameworks are optimized to solve a single specific defect problem, the deep learning-based seismic data recovery method considering data bias for enhanced generalization performance applied in this study is designed to generalize various defect patterns into a single DL model.

[0251] Experimental results showed that while both models demonstrated good performance on adjacent data, there was a significant difference in generalization performance on distant data with different geological structures.

[0252] In other words, the proposed data organization strategy demonstrated that it can stably interpolate even in new situations by learning the fundamental distribution of the data, rather than simply memorizing input-output relationships.

[0253] This study can be considered the first case to realize a universal deep learning-based trace interpolator that responds to various types of defects through a single learning process.

[0255] CONCLUSIONS

[0256] This study critically analyzed the limitations of existing deep learning-based earthquake data interpolation research and proposed a new approach aimed at improving generalization performance for unseen data.

[0257] This study systematically analyzed the dataset bias problem arising from existing research using only training data optimized for specific deficit types, and proposed training data composition strategies (selection, acquisition, and negative set bias mitigation) to mitigate it.

[0258] Field experiment results showed that while the existing method exhibited a sharp performance degradation in regions with different geological structures, the proposed method maintained high performance without additional training.

[0259] In other words, creeping overfitting was suppressed, and generalized interpolation performance was achieved. Although extensive verification across various fields (source waveform, frequency, noise characteristics, etc.) remains, the proposed framework demonstrates the potential to develop into a practical, general-purpose DL-based seismic data interpolation model.

[0261] Although various preferred embodiments of the present invention have been described above with some examples, the descriptions of various embodiments described in the "Specific details for carrying out the invention" section are merely illustrative, and those skilled in the art to which the present invention pertains will understand that the present invention can be modified in various ways or equivalent embodiments can be carried out based on the above description.

[0262] In addition, since the present invention can be implemented in various other forms, the present invention is not limited by the description above. The above description is provided merely to make the disclosure of the present invention complete and to fully inform those skilled in the art of the scope of the present invention, and it should be understood that the present invention is defined only by each claim of the claims.

Claims

Claim 1 (a) a step of configuring training data that reflects multiple types of defects and noise conditions, taking into account the bias of acquired field seismic data; (b) a step of training a deep learning model using the training data to enable recovery of multiple types of defects into a single model; and (c) a step of recovering the defect area of ​​actual seismic data using the trained deep learning model, wherein step (a) comprises integrating multiple types of damage, including regular defects, irregular defects, and continuous defects, into a single training dataset, performing data augmentation including flipping, stitching, and adding random noise to mitigate selection bias, capture bias, and negative set bias, and including complete seismic data without defects as part of the training data to learn amplitude preservation characteristics, thereby providing a deep learning-based seismic data recovery method that considers data bias for improved generalization performance. Claim 2 delete Claim 3 delete Claim 4 delete Claim 5 A deep learning-based seismic data recovery method considering data bias for enhanced generalization performance, wherein, in claim 1, the deep learning model of step (b) is based on a U-Net3+ structure including multi-scale feature extraction and skip connections between encoder and decoder, and is configured to perform deep supervision. Claim 6 A deep learning-based seismic data recovery method considering data bias for enhanced generalization performance, wherein, in claim 1, step (b) includes the step of repeatedly inputting training data while varying the intensity of Gaussian random noise to reflect a plurality of signal noise environments in the field. Claim 7 A deep learning-based seismic data recovery method considering data bias for enhanced generalization performance, wherein the recovery process of step (c) above is characterized by improving recovery accuracy by applying a hybrid loss function designed to minimize the residual between the output and input of a learned deep learning model. Claim 8 A deep learning-based seismic data recovery method considering data bias for enhanced generalization performance, characterized in that, in claim 1, it further includes a performance verification step for evaluating the quality of the recovered seismic data to verify the generalization performance of the model. Claim 9 A deep learning-based seismic data recovery method considering data bias for enhanced generalization performance, wherein, in claim 8, the evaluation process of the performance verification step is characterized by calculating the signal-to-noise ratio (SNR) and structural similarity (SSIM) between the restored seismic data and reference data to quantitatively verify the generalization performance. Claim 10 A computer program stored on a storage medium for executing a deep learning-based seismic data recovery method that considers data bias for improved generalization performance of any one of claims 1, 5 through 9 on a computer.

Images