Seismic data denoising method, device, equipment, medium and product

Abstract

The present application relates to the oil and gas artificial intelligence technical field, especially relates to a kind of seismic data denoising method, device, equipment, medium and product, and the method includes: to original seismic data is dimensioned extension, obtains the noisy seismic data matched with convolutional autoencoder;The noisy seismic data is input into convolutional autoencoder, to carry out denoising to the noisy seismic data, obtains denoising seismic data;The error of pre-set noiseless seismic data and the denoising seismic data is back propagated, updates the parameter of the convolutional autoencoder;Denoising seismic data is input into the convolutional autoencoder after back propagation, to carry out denoising to the denoising seismic data;It can improve the denoising effect of seismic data.

Description

Technical Field

[0001] This invention relates to the field of oil and gas artificial intelligence technology, and in particular to a method, apparatus, equipment, medium and product for denoising seismic data. Background Technology

[0002] Among the three requirements of "high signal-to-noise ratio, high resolution, and high fidelity" for seismic data, high signal-to-noise ratio is a fundamental requirement for indoor data processing and is of great significance for subsequent processing and comprehensive interpretation of seismic data. Due to the limitations of resolution and the presence of various noises in seismic exploration, ordinary noise processing techniques are not easy to identify thin layers and complex structures.

[0003] There is a technical problem in this field where the denoising effect of seismic data is poor. Summary of the Invention

[0004] This invention provides a method, apparatus, equipment, medium, and product for denoising seismic data, which solves the technical problem of poor denoising effect of seismic data.

[0005] In a first aspect, the present invention provides a method for denoising seismic data, the method comprising: S100, performing dimensionality expansion on the original seismic data to obtain noisy seismic data matching a convolutional autoencoder; S200, inputting the noisy seismic data into the convolutional autoencoder to denoise the noisy seismic data, thereby obtaining denoised seismic data; S300, backpropagating the error between preset noiseless seismic data and denoised seismic data to update the parameters of the convolutional autoencoder; S400, inputting the seismic data to be denoised into the backpropagated convolutional autoencoder to denoise the seismic data to be denoised.

[0006] In some embodiments, S100, the original seismic data is expanded in dimension to obtain noisy seismic data that matches the convolutional autoencoder, including: S101, windowing the original seismic data in the time-space domain to obtain time-window segmented data; S102, spatially transforming the time-window segmented data to obtain noisy seismic data that matches the convolutional autoencoder.

[0007] In some embodiments, S200, the noisy seismic data is input into a convolutional autoencoder to denoise the noisy seismic data and obtain denoised seismic data, including: S201, the noisy seismic data is input into a convolutional autoencoder to encode the noisy seismic data and obtain high-dimensional compressed features of the seismic data; S202, the high-dimensional compressed features are decoded to obtain denoised seismic data.

[0008] In some embodiments, S201, the noisy seismic data is input into a convolutional autoencoder to encode the noisy seismic data and obtain high-dimensional compressed features of the seismic data, including: S2011, convolving the noisy seismic data to obtain a deep feature map, wherein the convolution kernel includes a 5*5 convolution kernel or a 7*7 convolution kernel; S2012, pooling the deep feature map to obtain high-dimensional compressed features of the seismic data.

[0009] In some embodiments, S202, decoding the high-dimensional compressed features to obtain denoised seismic data includes: S2021, performing unpooling on the high-dimensional compressed features of the seismic data to obtain an extended feature map of the seismic data; S2022, performing deconvolution on the extended feature map to obtain reconstructed features of the seismic data; and S2023, performing spatial inverse transformation on the reconstructed features of the seismic data to obtain denoised seismic data.

[0010] In some embodiments, S300, the error between the preset noiseless seismic data and the denoised seismic data is backpropagated to update the parameters of the convolutional autoencoder, including: S301, calculating the error between the preset noiseless seismic data and the denoised seismic data; S302, backpropagating the error to update the weights and biases of the convolutional autoencoder.

[0011] Secondly, the present invention provides a denoising device for seismic data, comprising: a preprocessing module for performing dimensionality expansion on the original seismic data to obtain noisy seismic data matching a convolutional autoencoder; an encoding / decoding module for inputting the noisy seismic data into the convolutional autoencoder to denoise the noisy seismic data and obtain denoised seismic data; a backpropagation module for backpropagating the error between preset noiseless seismic data and denoised seismic data to update the parameters of the convolutional autoencoder; and a denoising module for inputting the seismic data to be denoised into the backpropagated convolutional autoencoder to denoise the seismic data to be denoised.

[0012] Thirdly, the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the seismic data denoising method of any of the above aspects.

[0013] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the seismic data denoising method of any of the above aspects.

[0014] Fifthly, the present invention provides a computer program product comprising a computer program that, when executed by a processor, implements the steps of the seismic data denoising method according to any of the above aspects.

[0015] This invention provides a method, apparatus, device, medium, and product for denoising seismic data. The method includes: S100, performing dimensionality expansion on the original seismic data to obtain noisy seismic data matched with a convolutional autoencoder; S200, inputting the noisy seismic data into the convolutional autoencoder to denoise the noisy seismic data, obtaining denoised seismic data; S300, backpropagating the error between preset noiseless seismic data and denoised seismic data to update the parameters of the convolutional autoencoder; S400, inputting the seismic data to be denoised into the backpropagated convolutional autoencoder to denoise the seismic data to be denoised; this invention can solve the technical problem of poor denoising effect of seismic data. Attached Figure Description

[0016] The invention will now be described in more detail with reference to embodiments and the accompanying drawings:

[0017] Figure 1 A flowchart illustrating a method for denoising seismic data provided in an embodiment of the present invention;

[0018] Figure 2 This is a schematic diagram of the structure of a seismic data denoising device provided in an embodiment of the present invention;

[0019] Figure 3 A schematic diagram of a technical process provided for an embodiment of the present invention;

[0020] Figure 4 This is a schematic diagram of the structure of a convolutional autoencoder provided in an embodiment of the present invention;

[0021] Figure 5 This is a schematic diagram of the test results of three-dimensional model data provided in an embodiment of the present invention;

[0022] Figure 6 This is a schematic diagram of the test results of three-dimensional actual data of a certain work area provided in an embodiment of the present invention;

[0023] Figure 7 This is a schematic diagram showing the comparison of noise reduction effects before and after small-angle superimposed profiles in a certain work area, provided by an embodiment of the present invention.

[0024] Figure 8 This is a schematic diagram showing the comparison of the noise reduction effects of amplitude slices along a certain work area before and after the present invention.

[0025] Figure 9 This is a schematic diagram showing the comparison of the effects of denoising on pre-stack noisy data provided in an embodiment of the present invention.

[0026] In the accompanying drawings, the same parts are referred to by the same reference numerals, and the drawings are not drawn to scale. Detailed Implementation

[0027] To enable those skilled in the art to better understand the present invention and to fully understand and implement the process of how the present invention uses technical means to solve technical problems and achieve corresponding technical effects, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The embodiments of the present invention and the various features therein can be combined with each other without conflict, and the resulting technical solutions are all within the protection scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0028] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0029] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0030] Among the three requirements of "high signal-to-noise ratio, high resolution, and high fidelity" for seismic data, high signal-to-noise ratio is a fundamental requirement for indoor data processing and is of great significance for subsequent processing and comprehensive interpretation of seismic data. Due to resolution limitations and the presence of various noises in seismic exploration, ordinary noise processing techniques are not effective in identifying thin layers and complex structures. This field suffers from the technical problem of poor denoising performance in seismic data.

[0031] To address the technical problem of poor denoising performance in seismic data, this invention proposes a method, apparatus, equipment, medium, and product for denoising seismic data. The implementation details of this invention are described below for ease of understanding and are not essential for implementing this solution.

[0032] Example 1

[0033] Figure 1 This is a flowchart illustrating a method for denoising seismic data provided in an embodiment of this application, as shown below. Figure 1 As shown in the technical solution of this embodiment, a method for denoising seismic data is provided. The method includes: S100, performing dimensionality expansion on the original seismic data to obtain noisy seismic data matching a convolutional autoencoder; S200, inputting the noisy seismic data into the convolutional autoencoder to denoise the noisy seismic data and obtain denoised seismic data; S300, backpropagating the error between preset noiseless seismic data and denoised seismic data to update the parameters of the convolutional autoencoder; S400, inputting the seismic data to be denoised into the backpropagated convolutional autoencoder to denoise the seismic data to be denoised.

[0034] With the shift in my country's oil and gas exploration objectives, complex surface and subsurface conditions have placed new demands on the quality of 3D seismic data, with a high signal-to-noise ratio being a fundamental requirement for indoor data processing. However, noise contamination is unavoidable during seismic data acquisition, resulting in pre-stack seismic data characterized by large volume and high noise levels. Existing model-driven seismic data denoising techniques all operate under the assumption that noise and effective signals conform to certain rules. When these assumptions are not met, problems such as damage to the effective signal or insufficient noise suppression arise. Therefore, the technical challenge in this field is how to perform data-driven denoising on seismic data.

[0035] In this embodiment, the original seismic data is first expanded in dimensionality by windowing it in the time-space domain to obtain time-windowed data. Then, the time-windowed data undergoes spatial transformation to obtain noisy seismic data that matches the convolutional autoencoder (CAE). The noisy seismic data is input into the CAE, and unsupervised learning combined with convolution and pooling operations is used for feature extraction. The noisy seismic data is first encoded to obtain high-dimensional compressed features, then decoded to obtain denoised seismic data. The error between the preset denoised seismic data and the denoised seismic data is then backpropagated to update the CAE parameters. Finally, the seismic data to be denoised is input into the backpropagated CAE for denoising. For example, deep features are extracted from the data through convolution and pooling operations in the encoding process, and data reconstruction is achieved through unpooling and deconvolution operations in the decoding process, thereby realizing denoising.

[0036] The technical solution in this embodiment expands the original seismic data in higher dimensions to match it with a convolutional autoencoder (CA). The CA is then used to encode and decode noisy seismic data, fully leveraging its data-driven advantages. It is independent of human interference and has no strict requirements or assumptions regarding data and noise. The performance of the CA is continuously optimized during error backpropagation and parameter updates. For the model denoising results, noise is effectively removed. For example, in layer slicing along the target layer, noisy river channels are blurred by noise; after denoising, the outline of the river system is basically revealed. Noise is effectively suppressed, and the signal-to-noise ratio (SNR) is improved, providing high-fidelity, high-SNR foundational data for subsequent seismic data processing, inversion, interpretation, and drilling, overcoming the limitations of traditional denoising techniques.

[0037] Example 2

[0038] Based on the above embodiments, in step S100, the original seismic data is expanded in dimension to obtain noisy seismic data that matches the convolutional autoencoder, including: in step S101, the original seismic data is windowed in the time-space domain to obtain time-window segmented data; in step S102, the time-window segmented data is spatially transformed to obtain noisy seismic data that matches the convolutional autoencoder.

[0039] In the process of denoising seismic data using convolutional autoencoders (CAEs), the format of the raw seismic data is often unsuitable for the CAE network structure. Due to the inherent characteristics of seismic data, its temporal and spatial distribution is complex, and direct input usually fails to enable CAEs to effectively extract features and perform denoising. For example, during seismic data acquisition, the acquired data has multiple dimensions and complex structures, which cannot be directly processed by CAEs. Therefore, there is a technical problem in this field of how to process raw seismic data to facilitate subsequent denoising operations.

[0040] In this embodiment, the dimensionality-upscaling of the original seismic data mainly involves two steps. First, the original seismic data is windowed in the time-space domain, dividing it into time-window segmented data. Windowing can be understood as dividing the data in time and space to better capture local features. Then, the time-window segmented data undergoes spatial transformation, converting it into noisy seismic data that matches the convolutional autoencoder. Spatial transformation mainly involves converting pixel values ​​to the

[01] value range for processing and inversely transforming the convolutional autoencoder denoising structure back to the original space. For example, for each pixel in the seismic data, its value range is adjusted to a suitable interval, making the data more compatible with the input requirements of the convolutional autoencoder. The data after dimensionality-upscaling allows the convolutional autoencoder to perform subsequent feature extraction and denoising operations more effectively.

[0041] The technical solution in this embodiment successfully achieves dimensionality expansion of the original seismic data by performing windowing and spatial transformation operations in the time-space domain, resulting in noisy seismic data that matches the convolutional autoencoder (CAE). This allows the CAE to better process the noisy seismic data, improving the efficiency and accuracy of feature extraction. For example, in subsequent convolution and pooling operations, it can more effectively discover local features of the data, reducing computational complexity. Simultaneously, transforming the data to a suitable numerical range and spatial structure also facilitates the CAE to extract deeper features, thus providing higher-quality input data for subsequent denoising. In seismic data denoising, the data after dimensionality expansion allows the CAE to better remove noise, improving the signal-to-noise ratio.

[0042] Example 3

[0043] Based on the above embodiments, in step S200, the noisy seismic data is input into a convolutional autoencoder to denoise the noisy seismic data and obtain denoised seismic data, including: in step S201, the noisy seismic data is input into a convolutional autoencoder to encode the noisy seismic data and obtain high-dimensional compressed features of the seismic data; in step S202, the high-dimensional compressed features are decoded to obtain denoised seismic data.

[0044] In seismic data processing for oil and gas exploration, pre-stack seismic data often contains significant noise, severely impacting data quality and subsequent processing and interpretation. For example, in subsurface structural regions, noise masks valuable information in seismic signals, making it difficult for conventional processing techniques to accurately identify thin layers and complex structures, resulting in issues such as signal loss or insufficient noise suppression. Therefore, when using convolutional autoencoders to denoise noisy seismic data, obtaining high-quality denoised seismic data is a pressing technical problem that needs to be solved in this field.

[0045] In this embodiment, noisy seismic data is input into a convolutional autoencoder (CAE) for encoding and decoding to achieve denoising. The encoding process involves inputting the noisy seismic data into the CAE and performing a convolution operation. The convolution kernel can be a 5x5 or 7x7 kernel, resulting in a deep feature map. The convolution operation extracts features by performing calculations between the kernel and the input data. Then, pooling is performed on the deep feature map to compress it, simplifying network computation and extracting high-level features. This results in high-dimensional compressed features of the seismic data. Decoding then involves unpooling the high-dimensional compressed features. Unpooling, for example, is upsampling, the inverse of pooling, resulting in an expanded feature map of the seismic data. This expanded feature map is then deconvolved, a process corresponding to convolution, and the output feature map is obtained through convolution summation and an activation function. Finally, spatial inverse transformation yields the denoised seismic data.

[0046] The technical solution in this embodiment effectively denoises noisy seismic data through encoding and decoding. During encoding, convolution and pooling operations extract deep features and compress the data. This embodiment selects a 5x5 or 7x7 convolution kernel, rather than a larger one. Choosing an appropriate kernel size ensures effective feature extraction while avoiding the problem of failing to extract local features due to an excessively large kernel. During decoding, depooling and deconvolution operations restore the high-dimensional compressed features to denoised seismic data. Through testing on 3D model data and actual seismic data, the method in this embodiment effectively removes noise from the seismic data, makes the phase axes of seismic profiles clearer, and improves the accuracy of fault and other structural characterizations through layer-by-layer slicing, thereby enhancing the signal-to-noise ratio and quality of the data and providing more reliable foundational data for subsequent seismic data analysis and structural interpretation.

[0047] Example 4

[0048] Based on the above embodiments, in step S201, the noisy seismic data is input into a convolutional autoencoder to encode the noisy seismic data and obtain high-dimensional compressed features of the seismic data, including: in step S2011, the noisy seismic data is convolved to obtain a deep feature map, wherein the convolution kernel includes a 5*5 convolution kernel or a 7*7 convolution kernel; in step S2012, the deep feature map is pooled to obtain high-dimensional compressed features of the seismic data.

[0049] Noisy seismic data has a complex structure and suffers from severe noise interference. How to encode it appropriately to extract high-dimensional compressed features that accurately reflect the characteristics of the seismic data is a crucial issue. For example, the presence of noise in seismic data blurs its features, making it difficult for ordinary processing methods to accurately extract effective deep features. Furthermore, the choice of convolutional kernel also affects the encoding effect; an inappropriate kernel may fail to fully extract data features or result in excessive computation. Therefore, there is a technical problem in this field of how to encode noisy seismic data to obtain accurate and effective high-dimensional compressed features.

[0050] In this embodiment, encoding noisy seismic data mainly involves two steps. First, the noisy seismic data is convolved. Based on factors such as the size of the input data, seismic data characteristics, processing speed, and memory, this embodiment selects a 5*5 or 7*7 convolution kernel. The convolution kernel performs a convolution operation with the input noisy seismic data. Each convolution kernel is configured with a bias, and after convolution with the input, an activation function input is generated. The activation function then yields a deep feature map. When the input to the convolutional layer consists of multiple feature maps, the input to the activation function is the sum of the multiple convolution results. Convolution operations can extract local features from the data, and the receptive field size is controlled by the size of the convolution kernel. Next, the deep feature map is pooled. The pooling layer is a downsampling layer; commonly used pooling methods include average pooling and max pooling. In this embodiment, a scaling factor of 2 is selected based on the time window size. The pooling layer compresses the input feature map, simplifying network computational complexity, while simultaneously extracting high-level features. Furthermore, pooling is translation-invariant to the feature map. Thus, high-dimensional compressed features of the seismic data are obtained.

[0051] The technical solution in this embodiment effectively encodes noisy seismic data by appropriately selecting the convolution kernel size and performing convolution and pooling operations. A suitable convolution kernel size achieves a balance between feature extraction and computational efficiency; for example, a 5x5 or 7x7 convolution kernel can effectively extract local features from seismic data. Convolution operations can deeply mine data features, while pooling operations further compress and filter features, extracting high-level features. In seismic data processing, the high-dimensional compressed features obtained by the encoding in this embodiment accurately reflect the main features of the seismic data, providing a good foundation for subsequent decoding and denoising. For example, in tests on model data and actual seismic data, the encoding method in this embodiment can effectively remove noise interference and extract clear seismic data features, enabling the subsequent decoding process to more accurately reconstruct high-quality denoised seismic data, thus improving the overall denoising effect and data quality.

[0052] Example 5

[0053] Based on the above embodiments, S202, the high-dimensional compressed features are decoded to obtain denoised seismic data, including: S2021, the high-dimensional compressed features of the seismic data are depooled to obtain an extended feature map of the seismic data; S2022, the extended feature map is deconvolved to obtain reconstructed features of the seismic data; S2023, the reconstructed features of the seismic data are spatially inversely transformed to obtain denoised seismic data.

[0054] In the process of denoising noisy seismic data using convolutional autoencoders, the high-dimensional compressed features of the encoded seismic data require a suitable decoding process to be restored to denoised seismic data. However, the structure of the high-dimensional compressed features differs significantly from the original seismic data, and how to accurately convert them into denoised seismic data is a pressing technical problem. For example, high-dimensional compressed features are obtained through operations such as convolution and pooling, and their feature representation differs from the original data, requiring the restoration of the original data structure and features. Simultaneously, ensuring the accuracy and integrity of the data during decoding, and avoiding information loss or distortion, is also a technical problem that needs to be solved in this field.

[0055] In this embodiment, decoding the high-dimensional compressed features mainly includes three steps. First, the high-dimensional compressed features of the seismic data are unpooled. The unpooling layer is also an upsampling layer. The high-level feature map obtained from the pooling layer is inversely transformed to obtain the feature map before pooling. Then, the expanded feature map of the seismic data obtained from unpooling is deconvolved. The deconvolution layer initializes k convolution kernels, and then obtains the output feature map through corresponding convolution summation and activation functions. If the encoding process uses valid convolution, the deconvolution layer uses full convolution; if the encoding process uses full convolution, the deconvolution layer uses valid convolution. The convolution layer and the deconvolution layer are also inverse processes. Finally, the reconstructed features of the seismic data are spatially inversely transformed, converting the data previously transformed to a specific value range and spatial structure back to the original space to obtain denoised seismic data.

[0056] The technical solution in this embodiment effectively decodes high-dimensional compressed features through unpooling, deconvolution, and spatial inverse transformation. Unpooling recovers feature details lost during pooling, while deconvolution, based on the inverse operation of the encoding process, restores the high-dimensional compressed features to a feature representation close to the original data. Spatial inverse transformation further restores the data to its original spatial structure. In seismic data denoising, this decoding method accurately converts high-dimensional compressed features into denoised seismic data, preserving key information and improving data quality. For example, in seismic data testing, the denoised seismic data obtained after decoding, compared to the original noisy data, shows effective noise removal, clearer phase axes in seismic profiles, and more accurate depiction of geological structures along layer slices, providing reliable data support for subsequent seismic data analysis and interpretation.

[0057] Example 6

[0058] Based on the above embodiments, in step S300, the error between the preset noiseless seismic data and the denoised seismic data is backpropagated to update the parameters of the convolutional autoencoder, including: S301, calculating the error between the preset noiseless seismic data and the denoised seismic data; and S302, backpropagating the error to update the weights and biases of the convolutional autoencoder.

[0059] In the process of denoising seismic data using convolutional autoencoders (CAEs), the parameters of the CAE have a significant impact on the denoising effect. With changes in the processed data and increasing denoising requirements, it is crucial to update the CAE parameters to better adapt to different noisy seismic data and achieve more accurate denoising. For example, in practical seismic data processing, seismic data from different regions have different characteristics and noise levels. If the CAE parameters remain unchanged, noise removal is usually ineffective, and may even lead to damage to the valid signal. Therefore, there is a technical problem in this field of how to update the CAE parameters to improve its denoising performance and adaptability.

[0060] In this embodiment, the error between preset noiseless seismic data and denoised seismic data is first calculated. By comparing the difference between the two, an error value reflecting the denoising effect is obtained. The error can be calculated using specific mathematical formulas; for example, different error calculation methods exist for the output layer and other layers. Then, the error is backpropagated. Backpropagation uses the BP algorithm to transmit residuals. Based on the magnitude and direction of the error, the weights and biases of the convolutional autoencoder are adjusted. There are corresponding update formulas for the convolutional kernel weights and convolutional layer biases. By continuously adjusting the weights and biases, the convolutional autoencoder can continuously update based on the training data, ultimately enabling the network to directly compute denoised data from noisy data.

[0061] The technical solution in this embodiment updates the parameters of the convolutional autoencoder by calculating errors and performing backpropagation, effectively improving the denoising performance and adaptability of the convolutional autoencoder. In seismic data processing, by continuously adjusting the parameters, the convolutional autoencoder can better adapt to different noisy seismic data. For example, for seismic data with high noise levels, after parameter updates, the convolutional autoencoder can more accurately identify and remove noise while retaining effective signals. In tests on model data and seismic data, the convolutional autoencoder after parameter updates showed a significant improvement in denoising performance, increased the signal-to-noise ratio of the seismic data, clearer phase axes in seismic profiles, and more accurate characterization of geological structures by layer-by-layer slices, providing a more reliable foundation for subsequent seismic data processing and analysis.

[0062] Example 7

[0063] Figure 2 This is a schematic diagram of the structure of a seismic data denoising device provided in an embodiment of this application, as shown below. Figure 2 As shown in the technical solution of this embodiment, a seismic data denoising device is provided. The device includes: a preprocessing module for performing dimensionality expansion on the original seismic data to obtain noisy seismic data matching a convolutional autoencoder; an encoding / decoding module for inputting the noisy seismic data into the convolutional autoencoder to denoise the noisy seismic data and obtain denoised seismic data; a backpropagation module for backpropagating the error between preset noiseless seismic data and denoised seismic data to update the parameters of the convolutional autoencoder; and a denoising module for inputting the seismic data to be denoised into the backpropagated convolutional autoencoder to denoise the seismic data to be denoised.

[0064] With the shift in my country's oil and gas exploration objectives, complex surface and subsurface conditions have placed new demands on the quality of 3D seismic data, with a high signal-to-noise ratio being a fundamental requirement for indoor data processing. However, noise contamination is unavoidable during seismic data acquisition, resulting in pre-stack seismic data characterized by large volume and high noise levels. Existing model-driven seismic data denoising techniques all operate under the assumption that noise and effective signals conform to certain rules. When these assumptions are not met, problems such as damage to the effective signal or insufficient noise suppression arise. Therefore, the technical challenge in this field is how to perform data-driven denoising on seismic data.

[0065] In this embodiment, the original seismic data is first expanded in dimensionality by windowing it in the time-space domain to obtain time-windowed data. Then, the time-windowed data undergoes spatial transformation to obtain noisy seismic data that matches the convolutional autoencoder (CAE). The noisy seismic data is input into the CAE, and unsupervised learning combined with convolution and pooling operations is used for feature extraction. The noisy seismic data is first encoded to obtain high-dimensional compressed features, then decoded to obtain denoised seismic data. The error between the preset denoised seismic data and the denoised seismic data is then backpropagated to update the CAE parameters. Finally, the seismic data to be denoised is input into the backpropagated CAE for denoising. For example, deep features are extracted from the data through convolution and pooling operations in the encoding process, and data reconstruction is achieved through unpooling and deconvolution operations in the decoding process, thereby realizing denoising.

[0066] The technical solution in this embodiment expands the original seismic data in higher dimensions to match it with a convolutional autoencoder (CA). The CA is then used to encode and decode noisy seismic data, fully leveraging its data-driven advantages. It is independent of human interference and has no strict requirements or assumptions regarding data and noise. The performance of the CA is continuously optimized during error backpropagation and parameter updates. For the model denoising results, noise is effectively removed. For example, in layer slicing along the target layer, noisy river channels are blurred by noise; after denoising, the outline of the river system is basically revealed. Noise is effectively suppressed, and the signal-to-noise ratio (SNR) is improved, providing high-fidelity, high-SNR foundational data for subsequent seismic data processing, inversion, interpretation, and drilling, overcoming the limitations of traditional denoising techniques.

[0067] Based on the above embodiments, in step S100, the original seismic data is expanded in dimension to obtain noisy seismic data that matches the convolutional autoencoder, including: in step S101, the original seismic data is windowed in the time-space domain to obtain time-window segmented data; in step S102, the time-window segmented data is spatially transformed to obtain noisy seismic data that matches the convolutional autoencoder.

[0068] In the process of denoising seismic data using convolutional autoencoders (CAEs), the format of the raw seismic data is often unsuitable for the CAE network structure. Due to the inherent characteristics of seismic data, its temporal and spatial distribution is complex, and direct input usually fails to enable CAEs to effectively extract features and perform denoising. For example, during seismic data acquisition, the acquired data has multiple dimensions and complex structures, which cannot be directly processed by CAEs. Therefore, there is a technical problem in this field of how to process raw seismic data to facilitate subsequent denoising operations.

[0069] In this embodiment, the dimensionality-upscaling of the original seismic data mainly involves two steps. First, the original seismic data is windowed in the time-space domain, dividing it into time-window segmented data. Windowing can be understood as dividing the data in time and space to better capture local features. Then, the time-window segmented data undergoes spatial transformation, converting it into noisy seismic data that matches the convolutional autoencoder. Spatial transformation mainly involves converting pixel values ​​to the

[01] value range for processing and inversely transforming the convolutional autoencoder denoising structure back to the original space. For example, for each pixel in the seismic data, its value range is adjusted to a suitable interval, making the data more compatible with the input requirements of the convolutional autoencoder. The data after dimensionality-upscaling allows the convolutional autoencoder to perform subsequent feature extraction and denoising operations more effectively.

[0070] The technical solution in this embodiment successfully achieves dimensionality expansion of the original seismic data by performing windowing and spatial transformation operations in the time-space domain, resulting in noisy seismic data that matches the convolutional autoencoder (CAE). This allows the CAE to better process the noisy seismic data, improving the efficiency and accuracy of feature extraction. For example, in subsequent convolution and pooling operations, it can more effectively discover local features of the data, reducing computational complexity. Simultaneously, transforming the data to a suitable numerical range and spatial structure also facilitates the CAE to extract deeper features, thus providing higher-quality input data for subsequent denoising. In seismic data denoising, the data after dimensionality expansion allows the CAE to better remove noise, improving the signal-to-noise ratio.

[0071] Based on the above embodiments, in step S200, the noisy seismic data is input into a convolutional autoencoder to denoise the noisy seismic data and obtain denoised seismic data, including: in step S201, the noisy seismic data is input into a convolutional autoencoder to encode the noisy seismic data and obtain high-dimensional compressed features of the seismic data; in step S202, the high-dimensional compressed features are decoded to obtain denoised seismic data.

[0072] In seismic data processing for oil and gas exploration, pre-stack seismic data often contains significant noise, severely impacting data quality and subsequent processing and interpretation. For example, in subsurface structural regions, noise masks valuable information in seismic signals, making it difficult for conventional processing techniques to accurately identify thin layers and complex structures, resulting in issues such as signal loss or insufficient noise suppression. Therefore, when using convolutional autoencoders to denoise noisy seismic data, obtaining high-quality denoised seismic data is a pressing technical problem that needs to be solved in this field.

[0073] In this embodiment, noisy seismic data is input into a convolutional autoencoder (CAE) for encoding and decoding to achieve denoising. The encoding process involves inputting the noisy seismic data into the CAE and performing a convolution operation. The convolution kernel can be a 5x5 or 7x7 kernel, resulting in a deep feature map. The convolution operation extracts features by performing calculations between the kernel and the input data. Then, pooling is performed on the deep feature map to compress it, simplifying network computation and extracting high-level features. This results in high-dimensional compressed features of the seismic data. Decoding then involves unpooling the high-dimensional compressed features. Unpooling, for example, is upsampling, the inverse of pooling, resulting in an expanded feature map of the seismic data. This expanded feature map is then deconvolved, a process corresponding to convolution, and the output feature map is obtained through convolution summation and an activation function. Finally, spatial inverse transformation yields the denoised seismic data.

[0074] The technical solution of this embodiment effectively denoises noisy seismic data by encoding and decoding it. During encoding, convolution and pooling operations extract deep features and compress the data. This embodiment selects a 5x5 or 7x7 convolution kernel, rather than a larger one. Choosing an appropriate kernel size ensures effective feature extraction while avoiding the problem of failing to extract local features due to an excessively large kernel. During decoding, depooling and deconvolution operations restore the high-dimensional compressed features to denoised seismic data. Through testing on 3D model data and actual seismic data, the device in this embodiment effectively removes noise from the seismic data, makes the phase axes of seismic profiles clearer, and improves the accuracy of fault and other structural characterizations through layer-by-layer slicing, thereby enhancing the signal-to-noise ratio and quality of the data and providing more reliable foundational data for subsequent seismic data analysis and structural interpretation.

[0075] Based on the above embodiments, in step S201, the noisy seismic data is input into a convolutional autoencoder to encode the noisy seismic data and obtain high-dimensional compressed features of the seismic data, including: in step S2011, the noisy seismic data is convolved to obtain a deep feature map, wherein the convolution kernel includes a 5*5 convolution kernel or a 7*7 convolution kernel; in step S2012, the deep feature map is pooled to obtain high-dimensional compressed features of the seismic data.

[0076] Noisy seismic data has a complex structure and suffers from severe noise interference. How to encode it appropriately to extract high-dimensional compressed features that accurately reflect the characteristics of the seismic data is a crucial issue. For example, the presence of noise in seismic data blurs its features, making it difficult for ordinary processing methods to accurately extract effective deep features. Furthermore, the choice of convolutional kernel also affects the encoding effect; an inappropriate kernel may fail to fully extract data features or result in excessive computation. Therefore, there is a technical problem in this field of how to encode noisy seismic data to obtain accurate and effective high-dimensional compressed features.

[0077] In this embodiment, encoding noisy seismic data mainly involves two steps. First, the noisy seismic data is convolved. Based on factors such as the size of the input data, seismic data characteristics, processing speed, and memory, this embodiment selects a 5*5 or 7*7 convolution kernel. The convolution kernel performs a convolution operation with the input noisy seismic data. Each convolution kernel is configured with a bias, and after convolution with the input, an activation function input is generated. The activation function then yields a deep feature map. When the input to the convolutional layer consists of multiple feature maps, the input to the activation function is the sum of the multiple convolution results. Convolution operations can extract local features from the data, and the receptive field size is controlled by the size of the convolution kernel. Next, the deep feature map is pooled. The pooling layer is a downsampling layer; commonly used pooling methods include average pooling and max pooling. In this embodiment, a scaling factor of 2 is selected based on the time window size. The pooling layer compresses the input feature map, simplifying network computational complexity, while simultaneously extracting high-level features. Furthermore, pooling is translation-invariant to the feature map. Thus, high-dimensional compressed features of the seismic data are obtained.

[0078] The technical solution in this embodiment effectively encodes noisy seismic data by appropriately selecting the convolution kernel size and performing convolution and pooling operations. A suitable convolution kernel size achieves a balance between feature extraction and computational efficiency; for example, a 5x5 or 7x7 convolution kernel can effectively extract local features from seismic data. Convolution operations can deeply mine data features, while pooling operations further compress and filter features, extracting high-level features. In seismic data processing, the high-dimensional compressed features obtained by the encoding in this embodiment accurately reflect the main features of the seismic data, providing a good foundation for subsequent decoding and denoising. For example, in tests on model data and actual seismic data, the encoding method in this embodiment can effectively remove noise interference and extract clear seismic data features, enabling the subsequent decoding process to more accurately reconstruct high-quality denoised seismic data, thus improving the overall denoising effect and data quality.

[0079] Based on the above embodiments, S202, the high-dimensional compressed features are decoded to obtain denoised seismic data, including: S2021, the high-dimensional compressed features of the seismic data are depooled to obtain an extended feature map of the seismic data; S2022, the extended feature map is deconvolved to obtain reconstructed features of the seismic data; S2023, the reconstructed features of the seismic data are spatially inversely transformed to obtain denoised seismic data.

[0080] In the process of denoising noisy seismic data using convolutional autoencoders, the high-dimensional compressed features of the encoded seismic data require a suitable decoding process to be restored to denoised seismic data. However, the structure of the high-dimensional compressed features differs significantly from the original seismic data, and how to accurately convert them into denoised seismic data is a pressing technical problem. For example, high-dimensional compressed features are obtained through operations such as convolution and pooling, and their feature representation differs from the original data, requiring the restoration of the original data structure and features. Simultaneously, ensuring the accuracy and integrity of the data during decoding, and avoiding information loss or distortion, is also a technical problem that needs to be solved in this field.

[0081] In this embodiment, decoding the high-dimensional compressed features mainly includes three steps. First, the high-dimensional compressed features of the seismic data are unpooled. The unpooling layer is also an upsampling layer. The high-level feature map obtained from the pooling layer is inversely transformed to obtain the feature map before pooling. Then, the expanded feature map of the seismic data obtained from unpooling is deconvolved. The deconvolution layer initializes k convolution kernels, and then obtains the output feature map through corresponding convolution summation and activation functions. If the encoding process uses valid convolution, the deconvolution layer uses full convolution; if the encoding process uses full convolution, the deconvolution layer uses valid convolution. The convolution layer and the deconvolution layer are also inverse processes. Finally, the reconstructed features of the seismic data are spatially inversely transformed, converting the data previously transformed to a specific value range and spatial structure back to the original space to obtain denoised seismic data.

[0082] The technical solution in this embodiment effectively decodes high-dimensional compressed features through unpooling, deconvolution, and spatial inverse transformation. Unpooling recovers feature details lost during pooling, while deconvolution, based on the inverse operation of the encoding process, restores the high-dimensional compressed features to a feature representation close to the original data. Spatial inverse transformation further restores the data to its original spatial structure. In seismic data denoising, this decoding method accurately converts high-dimensional compressed features into denoised seismic data, preserving key information and improving data quality. For example, in seismic data testing, the denoised seismic data obtained after decoding, compared to the original noisy data, shows effective noise removal, clearer phase axes in seismic profiles, and more accurate depiction of geological structures along layer slices, providing reliable data support for subsequent seismic data analysis and interpretation.

[0083] Based on the above embodiments, in step S300, the error between the preset noiseless seismic data and the denoised seismic data is backpropagated to update the parameters of the convolutional autoencoder, including: S301, calculating the error between the preset noiseless seismic data and the denoised seismic data; and S302, backpropagating the error to update the weights and biases of the convolutional autoencoder.

[0084] In the process of denoising seismic data using convolutional autoencoders (CAEs), the parameters of the CAE have a significant impact on the denoising effect. With changes in the processed data and increasing denoising requirements, it is crucial to update the CAE parameters to better adapt to different noisy seismic data and achieve more accurate denoising. For example, in practical seismic data processing, seismic data from different regions have different characteristics and noise levels. If the CAE parameters remain unchanged, noise removal is usually ineffective, and may even lead to damage to the valid signal. Therefore, there is a technical problem in this field of how to update the CAE parameters to improve its denoising performance and adaptability.

[0085] In this embodiment, the error between preset noiseless seismic data and denoised seismic data is first calculated. By comparing the difference between the two, an error value reflecting the denoising effect is obtained. The error can be calculated using specific mathematical formulas; for example, different error calculation methods exist for the output layer and other layers. Then, the error is backpropagated. Backpropagation uses the BP algorithm to transmit residuals. Based on the magnitude and direction of the error, the weights and biases of the convolutional autoencoder are adjusted. There are corresponding update formulas for the convolutional kernel weights and convolutional layer biases. By continuously adjusting the weights and biases, the convolutional autoencoder can continuously update based on the training data, ultimately enabling the network to directly compute denoised data from noisy data.

[0086] The technical solution in this embodiment updates the parameters of the convolutional autoencoder by calculating errors and performing backpropagation, effectively improving the denoising performance and adaptability of the convolutional autoencoder. In seismic data processing, by continuously adjusting the parameters, the convolutional autoencoder can better adapt to different noisy seismic data. For example, for seismic data with high noise levels, after parameter updates, the convolutional autoencoder can more accurately identify and remove noise while retaining effective signals. In tests on model data and seismic data, the convolutional autoencoder after parameter updates showed a significant improvement in denoising performance, increased the signal-to-noise ratio of the seismic data, clearer phase axes in seismic profiles, and more accurate characterization of geological structures by layer-by-layer slices, providing a more reliable foundation for subsequent seismic data processing and analysis.

[0087] Example 8

[0088] In this embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps of any of the above embodiments' seismic data denoising methods.

[0089] In the technical solution of this embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the seismic data denoising method of any of the above embodiments.

[0090] In the technical solution of this embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of any of the above embodiments' seismic data denoising methods.

[0091] The processor may include, but is not limited to, one or more processors or microprocessors. Each processor may be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic component, for performing the methods in the above embodiments. The computer-readable storage medium may be implemented by any type of volatile or non-volatile storage device or a combination thereof, and may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, EPROM memory, EEPROM memory, registers, computer storage media (e.g., hard disk, floppy disk, solid-state drive, removable disk, CD-ROM, DVD-ROM, Blu-ray disc, etc.).

[0092] Computer-readable storage media may also store at least one computer-executable program / instruction, such as computer-readable instructions. Computer-readable storage media include, but are not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Computer-readable storage media may include, for example, read-only memory (ROM), hard disk, flash memory, etc. For example, a non-transitory computer-readable storage medium may be connected to a computing device such as a computer, and then, when the computing device executes the computer-readable instructions stored on the computer-readable storage medium, the various methods described above can be performed.

[0093] In addition, the computer device may also include (but is not limited to) a data bus, an input / output (I / O) bus, a display, and input / output devices (e.g., a keyboard, mouse, speakers, etc.). The processor can communicate with external devices via the I / O bus through a wired or wireless network. In one embodiment, the at least one computer-executable instruction may also be compiled into or comprise a software product / computer program product, wherein one or more computer-executable instructions, when executed by the processor, perform the steps of the various functions and / or methods in the embodiments described herein.

[0094] Example 9

[0095] Based on the above embodiments, this embodiment provides an application example.

[0096] This application example provides a high-efficiency amplitude-preserving pre-stack seismic data denoising device and system based on a convolutional autoencoder.

[0097] This application example belongs to the field of oil and gas artificial intelligence technology, specifically involving signal processing and exploration geophysical processing in geosciences. This example proposes a highly efficient amplitude-preserving pre-stack seismic data denoising device and system based on a convolutional autoencoder. It employs a deep convolutional autoencoder (DCAE) technique to directly extract data-driven features from the raw seismic data, and then uses deep features in the depth feature space for data reconstruction and denoising, thereby achieving efficient amplitude-preserving processing of pre-stack seismic data. This technology can provide strong technical support for preprocessing such as oil and gas geophysical data cleaning, and provide high signal-to-noise ratio foundational data for subsequent structural interpretation, seismic inversion, reservoir prediction, drilling, and other work.

[0098] As my country's oil and gas exploration targets gradually shift from conventional oil and gas fields to "deep and non-deep" and "four new" fields, complex surface conditions, complex underground structures, and complex target reservoirs have placed new demands on the quality of 3D seismic data. Among the "high signal-to-noise ratio, high resolution, and high fidelity" requirements for seismic data, high signal-to-noise ratio is a fundamental requirement for indoor data processing and is crucial for subsequent processing and comprehensive interpretation. In the geophysical field, noise reduction in pre-stack data is mainly achieved through two key stages: acquisition and processing. In the acquisition stage, compared to 2D and conventional 3D seismic acquisition, "two-wide and one-high" seismic acquisition technology, which integrates well-seismic mixed acquisition, low-frequency controllable sources, omnidirectional and high-density techniques, improves the resolution and signal-to-noise ratio of seismic data. However, noise pollution is still unavoidable during acquisition, resulting in pre-stack seismic data characterized by "large volume and high noise." Due to resolution limitations and the presence of various noises in seismic exploration, ordinary processing techniques are insufficient to identify thin layers and complex structures, making noise suppression imperative. The field currently faces the technical problem of poor denoising effects in seismic data.

[0099] This application example method is directly data-driven, using a deep convolutional autoencoder (DCAE) to transform the feature space of the original seismic data, thereby obtaining data features in the depth feature space. It makes full use of the information in the original seismic data. This technique is flexible and efficient, does not rely on human interference, and has no strict requirements or assumptions about data and noise. It has better amplitude preservation effect than traditional denoising methods, thus providing strong technical support for noise removal.

[0100] Existing seismic data denoising techniques are model-driven, processing noise and effective signals based on certain assumptions. For example, mathematical transform denoising methods based on sparse representation assume that the seismic signal is sparse in a certain transform domain, while the seismic noise is non-sparse, utilizing the difference in their characteristics in the transform domain to suppress noise. However, when the data and noise do not meet the assumptions, traditional methods such as mathematical transforms are limited to a certain extent, resulting in technical problems such as damage to the effective signal or insufficient noise suppression. Therefore, it is necessary to invent a data-driven, intelligent, adaptive pre-stack seismic data denoising technique.

[0101] Convolutional autoencoder (CA) denoising technology is a data-driven technique that does not require any assumptions about the signal itself, breaking through the theoretical assumptions and preconditions of traditional methods. This CADe technology leverages the flexible and efficient learning system of deep learning, as well as its deep spatial feature transformation and learning capabilities. It first encodes the noisy pre-stack data / raw seismic data into feature vectors through nonlinear dimensionality reduction, and then decodes and decompresses and upscales the feature vectors in the feature space to reconstruct noise-free seismic data of the same size as the input.

[0102] In the process of learning a set of vectors that sparsely represent the main features of pre-stack data, the convolutional autoencoder (CAE) achieves effective signal representation of seismic data. Through high-dimensional nonlinear mapping using the CAE, intelligent denoising of seismic data can be effectively performed, while maintaining amplitude. Furthermore, this CAE denoising technique has the ability to remember and transfer learn from processed data, making it more suitable for automated denoising of large-scale big data compared to shallow learners such as fully connected neural networks. After applying transfer learning, this CAE denoising technique can provide high-quality foundational data for subsequent seismic data processing, multi-scale fault interpretation, geological sweet spot prediction, seismic-guided drilling, and fine reservoir characterization.

[0103] Convolutional autoencoders (CAEs) utilize the unsupervised learning approach of traditional autoencoders. Unsupervised learning allows for the learning of sample features without labels. CAEs combine convolution and pooling operations from convolutional neural networks to achieve feature extraction, and finally, through stacking, create a deep neural network. Its encoding process extracts features from the original data through convolution and pooling, while the decoding process reconstructs the data through unpooling and deconvolution. With the advancement of computer processing power, stacking CAEs to process large datasets has become possible, making it a research hotspot in many scientific fields. A CAE is a special type of convolutional neural network and also a special type of autoencoder. Compared to autoencoders and denoising autoencoders, CAEs have a simpler structure, fewer network parameters, and better robustness.

[0104] The overall network structure of a Convolutional Autoencoder (CAE). CAEs utilize backpropagation for error propagation and weight updates. First, weights and biases are randomly assigned to the pre-defined network structure, or optimized weights and biases from earlier data are used as the initial model. Training data is then imported, and forward propagation is used to calculate the output value. Backpropagation updates the weights and biases through error propagation. Its main structural ideas are: local region awareness, weight sharing, spatial or temporal upsampling, and deep feature extraction. Local region awareness allows the encoding process to discover local features in the data; weight sharing reduces the number of parameters required for training, resulting in a more generalizable model and faster training speed; spatial or temporal upsampling aims to obfuscate the specific locations of features, enabling CAEs to recognize deformed and distorted images; deep feature extraction involves multiple feature space transformations during the encoding process to extract deep features from the original data, allowing the network to extract representative deep features.

[0105] Convolutional autoencoder and data initialization. Data initialization involves initializing the seismic data into a mode suitable for the autoencoder using a windowing method. Convolutional autoencoder initialization involves manually setting / presetting the number of convolutional layers and kernel size in the encoding process, as well as the number of deconvolutional layers and kernel size in the decoding process, and assigning initial values ​​to the weights and biases involved in each part.

[0106] Forward pass operation. The convolutional autoencoder mainly consists of an encoding process and a decoding process. The encoding process includes an input layer, a convolutional layer, and a pooling layer, which yields high-dimensional compressed features of the seismic data. The decoding process includes an unpooling layer, a deconvolutional layer, and an output layer.

[0107] Input layer. The input layer data consists of preprocessed raw seismic data. The main preprocessing operations are: expanding the dimensionality of each pixel in the seismic data by opening a window of the same size in the time-space domain, performing spatial transformation on the pixel values ​​to convert them to the value range of [0 1] for processing, and performing inverse transformation of the convolutional autoencoder denoising structure back to the original space. The purpose of the transformation is to make the data more suitable for the network structure of the convolutional autoencoder.

[0108] Convolutional Layer. A convolutional layer initializes k convolutional kernels W, each configured with a bias b. These kernels are convolved with the input x to generate k activation function inputs, which in turn produce k feature maps. When the input to the convolutional layer consists of multiple feature maps, the input to the activation function is the sum of the results of the multiple convolutions. The expression is as follows:

[0109]

[0110] Where i is the number of neurons in the input layer, i.e., the number of input feature maps in the previous layer, j is the number of output feature maps in this convolutional layer, and Lcov represents the number of convolutional layers. This is the output feature map of the i-th neuron node in the L-th layer. This is the output feature map of the i-th neuron node in the Lcov-1 layer, and when Lcov = 1, it is the original input data. Let Lcov be the convolution kernel connecting the i-th neuron in Lcov-1 layer and the j-th neuron in Lcov layer. σ represents the activation function, corresponding to the bias of the convolution kernel. Typically, the receptive field size is controlled by the size of the convolution kernel; a larger kernel uses more data for computation, but larger is not always better. When the kernel is too large, it cannot extract effective local features. Therefore, the kernel size needs to be selected considering factors such as the size of the input data, the characteristics of the seismic data, processing speed, and memory. In this application example, a 5x5 or 7x7 convolution kernel is selected. The kernel is used to convolve the time windows of the input data, ultimately obtaining a two-dimensional feature map for each time window through feature space transformation.

[0111] Pooling layers. A pooling layer is a feature mapping layer, also known as a downsampling layer. Pooling layers serve two purposes: first, they compress the input feature map, reducing its size and simplifying network computation; second, they perform feature compression transformations to extract higher-level features. Pooling is translation-invariant to feature maps. Two commonly used pooling methods are average pooling and max pooling.

[0112] In the convolutional autoencoder of this application example, the scaling factor selected based on the time window size is 2, and when the number of layers is high, the arrangement of convolutional layers and pooling layers can be flexibly selected, for example, three convolutional layers followed by one pooling layer.

[0113] Unpooling layer. The unpooling layer is also known as the upsampling layer. It is a process that uses the high-level feature map obtained from the pooling layer to perform an inverse transformation to obtain the feature map before pooling. The pooling layer and the unpooling layer are inverse processes.

[0114] Deconvolutional layer. The deconvolutional layer initializes k convolutional kernels W', and then obtains the output feature map through corresponding convolution summations and activation functions. The calculation approach is the same as the convolutional layer, except that if the encoding process uses valid convolutions, the deconvolutional layer uses full convolutions, and vice versa. The convolutional and deconvolutional layers are also inverse processes. The expression is as follows:

[0115]

[0116] When Ldec is the last layer, h is the decoded output y of the encoder. This is the output feature map of the i-th neuron node in the Ldec-1 deconvolution layer. This is the convolution kernel connecting the i-th neuron of the Ldec-1 deconvolution layer to the j-th neuron of the Ldec layer. σ represents the activation function, corresponding to the bias of the convolution kernel in the deconvolution layer.

[0117] Output layer. The output layer is the last layer in the decoding process. This layer receives the denoised image y derived from the inverse transformation of deep features, calculates the error between the denoised image y and the noiseless image x, and then passes the result to the previous layer. A data-driven nonlinear mapping between noisy and noiseless data is established, thereby optimizing the denoising function (CAE).

[0118] Error backpropagation and weight update. Similar to convolutional neural networks, convolutional autoencoders also use the backpropagation algorithm to propagate residuals and adjust the weights W and biases b, allowing the convolutional autoencoder to continuously update based on the training data. Ultimately, this enables the network to directly compute denoised data from noisy data. The expressions for error backpropagation and weight update are as follows:

[0119] Error propagation.

[0120] 1. For the output layer:

[0121]

[0122] Where i represents the number of output images.

[0123] 2. For other layers:

[0124] It mainly consists of two parts: one is crossing the activation function, and the other is passing the error between layers by multiplying by weights. The expression is as follows:

[0125] δ L =(W L+1 ) T *δ L+1 *f'(u) (4)

[0126] Where u is the output of layer L, δ L This represents the error of layer L.

[0127] Weights updated.

[0128] 1. For the convolution kernel weights ω:

[0129]

[0130] 2. For the convolutional layer bias b:

[0131]

[0132] This application example proposes a data-driven intelligent amplitude-preserving signal-to-noise enhancement technique for seismic data. This technique deeply mines and captures the detailed information contained in the original data. Based on data-driven principles and leveraging a highly efficient and flexible large-scale deep learning system, it establishes a complex nonlinear mapping relationship between noisy and blurred data and denoised, clear data. Through transfer learning and self-learning, it adaptively improves the quality of seismic data, providing high-fidelity, high signal-to-noise ratio foundational data for subsequent seismic data processing, inversion, interpretation, and drilling. This method overcomes the reliance on various assumptions in conventional denoising techniques, requiring no assumptions and thus potentially breaking through the bottleneck of low fidelity in complex data areas from the outset. Compared to conventional methods, this application example significantly reduces human interference, and its data-driven potential is greater than that of conventional model-driven methods.

[0133] Figure 3 This is a technical flowchart for this application example.

[0134] To verify the practical application effect of the method in this application example, this application example selected a three-dimensional pushover model data with a common angular domain and a three-dimensional actual data with a common angular domain from a field area to test the denoising technology of this application example.

[0135] Figure 4 This is a structural diagram of the convolutional autoencoder used in this test.

[0136] Figure 5 The test results are for the 3D model data, where Figure 5 (a) Figure 5(b) The results of adding and removing noise for Inline131 are shown respectively; Figure 5 (c) Figure 5 (d) shows the noise-included and denoised results for Inline250, respectively; Figure 5 (e) Figure 5 (f) are slices of the amplitude along a target layer before and after denoising.

[0137] Figure 6 The test results are based on actual 3D data of a certain region, among which... Figure 6 (a) Figure 6 (b) The results of Inline155 with and without noise are respectively; Figure 6 (c) Figure 6 (d) shows the noise-injected and noise-removed results for Inline201, respectively; Figure 6 (e) Figure 6 (f) are slices of the amplitude along a target layer before and after denoising.

[0138] Figure 7 This is a schematic diagram comparing the noise reduction effects of small-angle superimposed profiles before and after noise reduction in this work area. Figure 7 a, 7b, and 7c represent the seismic profile before denoising, the seismic profile after denoising, and the noise removed, respectively. Figure 7 A comparison of 7b shows that after the convolutional autoencoder denoising process, the seismic phase axis of the seismic profile is clearer and the lateral continuity of the profile is better. Figure 7 c indicates that the convolutional autoencoder performs well in both noise reduction and amplitude preservation.

[0139] Figure 8 This is a schematic diagram comparing the noise reduction effects of amplitude slices along the layer in this work area before and after the process. Among them, Figure 8 a, 8b, and 8c represent the amplitude slices before denoising, the amplitude slices after denoising, and the removed noise, respectively. Figure 8 A comparison of 8b and 8b shows that after the convolutional autoencoder denoising process, the slices along the layers more clearly depict the extension direction and development scale of the fault.

[0140] Figure 9 This is a diagram comparing the denoising effects before and after denoising the pre-stack noisy data. Among them, Figure 9 a and 9b represent the pre-stack noisy seismic data and their denoising results, respectively.

[0141] From the model denoising results, this method effectively removes noise, especially for slices along the target layer of the model. The noisy river channels in the data are blurred by noise interference, but after denoising, the outline of the river system (including oxbow lakes and the direction of river extension) is basically revealed, which is more conducive to subsequent interpretation. From the denoising results of actual data, this method also effectively suppresses noise in the actual data, improving the signal-to-noise ratio.

[0142] This application example proposes a data-driven intelligent amplitude-preserving signal-to-noise enhancement technique for seismic data. This technique can deeply mine and capture the detailed information contained in the original data. Based on data-driven principles and leveraging a highly efficient and flexible large-scale deep learning system, it establishes a complex nonlinear mapping relationship between noisy and blurred data and denoised, clear data. Through transfer learning and self-learning, it can adaptively improve the quality of seismic data, providing high-fidelity, high signal-to-noise ratio foundational data for subsequent seismic data processing, inversion, interpretation, and drilling. This method overcomes the reliance on various assumptions in conventional denoising techniques, requiring no assumptions and thus potentially breaking through the bottleneck of low fidelity in complex data areas from the source. Compared to conventional methods, it significantly reduces human interference, and its unique data-driven potential surpasses that of conventional model-driven methods.

[0143] In the embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0144] It should be noted that, in this invention, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element limited by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0145] While the embodiments disclosed in this invention are as described above, the above content is merely for the purpose of facilitating understanding of this invention and is not intended to limit the invention. Any person skilled in the art to which this invention pertains may make any modifications and changes in form and detail of the implementation without departing from the spirit and scope disclosed in this invention; however, the scope of patent protection of this invention shall still be determined by the scope defined in the appended claims.

Claims

1. A method for denoising seismic data, characterized in that, The method includes: S100 performs dimensionality expansion on the original seismic data to obtain noisy seismic data that matches the convolutional autoencoder; S200, the noisy seismic data is input into a convolutional autoencoder to denoise the noisy seismic data and obtain denoised seismic data; S300, backpropagate the error between the preset noiseless seismic data and the denoised seismic data, and update the parameters of the convolutional autoencoder. S400, the seismic data to be denoised is input into the convolutional autoencoder after backpropagation to denoise the seismic data to be denoised.

2. The method for denoising seismic data according to claim 1, characterized in that, S100 involves upscaling the original seismic data to obtain noisy seismic data that matches the convolutional autoencoder, including: S101, Windowing is performed on the raw seismic data in the time-space domain to obtain time-window segmented data; S102, Spatial transformation is performed on the time window segmentation data to obtain noisy seismic data that matches the convolutional autoencoder.

3. The method for denoising seismic data according to claim 1, characterized in that, In step S200, the noisy seismic data is input into a convolutional autoencoder to denoise the noisy seismic data, resulting in denoised seismic data, including: S201, The noisy seismic data is input into a convolutional autoencoder to encode the noisy seismic data and obtain the high-dimensional compressed features of the seismic data; S202, decode the high-dimensional compressed features to obtain denoised seismic data.

4. The method for denoising seismic data according to claim 3, characterized in that, In step S201, the noisy seismic data is input into a convolutional autoencoder to encode the noisy seismic data, thereby obtaining high-dimensional compressed features of the seismic data, including: S2011, Convolve the noisy seismic data to obtain a deep feature map, wherein the convolution kernel includes a 5*5 convolution kernel or a 7*7 convolution kernel; S2012, the deep feature map is pooled to obtain high-dimensional compressed features of the seismic data.

5. The method for denoising seismic data according to claim 3, characterized in that, S202 involves decoding the high-dimensional compressed features to obtain denoised seismic data, including: S2021, unpooling is performed on the high-dimensional compressed features of the seismic data to obtain an extended feature map of the seismic data; S2022, Deconvolve the extended feature map to obtain the reconstructed features of the seismic data; S2023, the reconstructed features of the earthquake data are spatially inversely transformed to obtain denoised earthquake data.

6. The method for denoising seismic data according to claim 1, characterized in that, In step S300, the error between the preset noiseless seismic data and the denoised seismic data is backpropagated to update the parameters of the convolutional autoencoder, including: S301, Calculate the error between the preset noiseless seismic data and the denoised seismic data; S302, the error is backpropagated to update the weights and biases of the convolutional autoencoder.

7. A noise reduction device for seismic data, characterized in that, The device includes: The preprocessing module is used to perform dimensionality expansion on the raw seismic data to obtain noisy seismic data that matches the convolutional autoencoder. The encoding / decoding module is used to input the noisy seismic data into a convolutional autoencoder to denoise the noisy seismic data and obtain denoised seismic data. The backpropagation module is used to backpropagate the error between the preset noiseless seismic data and the denoised seismic data, and update the parameters of the convolutional autoencoder. The denoising module is used to input the seismic data to be denoised into the backpropagation convolutional autoencoder to denoise the seismic data.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method for denoising seismic data according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method for denoising seismic data as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method for denoising seismic data as described in any one of claims 1 to 6.

Images