High-dimensional seismic data denoising method and device

Abstract

The invention discloses a high-dimensional seismic data denoising method and device. The method comprises the following steps: segmenting a high-dimensional seismic data volume into a plurality of data blocks by using a sliding window; taking the data block with the corresponding de-noised data as a sample to obtain a sample set, and training a Transform model to obtain a de-noising model; and inputting a plurality of data blocks obtained by segmenting the high-dimensional seismic data volume into a denoising model, and obtaining a denoised seismic data volume of the high-dimensional seismic data volume according to an output result of the model. According to the method, intelligent suppression of high-dimensional seismic data noise can be realized by using high-dimensional features of seismic data based on a deep neural network Transform model, and the denoising precision is high.

Description

Technical Field

[0001] This invention relates to the field of seismic data processing technology in oil and gas geophysical exploration, and particularly to a method and apparatus for denoising high-dimensional seismic data. Background Technology

[0002] Due to the complex underground geological structure and surface conditions, oil and gas exploration in my country faces numerous challenges, including difficulties in construction organization, low signal-to-noise ratios in data, and difficulties in seismic imaging and interpretation. How to accurately determine the structure of underground media using seismic exploration is a constant pursuit for researchers, especially in today's era of high oil extraction costs and volatile oil prices. How to effectively overcome the bottlenecks of traditional geophysical data processing methods, identify noise, interpret data clearly, and make full use of data to effectively reduce the difficulty and risk of exploration and development, improve exploration accuracy and oil and gas recovery rates, and control development costs are pressing issues that the petroleum industry urgently needs to address.

[0003] Seismic data contains various types of noise that easily couple with valid signals. Identifying and separating noise while preserving valid signals has been a long-standing concern in the field of exploration geophysics. Seismic data denoising, as a key step in the preprocessing of datasets for oil and gas resource exploration, is crucial for improving the resolution of seismic imaging. Traditional seismic exploration data denoising techniques are based on signal processing and have encountered several bottlenecks: 1. The volume of exploration data is exploding, making it difficult for traditional techniques to uncover the inherent connections within large seismic datasets. 2. Over-reliance on signal and noise model assumptions makes it difficult for existing algorithms to effectively denoise complex noise types. 3. Traditional algorithms rely on professional parameter adjustments, limiting the automation of data processing. The exploration industry urgently needs to research "intelligent denoising," utilizing artificial intelligence and deep learning technologies to achieve breakthroughs in the efficiency, effectiveness, and automation of exploration data denoising.

[0004] In seismic exploration data processing, seismic data is typically processed by extracting common shot points, receiver points, or offset gathers based on shot points, receiver points, or offsets. Mathematically, the extracted common shot gathers (or common receiver points, common offset gathers) are a subset of the original seismic data; physically, this represents a dimensionality reduction of the original data (in 3D exploration, 5D seismic data becomes 3D). This prevents the full utilization of the correlation features of seismic data in high-dimensional space. For example, processing based on common shot gathers cannot consider the correlation between adjacent shot data, which to some extent disrupts the structure of the original data. Fully utilizing the correlation characteristics of different azimuths and offsets in high-dimensional seismic data is the physical basis for the superiority of high-dimensional seismic data processing (such as denoising and interpolation reconstruction) over low-dimensional cases. While 5D seismic data interpolation based on traditional methods has a long history of research, research on 5D data denoising based on traditional methods is relatively scarce. The problems and bottlenecks encountered by traditional methods are generally clear, but 5D intelligent denoising based on deep learning is still largely unexplored territory. Summary of the Invention

[0005] To enrich product types, process routes, and options, this invention provides a method and apparatus for denoising high-dimensional seismic data, achieving intelligent suppression of noise in high-dimensional seismic data.

[0006] In a first aspect, embodiments of the present invention provide a method for denoising high-dimensional seismic data, comprising:

[0007] Use a sliding window to divide the high-dimensional seismic data volume into multiple data blocks;

[0008] The data block with corresponding denoised data is taken as a sample to obtain the sample set;

[0009] The Transformer model is trained using the sample set to obtain the denoising model;

[0010] The multiple data blocks are input into the denoising model, and the denoised seismic data volume of the high-dimensional seismic data volume is obtained based on the output of the model.

[0011] Optionally, before training the Transformer model using the sample set, the method further includes:

[0012] The seismic traces contained in some of the obtained data blocks are processed by assigning zeros to their waveform data. Then, the multiple data blocks are input into the Transformer model, and self-supervised pre-training is performed by interpolating the waveform data of the zero-assigned seismic traces.

[0013] Optionally, the method of using a sliding window to divide the high-dimensional seismic data volume into multiple data blocks includes:

[0014] Set the window size and sliding step size, where the side length of the window is greater than the step size;

[0015] In the shot point domain or receiver point domain of the high-dimensional seismic data volume, a sliding window is used to cut a data block for each current position of the window, and finally multiple data blocks are obtained.

[0016] Optionally, after setting the window size and sliding step, the method further includes:

[0017] The number of seismic traces contained in each data block is determined based on the window size; correspondingly,

[0018] After the data block is obtained by cutting, the process also includes:

[0019] Determine if the number of seismic traces in the data block is less than the determined number of seismic traces. If so, pad the data block with seismic traces by assigning 0 to the seismic waveform data.

[0020] Optionally, the high-dimensional seismic data volume is a five-dimensional seismic data volume containing four spatial dimensions and one time dimension;

[0021] The four spatial dimensions include the two-dimensional coordinates, offset, and azimuth of the gun point and receiver point.

[0022] Optionally, the Transformer model includes an input layer, an encoder, a decoder, and an output layer connected in sequence;

[0023] The input layer is used to perform a one-dimensional convolution operation on the input data, converting it into a vector of fixed dimensions;

[0024] The encoder contains multiple self-attention modules. Each self-attention module calculates the similarity of the query, key, and value matrices of the input data to capture the global dependencies of high-dimensional seismic data.

[0025] The decoder is used to recover signals from seismic data and suppress noise by applying a self-attention mechanism to the temporal and spatial dimensions.

[0026] The output layer is used to pass the output of the decoder to a one-dimensional convolutional layer to generate denoised seismic data.

[0027] Optionally, the Transformer model further includes a position encoding layer between the input layer and the encoder. This layer determines the position encoding of the data block based on the two-dimensional coordinates of the shot and receiver points contained within the data block, using sine and cosine functions. The position encoding is used to determine the relative position information of the data block. Correspondingly,

[0028] The decoder is used to recover signals from seismic data and suppress noise by applying a self-attention mechanism to the time and space dimensions and combining it with location encoding.

[0029] Optionally, the encoder employs a multi-head attention mechanism, where each head learns different feature patterns from the data.

[0030] Optionally, the decoder employs residual connections and layer normalization techniques;

[0031] The decoder includes multiple self-attention modules.

[0032] Optionally, before dividing the high-dimensional seismic data volume into multiple data blocks using a sliding window, the method further includes:

[0033] The waveform data of the high-dimensional seismic data volume is normalized.

[0034] Secondly, embodiments of the present invention provide a high-dimensional seismic data denoising device, comprising:

[0035] The data segmentation module is used to segment high-dimensional seismic data volumes into multiple data blocks using a sliding window.

[0036] The sample set creation module is used to take a data block with corresponding denoised data as a sample to obtain a sample set;

[0037] The model training module is used to train the Transformer model using the sample set to obtain a denoising model.

[0038] The seismic data denoising module is used to input the multiple data blocks into the denoising model and obtain the denoised seismic data volume of the high-dimensional seismic data volume based on the output of the model.

[0039] Thirdly, embodiments of the present invention provide a computer storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described high-dimensional seismic data denoising method.

[0040] Fourthly, this disclosure provides a server, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described high-dimensional seismic data denoising method.

[0041] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following:

[0042] (1) The high-dimensional seismic data denoising method provided in this embodiment of the invention targets a high-dimensional seismic data volume that has undergone denoising processing in some blocks. First, the high-dimensional seismic data volume is divided into multiple data blocks; the data blocks with corresponding denoised data are used as samples to obtain a sample set, and a Transformer model is trained to obtain a denoising model; then, the multiple data blocks obtained by cutting the high-dimensional seismic data volume are input into the denoising model, and the denoised seismic data volume is obtained according to the output of the model. This method realizes intelligent suppression of noise in high-dimensional seismic data and has high denoising accuracy; it is of great significance to the intelligent development of geophysics. In actual exploration tasks, it can improve work efficiency, shorten work cycle, reduce labor costs, reduce dependence on human experience, enhance the reliability of data-driven analysis, improve the ability to solve complex problems, and enhance application effects.

[0043] (2) The high-dimensional seismic data denoising method provided in this embodiment of the invention uses a sliding window to divide the high-dimensional seismic data volume into multiple data blocks. The advantage of this segmentation method is that it maintains the structure of the high-dimensional data while making full use of hardware resources, improving the computational efficiency and denoising effect of the model. When processing each data block, the model can make full use of the global correlation between different spatial dimensions, thereby improving the accuracy of denoising. Compared with traditional two-dimensional and three-dimensional seismic data segmentation methods, this segmentation method extracts richer spatial and temporal information through a multi-dimensional sliding window. Each data block not only contains short-term information in the time series, but also covers the multi-dimensional long-distance dependencies within the entire window, further improving the performance of the Transformer model in the denoising task.

[0044] (3) In order to ensure the learning ability of the model, the high-dimensional seismic data denoising method provided in this embodiment of the invention has a sliding window side length greater than the sliding step size, so that the adjacent data blocks obtained by cutting overlap. This overlap helps the model to better capture the correlation between data blocks and improve the overall denoising effect.

[0045] (4) The high-dimensional seismic data denoising method provided in this embodiment of the invention performs self-supervised pre-training without noise labels, and performs interpolation training after randomly occluding part of the data. The model successfully recovers the occluded data. After pre-training, the model performs more stably and converges faster in subsequent denoising tasks.

[0046] (5) The high-dimensional seismic data denoising method provided in this embodiment of the invention, after adding position encoding, enables the model to better capture the correlation between spatial dimensions, thereby ensuring that the model can make full use of its spatial structure information and improve the denoising accuracy when processing five-dimensional seismic data.

[0047] (6) The high-dimensional seismic data denoising method provided in this embodiment of the invention, Transformer, can effectively capture complex signal patterns in seismic data by utilizing a self-attention mechanism, which is crucial for analyzing seismic data with time and space dimensions. This mechanism enables the model to comprehensively consider the entire dataset and identify key factors affecting signal quality. Different attention "heads" can focus on different aspects or features in the sequence. For example, in seismic data processing, some "heads" may focus on capturing the propagation velocity features of seismic waves, while other "heads" may focus on changes in amplitude or frequency. This design enables the model to capture diverse features, allowing the model to more accurately perform noise suppression and signal recognition tasks when processing complex signals such as seismic data.

[0048] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings.

[0049] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0050] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0051] Figure 1 This is a flowchart of the high-dimensional seismic data denoising method in Embodiment 1 of the present invention;

[0052] Figure 2 This is an example diagram of a sliding window in the shot point domain in Embodiment 1 of the present invention;

[0053] Figure 3 This is an example diagram of a sliding window in the receiving point domain in Embodiment 1 of the present invention;

[0054] Figure 4 This is a schematic diagram of the multi-head self-attention mechanism in Embodiment 1 of the present invention;

[0055] Figure 5 This is a comparison chart of the denoising results of the traditional method and the Transformer method in Embodiment 1 of the present invention;

[0056] Figure 6a This is a position encoding heatmap for the 2D scenario in Embodiment 1 of the present invention;

[0057] Figure 6b This is a position encoding heatmap for the 4D scenario in Embodiment 1 of the present invention;

[0058] Figure 7 This is a flowchart of the overall denoising process using the five-dimensional Transformer network in Embodiment 2 of the present invention;

[0059] Figure 8 This is a demonstration of the results of self-supervised pre-training in Embodiment 2 of the present invention;

[0060] Figure 9 This is a demonstration of the denoising results of 2D seismic data in Embodiment 2 of the present invention;

[0061] Figure 10 This is a demonstration of the denoising results of 5D seismic data in Embodiment 2 of the present invention;

[0062] Figure 11 This is a comparison of the denoising results with and without positional coding in Embodiment 2 of the present invention;

[0063] Figure 12 This is a demonstration of the actual OVT domain data denoising results in Embodiment 2 of the present invention;

[0064] Figure 13 This is a schematic diagram of the high-dimensional seismic data denoising device in an embodiment of the present invention. Detailed Implementation

[0065] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0066] It should be understood that the terminology used in this invention is merely for describing particular embodiments and is not intended to limit the invention. Furthermore, with respect to numerical ranges in this invention, it should be understood that each intermediate value between the upper and lower limits of the range is also specifically disclosed. Every smaller range between any stated value or intermediate value within a stated range, and any other stated value or intermediate value within said range, is also included in this invention. The upper and lower limits of these smaller ranges may be independently included or excluded from the range.

[0067] Unless otherwise stated, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. While only preferred methods and materials have been described herein, any methods and materials similar or equivalent to those described herein may be used in the implementation or testing of this invention. All references to this specification are incorporated by way of citation to disclose and describe methods and / or materials associated with those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

[0068] To address the problems of lack of high-dimensional denoising network frameworks, difficulty in effectively utilizing high-dimensional information, and difficulty in achieving intelligent denoising of high-dimensional seismic data, this invention proposes a high-dimensional seismic data denoising method and apparatus that can utilize the high-dimensional characteristics of seismic data to achieve intelligent suppression of seismic noise with high denoising accuracy.

[0069] First, we will introduce the problem of seismic data denoising under high-dimensional spatial convolutional network models.

[0070] The denoising model using deep neural networks can be described by the following model. Noisy data can generally be described as follows:

[0071] y = x + n (1)

[0072] Here, y represents noisy data, x represents clean data, and n is the noise included in the data, which can be random noise or coherent noise. The general idea of ​​the denoising problem is to recover clean data x from noisy data y. If we introduce deep neural networks into this problem, to express this general idea:

[0073]

[0074] in, Let represent a deep neural network with optimal parameters θ. Determining these parameters can be expressed as an optimization problem:

[0075] θ0=argmin θ [E(f θ (y);x)+R(y)] (3)

[0076] Here, function E measures the distance between two objects. Typically, the mean squared error (MSE) loss is used as this function E. R(y) is a regularization term that can be manually chosen and is usually already included in the neural network. In denoising problems, a training strategy called residual learning has proven effective (Zhang et al., 2017), which means:

[0077]

[0078] Transform the above formula into:

[0079] θ0=argmin θ [E(f θ (y);yx)+R(y)] (5)

[0080] Using a training set containing noisy data y and its corresponding clean label x, the neural network model can be trained under supervision according to the above formula.

[0081] Example 1

[0082] Embodiment 1 of the present invention provides a method for denoising high-dimensional seismic data, the process of which is as follows: Figure 1 As shown, it includes the following steps:

[0083] Step S11: Use a sliding window to divide the high-dimensional seismic data volume into multiple data blocks.

[0084] In this embodiment of the invention, the high-dimensional seismic data volume is a five-dimensional seismic data volume containing four spatial dimensions and one time dimension. The four spatial dimensions include the two-dimensional spatial coordinates (i.e., x and y coordinates) of the shot point and receiver point, the offset, and the azimuth.

[0085] First, set the window size and sliding step size. To ensure that adjacent data blocks have overlapping parts, the window side length needs to be set to be greater than the step size.

[0086] In the shot point domain or receiver point domain of a high-dimensional seismic data volume, a sliding window with a pre-set step size is used to cut a data block at each current position of the window, ultimately resulting in multiple data blocks. See also Figure 2 and Figure 3 The figures shown are examples of sliding windows in the shot point domain and receiver point domain, respectively.

[0087] Furthermore, the number of seismic traces contained in each data block is determined based on the window size. After a data block is obtained, it is checked whether the number of seismic traces contained in the data block is less than the determined number of seismic traces. If so, the seismic traces in the data block are padded with 0s by assigning 0s to the seismic waveform data. This ensures the consistency of the data structure of each data block.

[0088] To address the denoising issue in five-dimensional seismic data, a seismic data extraction and sample generation method based on SEGY files was used. Each track in the file contains the seismic waveform and its associated geographic information (such as the coordinates of receivers and shot points).

[0089] During data reading, the function extracts the seismic waveform data for each trace (if enabled) and its corresponding header information, including the x and y coordinates of the receiver and shot points. Each data block contains the seismic waveform, the coordinates of the receiver and shot points, the trace number, and the time sampling interval.

[0090] Seismic waveform data: The actual seismic signals acquired.

[0091] Coordinates of receiver and shot points: including x and y coordinates, used to describe the geographical distribution of the samples.

[0092] Track number: A unique identifier for each track, used to locate and track the source of data.

[0093] Time sampling interval: describes the temporal resolution of each data point, facilitating subsequent time series analysis.

[0094] It's important to note that the coordinate information read here refers to the x and y coordinates of the receiver and shot points, rather than the azimuth and offset of typical five-dimensional seismic data. This is necessary for the subsequent introduction of positional encoding in five-dimensional coordinates. In fact, these two methods of representing five-dimensional coordinates can be converted one-to-one within the same coordinate system.

[0095] Five-dimensional seismic data files are typically in the GB range. To address the challenge of data volumes that are too large to be directly input into models, a segmentation operation is used to rationally divide the data into blocks, facilitating efficient input into the Transformer model. Within a sliding window, seismic waveform records between all source and receiver points are extracted; these records include time-dimension information. Each data block within a window maintains the integrity of the four-dimensional spatial information while including the time dimension.

[0096] Step S12: Take the data block with corresponding denoised data as a sample to obtain the sample set.

[0097] Each sample includes noisy seismic waveform data, the corresponding denoised seismic waveform data, relevant spatial location information (such as the x and y coordinates of the receiver and shot points), and the temporal sampling interval. This sample generation method ensures that the training data has sufficient spatial and temporal diversity, making it suitable for deep learning denoising tasks on five-dimensional seismic data.

[0098] This dataset construction method is applicable to different seismic zones. By simply changing the input actual seismic data, it can automatically extract noisy samples and generate denoised sample pairs. This portability ensures that the method in this embodiment can generate high-quality training datasets under different seismic environments and data conditions, and has broad application potential.

[0099] Step S13: Train the Transformer model using the sample set to obtain the denoising model.

[0100] The Transformer architecture is an innovation in deep learning, initially achieving significant breakthroughs in natural language processing. The core idea of ​​the Transformer is its self-attention mechanism, which allows the model to consider information from other elements while processing each element of a sequence, unlike traditional recurrent network structures. The advantage of self-attention lies in its ability to capture long-range dependencies in a sequence and its extreme suitability for parallel computation, thus significantly improving processing efficiency.

[0101] The goal of self-attention is to learn a representation for each element in a sequence that can be associated with other elements in the sequence. The core concept of self-attention is to transform each input element into a set of "keys," "values," and "queries," which are vector representations obtained by multiplying by a weight matrix. For a given input sequence (such as a time-series sample from seismic data), where each is a feature vector, query (Q), key (K), and value (V) matrices can be computed:

[0102] Q = XW Q K = XW K V = XW V (4)

[0103] Here W Q W K and W V It is a trainable weight matrix. Once Q, K, and V are available, the self-attention mechanism will compute its compatibility with the keys at all positions in the sequence. This is achieved through "dot products":

[0104]

[0105] The softmax function here ensures that the weights are positive and sum to 1, and the dot product result is divided by... (d K The dimension of the key / query vector helps stabilize the gradient. This means that values ​​have a greater impact at locations with higher attention weights. A key advantage of self-attention mechanisms is their parallel performance, as attention scores and outputs can be computed simultaneously, unlike recursive structures that require waiting for the previous time step to complete. Furthermore, self-attention mechanisms can capture long-range dependencies, making them crucial in time-series data processing tasks such as seismic data denoising.

[0106] Another core concept in Transformer is "Multi-Head Attention," which allows the model to learn information from different locations simultaneously, improving its ability to learn complex patterns. See also Figure 4 As shown, in a multi-head attention setup, the self-attention calculation described above is performed in parallel multiple times, each time using a different set of weight matrices:

[0107] MiltiHead(Q,K,V)=Concat(head1,...,head h W O (6)

[0108] Among them, head i =Attention(QW i QKW i K VW i V Each "head" captures different information, which is then concatenated and multiplied by another weight matrix W. O This yields the final output. In this way, the Transformer can process data in a highly parallelized manner.

[0109] In the field of seismic data denoising, the advantages of Transformer are obvious. First, by utilizing a self-attention mechanism, Transformer can effectively capture complex signal patterns in seismic data, which is crucial for analyzing seismic data with temporal and spatial dimensions. This mechanism allows the model to comprehensively consider the entire dataset and identify key factors affecting signal quality. Second, thanks to Transformer's parallel processing capabilities, it can quickly output results when processing large-scale seismic datasets. This processing method outperforms convolutional neural network methods in both efficiency and accuracy when analyzing long-term series data. Furthermore, Transformer can identify and remember important feature patterns in seismic signals and effectively remove noise using these patterns.

[0110] Figure 5 This is a comparison of some denoising results. The five-dimensional seismic data denoising model in this embodiment is based on the Transformer framework. Its core idea is to use a self-attention mechanism to process five-dimensional high-dimensional data (4 spatial dimensions and 1 temporal dimension). Unlike traditional convolutional neural networks (CNNs) and recurrent neural networks (RNNs), the Transformer can not only capture long-range dependencies but also has efficient parallel processing capabilities, giving it a significant advantage in processing large-scale, complex seismic data. Figure 5 ).

[0111] The Transformer model consists of an input layer, a positional encoding layer, an encoder, a decoder, and an output layer connected in sequence.

[0112] The 5D denoising Transformer uses raw 5D seismic traces as input, which contain four spatial dimensions and one temporal dimension. These traces are extracted from 5D seismic data, including their location coordinates. The data is first embedded into a sequence of specific dimensions through 1D convolutional layers, then 5D location encodings are added separately before being fed into the Transformer encoder and decoder. The parameters of the Transformer blocks in the encoder and decoder are continuously updated until the model's output matches the corresponding label.

[0113] (1) Input layer, used to perform one-dimensional convolution operation on the input data and convert it into a vector of fixed dimensions.

[0114] Due to the high dimensionality of seismic data, the input layer embeds the segmented five-dimensional data blocks using 1D convolutional layers. This process transforms the input time series and spatial coordinates into fixed-dimensional vector representations by performing a one-dimensional convolution operation on the seismic waveform data. These vector representations are then fed into other network layers of the Transformer model for further processing.

[0115] (2) Location coding layer

[0116] The self-attention mechanism of the Transformer model does not contain built-in sequence order information. Therefore, in some embodiments, the Transformer model also sets up a position encoding layer between the input layer and the encoder. This layer is used to determine the position encoding of the data block based on the two-dimensional spatial coordinates of the shot point and receiver point contained in the data block, using the sin and cos functions. The position encoding is used to determine the relative position information of the data block.

[0117] Traditional Transformer positional encoding is typically used to process one-dimensional sequences (such as word sequences in natural language processing), but this embodiment extends it to four-dimensional spatial coordinates. Specifically, based on the relative positions of the source and receiver points, continuous positional codes are generated and added to the embedded data to ensure that the model can preserve and utilize spatial information.

[0118] By incorporating location encoding, the model can better capture the correlation between spatial dimensions, thereby ensuring that the model can fully utilize the spatial structure information when processing five-dimensional seismic data and improve denoising accuracy.

[0119] To facilitate accurate spatial description of position information using the sine and cosine functions, the input position information here consists of four coordinates: the x and y coordinates of the shot point and the detector, rather than the traditional azimuth and offset forms. In reality, these two representations correspond one-to-one in space.

[0120] The sequences input into the Transformer block are all equivalent, regardless of their initial positions.

[0121] Therefore, adding position encoding can be used to recover this position information. In the 1D case, sin-cos position encoding can be defined as:

[0122]

[0123] Its purpose is to represent the location of seismic traces within the original data block. Here, i represents the data location, j represents the encoded location, and L represents the total encoded length. Considering that spatial coordinates are usually not points exactly on the grid, the location encoding should be extended to a continuous form:

[0124]

[0125] Where x represents the relative location of the study area.

[0126]

[0127] In the 2D case, if the total encoding length remains L, the encoding length for each dimension is reduced to L / 2. If the dimension to be encoded is expanded to 4, the encoding length for each dimension will be reduced to L / 4, as shown in Figure 6. The horizontal axis represents the encoding position, and the vertical axis represents the position index of the input sequence. The length of the horizontal axis represents the length of the positional encoding, and the length of the vertical axis represents the total number of input sequences.

[0128] (3) The encoder consists of multiple self-attention modules. Each self-attention module calculates the similarity of the query, key, and value matrices of the input data to capture the global dependencies of high-dimensional seismic data. Through this mechanism, the model can extract useful information from all dimensions of the input data (including time and space dimensions).

[0129] Furthermore, the encoder employs a multi-head attention mechanism, where each head learns different feature patterns in the data. For example, some attention heads might focus on frequency variations in seismic waves, while others might focus on amplitude or propagation velocity characteristics of the waveform. Ultimately, the outputs of these heads are concatenated and fed into the next layer.

[0130] (4) The decoder is used to recover signals in seismic data and suppress noise by applying a self-attention mechanism to the time and space dimensions.

[0131] Furthermore, the decoder is used to recover signals in seismic data and suppress noise by applying a self-attention mechanism to the temporal and spatial dimensions, combined with location encoding.

[0132] The decoder employs residual connections and layer normalization techniques. These techniques help stabilize the model's training process and accelerate convergence. Residual connections add the output of each layer to the input, ensuring that information is not lost during transmission, while layer normalization improves the model's training stability by adjusting the distribution of activation values.

[0133] The decoder also includes multiple self-attention modules, the purpose of which is to combine the information from the encoder output to generate denoised seismic data. The decoder recovers and restores the signal in the seismic data and suppresses noise by applying the self-attention mechanism to the temporal and spatial dimensions and combining it with location encoding.

[0134] (5) The output layer is used to pass the decoder output to the one-dimensional convolutional layer to generate denoised seismic data. Through this process, the model can not only remove noise from the input data, but also retain and recover the original seismic signal.

[0135] Mean squared error (MSE) was used as the loss function to measure the difference between the denoised model output and the actual noise-free data. Furthermore, considering the unique characteristics of seismic data, the signal-to-noise ratio (SNR) was also introduced as an evaluation metric to further optimize the model's performance.

[0136] During training, numerous experiments were conducted with different loss functions, optimizers, and learning rates to systematically compare the impact of different parameter settings on model convergence speed and denoising performance. Ultimately, the Adam optimizer was selected, and the learning rate was dynamically adjusted under different noise levels to ensure rapid and stable model convergence.

[0137] Due to the significant regional variations in earthquake data, a fine-tuning strategy using transfer learning was adopted. First, the model was pre-trained on simulated data, and then fine-tuned using data from the actual survey area, adjusting the decoder weights to ensure the model could adapt to the new data distribution.

[0138] Step S14: Input multiple data blocks into the denoising model, and obtain the denoised seismic data volume of the high-dimensional seismic data volume based on the output of the model.

[0139] The high-dimensional seismic data denoising method provided in Embodiment 1 of this invention targets a high-dimensional seismic data volume that has undergone denoising processing in certain blocks. First, the high-dimensional seismic data volume is divided into multiple data blocks. Each data block with corresponding denoised data is used as a sample to obtain a sample set, which is then used to train a Transformer model to obtain a denoising model. Then, the multiple data blocks obtained from cutting the high-dimensional seismic data volume are input into the denoising model, and the denoised seismic data volume is obtained based on the model's output. This method achieves intelligent suppression of seismic noise with high denoising accuracy. It has theoretical significance for the intelligent development of geophysics. In practical exploration tasks, it can improve work efficiency, shorten work cycles, reduce labor costs, reduce reliance on human experience, enhance the reliability of data-driven analysis, improve the ability to solve complex problems, and enhance application effects.

[0140] The high-dimensional seismic data denoising method provided in Embodiment 1 of this invention uses a sliding window to segment the high-dimensional seismic data volume into multiple data blocks. The advantage of this segmentation method is that it maintains the structure of the high-dimensional data while fully utilizing hardware resources, improving the computational efficiency and denoising effect of the model. When processing each data block, the model can fully utilize the global correlation between different spatial dimensions, thereby improving the accuracy of denoising. Compared with traditional two-dimensional and three-dimensional seismic data segmentation methods, this segmentation method extracts richer spatial and temporal information through a multi-dimensional sliding window. Each data block not only contains short-term information from the time series but also covers multi-dimensional long-distance dependencies within the entire window, further improving the performance of the Transformer model in denoising tasks.

[0141] To ensure the model's learning ability, the high-dimensional seismic data denoising method provided in Embodiment 1 of this invention uses a sliding window with a side length greater than the sliding step size, which causes the adjacent data blocks obtained from the cutting to partially overlap. This overlap helps the model better capture the correlation between data blocks and improves the overall denoising effect.

[0142] The high-dimensional seismic data denoising method provided in Embodiment 1 of this invention utilizes a self-attention mechanism in the Transformer to effectively capture complex signal patterns in seismic data. This is crucial for analyzing seismic data with temporal and spatial dimensions. This mechanism enables the model to comprehensively consider the entire dataset and identify key factors affecting signal quality. Different attention "heads" can focus on different aspects or features in the sequence. For example, in seismic data processing, some "heads" may focus on capturing the propagation velocity characteristics of seismic waves, while others may focus on amplitude or frequency changes. This design allows the model to capture diverse features, enabling it to more accurately perform noise suppression and signal recognition tasks when processing complex signals such as seismic data.

[0143] In some embodiments, before training the Transformer model using the sample set, the method further includes:

[0144] The seismic traces contained in some of the obtained data blocks are processed by assigning zeros to their waveform data. Then, the multiple data blocks are input into the Transformer model, and self-supervised pre-training is performed by interpolating the waveform data of the zero-assigned seismic traces.

[0145] Furthermore, setting the waveform data of seismic traces to 0 can be done by setting the waveform data of all seismic traces in the entire data block to 0, or by setting the waveform data of a portion of the seismic traces in a data block to 0.

[0146] Self-supervised pre-training was performed without noise labels. After randomly occluding a portion of the data, interpolation training was conducted, and the model successfully recovered the occluded data. After pre-training, the model performed more stably and converged faster in subsequent denoising tasks.

[0147] Furthermore, a Boolean variable can be used to control whether seismic waveform data is read. If set to False, corresponding zero data will be generated, thus achieving the assignment of zero to the seismic trace waveform data.

[0148] In some embodiments, before using a sliding window to divide a high-dimensional seismic data volume into multiple data blocks, the method further includes: normalizing the waveform data of the high-dimensional seismic data volume.

[0149] Example 2

[0150] Embodiment 2 of the present invention provides a specific application of a high-dimensional seismic data denoising method, see [link to documentation]. Figure 7 The diagram shown is the overall flowchart of the five-dimensional Transformer network for denoising.

[0151] The specific implementation steps of this embodiment include:

[0152] (1) Actual data extraction and sample generation.

[0153] (2) Training of deep learning networks based on self-attention mechanism.

[0154] (3) Input the data into the network model and suppress noise.

[0155] This embodiment demonstrates the denoising performance of the model in different experimental scenarios, including denoising results for 2D and 5D seismic data and the application of the model in real-world data. Through a series of comparative experiments, the effectiveness and superiority of the Transformer-based five-dimensional seismic data denoising model are verified.

[0156] 1. 2D seismic data denoising experiment

[0157] In experiments with 2D seismic data, self-supervised pre-training was first performed, followed by supervised fine-tuning training. By training on 2D seismic data with random noise, the model was able to effectively remove noise and recover the original signal.

[0158] Self-supervised pre-training was performed without noise labels. After randomly occluding a portion of the data, interpolation training was conducted, and the model successfully recovered the occluded data. After pre-training, the model performed more stably and converged faster in subsequent denoising tasks.

[0159] See Figure 8As shown in the figure, the comparison of the results of self-supervised training is as follows: from left to right, the original seismic data, the seismic data after random partial occlusion, the denoised seismic data obtained without pre-training, and the denoised seismic data obtained after pre-training.

[0160] Denoising results: The model successfully removed noise from the 2D seismic data, including random noise and some linear noise. See also Figure 9 As shown, from left to right, the data consists of the original data, the data with added random noise, the data denoised by the network, and the data with noise removed by the network. By comparing the data before and after the noise, it can be clearly seen that the data after denoising has recovered more structural information and the signal-to-noise ratio (SNR) has been significantly improved.

[0161] 2. 5D seismic data denoising experiment

[0162] After verifying the denoising effect on 2D data, the model was extended to 5D seismic data. Due to the complexity and massive scale of 5D data, segmentation and sliding window techniques were used to process the data, ensuring that the model could run effectively under hardware conditions.

[0163] In a 5D seismic data experiment, the denoising performance of Transformer-based and traditional DnCNN models was compared. See [link / reference]. Figure 10 As shown, the data are presented in sequence as follows: (a) original data, (b) data with added random noise, (c) denoising result of DnCNN network, (d) noise removed by DnCNN network, (e) denoising result of this embodiment, and (f) noise removed by this embodiment. The experimental results show that the Transformer model can not only significantly improve SNR, but also perform better in the recovery of multi-dimensional signals.

[0164] Impact of Position Encoding: This embodiment also investigates the role of position encoding in 5D data denoising. See [link to relevant documentation] Figure 11 As shown, the results are (a) denoising result without position coding, (d) noise removed without position coding, (c) denoising result with position coding, and (d) noise removed with position coding. The experimental results show that after adding position coding, the model can better capture the correlation between spatial dimensions, thereby improving the denoising accuracy.

[0165] 3. Actual data denoising experiment

[0166] In actual data testing, transfer learning was employed to apply the model pre-trained on simulated data to seismic data from the actual survey area. Because the noise types and data distribution of the actual data differ significantly from those of the simulated data, the model's decoder was fine-tuned to adapt to the new data characteristics.

[0167] OVT domain denoising results: Figure 12As shown, the data are presented in the following order: (a) noisy seismic data, (b) denoised data without using the self-supervised pre-trained network, (c) noise removed without using the self-supervised pre-trained network, (d) denoised data using the self-supervised pre-trained network, and (e) noise removed using the self-supervised pre-trained network. The fine-tuned model successfully removed a significant amount of noise from the OVT domain seismic data, significantly improving the signal-to-noise ratio. The model demonstrates strong adaptability when handling data with different noise levels, effectively removing both coherent and random noise.

[0168] Based on the inventive concept of this invention, embodiments of this invention also provide a high-dimensional seismic data denoising device, the structure of which is as follows: Figure 13 As shown, it includes:

[0169] Data segmentation module 131 is used to segment a high-dimensional seismic data volume into multiple data blocks using a sliding window;

[0170] The sample set establishment module 132 is used to take a data block with corresponding denoised data as a sample to obtain a sample set;

[0171] Model training module 133 is used to train the Transformer model using the sample set to obtain a denoised model;

[0172] The seismic data denoising module 134 is used to input the multiple data blocks into the denoising model and obtain the denoised seismic data volume of the high-dimensional seismic data volume based on the output of the model.

[0173] In some embodiments, the above-described apparatus further includes a model pre-training module 135, which is used to perform waveform data zeroing on the seismic traces contained in a portion of the obtained plurality of data blocks before the model training module 133 trains the Transformer model using the sample set, and then inputs the plurality of data blocks into the Transformer model to perform self-supervised pre-training by interpolating the waveform data of the zero-set seismic traces.

[0174] In some embodiments, the data segmentation module 131, which uses a sliding window to segment the high-dimensional seismic data volume into multiple data blocks, is used for:

[0175] Set the window size and sliding step size, with the window side length being greater than the step size; in the shot point domain or receiver point domain of the high-dimensional seismic data volume, slide the window according to the step size, and for each current position of the window, cut to obtain a data block, ultimately obtaining multiple data blocks.

[0176] In some embodiments, after setting the window size and sliding step size, the data segmentation module 131 is further configured to:

[0177] The number of seismic traces contained in each data block is determined based on the window size; correspondingly,

[0178] The data segmentation module 131, after obtaining a data block through segmentation, is further used for:

[0179] Determine if the number of seismic traces in the data block is less than the determined number of seismic traces. If so, pad the data block with seismic traces by assigning 0 to the seismic waveform data.

[0180] In some embodiments, the above-described apparatus further includes a normalization processing module 136, configured to:

[0181] Before the data segmentation module 131 uses a sliding window to segment the high-dimensional seismic data volume into multiple data blocks, it performs normalization processing on the waveform data of the high-dimensional seismic data volume.

[0182] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0183] Based on the inventive concept of the present invention, embodiments of the present invention also provide a computer storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described high-dimensional seismic data denoising method.

[0184] Based on the inventive concept of the present invention, embodiments of the present invention also provide a server, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described high-dimensional seismic data denoising method.

[0185] Unless otherwise specifically stated, terms such as processing, calculation, operation, determination, display, etc., may refer to the actions and / or processes of one or more processing or computing systems or similar devices that represent the manipulation and conversion of data representing physical (e.g., electronic) quantities within the registers or memory of the processing system into other data similarly representing physical quantities within the memory, registers, or other such information storage, transmission, or display devices of the processing system. Information and signals can be represented using any of a variety of different techniques and methods. For example, data, instructions, commands, information, signals, bits, symbols, and chips mentioned throughout the above description can be represented by voltage, current, electromagnetic waves, magnetic fields or particles, light fields or particles, or any combination thereof.

[0186] It should be understood that the specific order or hierarchy of steps in the disclosed process is an example of an exemplary method. Based on design preferences, it should be understood that the specific order or hierarchy of steps in the process may be rearranged without departing from the scope of this disclosure. The appended method claims provide elements of various steps in an exemplary order and are not intended to limit the scope to the specific order or hierarchy described.

[0187] In the detailed description above, various features are combined together in a single embodiment to simplify this disclosure. This approach to disclosure should not be construed as reflecting an intention that embodiments of the claimed subject matter require more features than those stated in each claim. Rather, as reflected in the appended claims, the invention is presented with fewer features than all of the features in a single disclosed embodiment. Therefore, the appended claims are hereby clearly incorporated into the detailed description, wherein each claim stands alone as a preferred embodiment of the invention.

[0188] Those skilled in the art will also understand that the various illustrative logic blocks, modules, circuits, and algorithm steps described in conjunction with the embodiments herein can be implemented as electronic hardware, computer software, or a combination thereof. To clearly illustrate the interchangeability between hardware and software, the various illustrative components, blocks, modules, circuits, and steps described above are generally described in terms of their functionality. Whether such functionality is implemented as hardware or software depends on the specific application and the design constraints imposed on the overall system. Those skilled in the art can implement the described functionality in alternative ways for each specific application; however, such implementation decisions should not be construed as departing from the scope of this disclosure.

[0189] The steps of the methods or algorithms described in conjunction with the embodiments herein can be directly embodied in hardware, software modules executed by a processor, or a combination thereof. The software modules can reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium well known in the art. An exemplary storage medium is connected to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and storage medium can reside in an ASIC. The ASIC can reside in a user terminal. Alternatively, the processor and storage medium can exist as discrete components in the user terminal.

[0190] For software implementation, the techniques described in this application can be implemented using modules (e.g., procedures, functions, etc.) that perform the functions described in this application. This software code can be stored in memory units and executed by a processor. The memory units can be implemented within the processor or outside the processor; in the latter case, they are communicatively coupled to the processor via various means, as is well known in the art.

[0191] The foregoing description includes examples of one or more embodiments. It is certainly impossible to describe all possible combinations of components or methods in order to describe the above embodiments, but those skilled in the art will recognize that further combinations and arrangements of the various embodiments are possible. Therefore, the embodiments described herein are intended to cover all such changes, modifications, and variations that fall within the scope of the appended claims. Furthermore, the term "comprising" as used in the specification or claims is interpreted in a manner similar to the term "including," as interpreted when used as a conjunction in the claims. Additionally, the use of any term "or" in the specification of the claims is intended to mean "non-exclusive or."

Claims

1. A method for high-dimensional seismic data denoising, characterized in that, include: Use a sliding window to divide the high-dimensional seismic data volume into multiple data blocks; The data block with corresponding denoised data is taken as a sample to obtain the sample set; The Transformer model is trained using the sample set to obtain the denoising model; The multiple data blocks are input into the denoising model, and the denoised seismic data volume of the high-dimensional seismic data volume is obtained based on the output of the model.

2. The method as described in claim 1, characterized in that, Before training the Transformer model using the aforementioned sample set, the following steps are also included: The seismic traces contained in some of the obtained data blocks are processed by assigning zeros to their waveform data. Then, the multiple data blocks are input into the Transformer model, and self-supervised pre-training is performed by interpolating the waveform data of the zero-assigned seismic traces.

3. The method as described in claim 1, characterized in that, The method of using a sliding window to divide a high-dimensional seismic data volume into multiple data blocks includes: Set the window size and sliding step size, where the side length of the window is greater than the step size; In the shot point domain or receiver point domain of the high-dimensional seismic data volume, a sliding window is used to cut a data block for each current position of the window, and finally multiple data blocks are obtained.

4. The method as described in claim 3, characterized in that, After setting the window size and sliding step, the following is also included: The number of seismic traces contained in each data block is determined based on the window size; correspondingly, After the data block is obtained by cutting, the process also includes: Determine if the number of seismic traces in the data block is less than the determined number of seismic traces. If so, pad the data block with seismic traces by assigning 0 to the seismic waveform data.

5. The method as described in claim 1, characterized in that, The high-dimensional seismic data volume is a five-dimensional seismic data volume containing four spatial dimensions and one time dimension; The four spatial dimensions include the two-dimensional coordinates, offset, and azimuth of the gun point and receiver point.

6. The method as described in claim 1, characterized in that, The Transformer model includes an input layer, an encoder, a decoder, and an output layer connected in sequence. The input layer is used to perform a one-dimensional convolution operation on the input data, converting it into a vector of fixed dimensions; The encoder contains multiple self-attention modules. Each self-attention module calculates the similarity of the query, key, and value matrices of the input data to capture the global dependencies of high-dimensional seismic data. The decoder is used to recover signals from seismic data and suppress noise by applying a self-attention mechanism to the temporal and spatial dimensions. The output layer is used to pass the output of the decoder to a one-dimensional convolutional layer to generate denoised seismic data.

7. The method as described in claim 6, characterized in that, The Transformer model further includes a position encoding layer between the input layer and the encoder. This layer determines the position encoding of the data block based on the two-dimensional coordinates of the shot and receiver points contained within the data block, using sine and cosine functions. The position encoding is used to determine the relative position information of the data block. Correspondingly, The decoder is used to recover signals from seismic data and suppress noise by applying a self-attention mechanism to the time and space dimensions and combining it with location encoding.

8. The method as described in claim 6, characterized in that, The encoder employs a multi-head attention mechanism, where each head learns different feature patterns from the data.

9. The method as described in claim 6, characterized in that, The decoder employs residual connection and layer normalization techniques. The decoder includes multiple self-attention modules.

10. The method according to any one of claims 1 to 9, characterized in that, Before using a sliding window to divide the high-dimensional seismic data volume into multiple data blocks, the method also includes: The waveform data of the high-dimensional seismic data volume is normalized.

11. A high-dimensional seismic data denoising device, characterized in that, include: The data segmentation module is used to segment high-dimensional seismic data volumes into multiple data blocks using a sliding window. The sample set creation module is used to take a data block with corresponding denoised data as a sample to obtain a sample set; The model training module is used to train the Transformer model using the sample set to obtain a denoising model. The seismic data denoising module is used to input the multiple data blocks into the denoising model and obtain the denoised seismic data volume of the high-dimensional seismic data volume based on the output of the model.

12. A computer storage medium, characterized in that, The computer storage medium stores computer-executable instructions, which, when executed by a processor, implement the high-dimensional seismic data denoising method according to any one of claims 1 to 10.

13. A server, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the high-dimensional seismic data denoising method according to any one of claims 1 to 10.

Images