Seismic data five-dimensional interpolation method based on deep learning

Abstract

The application discloses a kind of data five-dimensional interpolation methods based on deep learning, comprising: according to 3D convolution constructs 5D convolution, and designs 5D convolution layer;Five-dimensional fully convolutional neural network model FCNN5D is built, wherein five-dimensional fully convolutional neural network FCNN5D includes stacked multiple 5D convolution layers;Randomly intercept N from simulation data Size N1×N2×N3×N4×N5 5D data block as data set D, wherein a part is as training set D train , a part is as verification set D val ;Training sample in step 3 is input to FCNN5D in step 2 and is trained, and loss function is calculated, network model parameter is trained by minimizing loss function, and the trained network parameter is verified FCNN5D performance on verification data set, and the highest performance parameter is selected as optimized FCNN5D parameter;FCNN5D in step 4 is used to interpolate other work area data.The application can construct high-dimensional convolution, extract all dimensional information of high-dimensional data, and have higher precision.

Description

Technical Field

[0001] This invention belongs to the technical field of data processing, specifically relating to a five-dimensional interpolation method for seismic data based on deep learning. Background Technology

[0002] Due to limitations imposed by the surface environment, obstacles, and restricted mining areas, the distribution of seismic sources and geophones is often uneven or impossible, leading to missing data. Spatially irregular sampling can cause spatial frequency distortion or offset arcing during data processing. Therefore, utilizing data interpolation techniques for high-precision, high-resolution, and efficient regularization reconstruction of seismic gathers is crucial for obtaining high-quality seismic data, reducing exploration costs, and improving imaging accuracy.

[0003] Three-dimensional seismic data exploration brings the data dimension to five dimensions (5D). The representation of these five dimensions differs across different data domains: the shot-receiver domain typically includes time, source coordinates (x and y), and receiver coordinates (x and y); the common midpoint domain typically includes time, common midpoint coordinates (x and y), and common offset coordinates (x and y), or time, common midpoint coordinates (x and y), shot-receiver distance, and azimuth. Five-dimensional seismic data contains more spatial information. Compared to conventional two-dimensional or three-dimensional interpolation, five-dimensional interpolation can fully utilize the correlation characteristics and spatial information features of different azimuths, offsets, etc., in high-dimensional data, enabling more accurate prediction of missing traces. Currently, five-dimensional interpolation has become an important means of improving imaging accuracy in the reprocessing of old data.

[0004] In recent years, many traditional methods have been applied to 5D seismic data interpolation, among which parameter selection has a significant impact on the interpolation results. For example, Fourier transform domain methods: industrially, methods such as Minimum Weighted Norm Interpolation (MWNI) and Leakage-Resistant Fourier Transform (ALFT) are commonly used, requiring adjustment of regularization parameters; prediction filter methods include Non-Stationary Prediction Error Filter (NPEF), requiring adjustment of window size, regularization parameters, etc.; methods based on rank reduction theory include Damped Matrix Rank Reduction (DRR) and Parallel Matrix Factorization (PMF), requiring adjustment of parameters such as window size and rank.

[0005] Deep learning-based methods train network models using large amounts of data to find the mapping between missing and complete data. After training, the trained network is used to directly interpolate the missing data, which is time-efficient and requires no parameter tuning. However, existing frameworks like PyTorch and TensorFlow only provide low-dimensional convolutions. Therefore, constructing high-dimensional convolutions is a challenge for using deep learning methods to interpolate high-dimensional data. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing a five-dimensional interpolation method for seismic data based on deep learning. This method can construct high-dimensional convolutions, extract all dimensional information from high-dimensional data, and has high accuracy.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A five-dimensional interpolation method for seismic data based on deep learning includes the following steps: Step 1: Construct 5D convolution based on 3D convolution, and design 5D convolutional layers; Step 2: Construct a five-dimensional fully convolutional neural network model FCNN5D, which consists of multiple stacked 5D convolutional layers; Step 3: Randomly extract data from the simulation data The size is 5D data blocks as dataset One part of it is used as the training set. A portion of it serves as a verification set. ; Step 4: Input the training samples from Step 3 into the FCNN5D constructed in Step 2 for training, calculate the loss function, train the network model parameters by minimizing the loss function, and then verify the performance of FCNN5D on the validation dataset using the trained network parameters. Select the parameters corresponding to the best performance of FCNN5D on the validation set as the optimized FCNN5D parameters. Step 5: Use the optimized FCNN5D from Step 4 to interpolate the data from other work areas.

[0008] Furthermore, the specific method for step 1 is as follows: Record the network Layer, number The 5D convolution operator for the layer is Size is , No. The input feature layer of the layer is Size is ,in For batch dimension, Input the number of channels; when , At that time, the size of the 5D convolution operator is After removing a single dimension, it is recorded as The corresponding 5D input feature layer size is After removing a single dimension, it is recorded as The output is Size is , The 5D convolution process is represented as follows:

[0009] Record No. The activation function of a layer is denoted as: Then the first A convolutional layer is represented as , No. A convolutional layer is represented as .

[0010] Furthermore, the method for constructing dataset D in step 3 is as follows: The original complete data block is used as a label and normalized to the range [-1, 1]. Then, a portion of the seismic traces in each data block are randomly selected and set to 0 to simulate random missing data. The normalized complete data block is denoted as […]. The missing data block is The dataset is composed of both complete and missing data, represented as follows: .

[0011] Furthermore, the normalization method involves dividing each complete data point by the maximum value of the absolute signal value.

[0012] Furthermore, in step 4, the loss function for: ; in, The square of the Frobenius norm. The number of samples in the training set. The network parameters to be trained, For missing data blocks, This is a complete data block.

[0013] Furthermore, in step 4, batch training is used when training FCNN5D. That is, a batch of sample pairs is input into FCNN5D in each iteration, and a gradient descent-like algorithm is used to minimize the loss function to train the FCNN5D parameters. The trained parameters are: .

[0014] Furthermore, gradient descent algorithms include, but are not limited to, batch gradient descent, stochastic gradient descent, and the Adam algorithm.

[0015] Furthermore, in step 4, the average signal-to-noise ratio (SNR) is used on the validation set to evaluate the performance of the trained FCNN5D parameters. The formula for calculating SNR is:

[0016] in, Indicates complete data. This represents the FCNN5D interpolation result; Among them, a higher signal-to-noise ratio indicates a better interpolation effect. After multiple rounds of training, the parameters corresponding to the round with the largest average SNR on the validation set are selected as the optimal parameters for FCNN5D.

[0017] Furthermore, let the optimized FCNN5D be denoted as... ,use Missing data for other work areas Perform interpolation, and record the interpolation result as... :

[0018] in, and All data are normalized. For the sampling tensor, " " represents a tensor consisting only of the digit 1, with a size of same," " is the Hadamard product; right Further inverse normalization is performed to obtain the interpolation result of the original data.

[0019] Compared with existing technologies, the beneficial effects of this invention are as follows: The five-dimensional interpolation method for seismic data based on deep learning proposed in this invention is based on the mathematical definition of convolution. It uses 3D convolution to construct 5D convolution and designs 5D convolutional layers to extract information from all dimensions of the data. It then builds FCNN5D to complete the mapping from missing data to complete data. Compared with existing three-dimensional convolutional neural networks, this invention can extract information from all dimensions of five-dimensional data, has higher accuracy, and is a more intelligent and accurate interpolation method. In addition, this invention is a data-driven method and can be implemented based on the PyTorch or Tensorflow framework. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the 5D convolution construction method according to an embodiment of the present invention; Figure 2 A schematic diagram of the FCNN5D structure constructed according to an embodiment of the present invention; Figure 3 This refers to the original data of routes 26 to 41 of the narrow azimuth flight path data simulated in this embodiment of the invention; Figure 4 The interpolation results are for a common shot point gather in routes 26 to 41 of the present invention, where (a) is complete data and (b) is missing data with 80% of the seismic traces randomly missing. Figure 5The diagram shows a comparison between the interpolation results of this embodiment and the interpolation results of the UNet3D method, where (a) is the interpolation result of the UNet3D method, (b) is the interpolation result of the method in this embodiment, (c) is the residual of the interpolation result of the UNet3D method, and (d) is the residual of the interpolation result of the method in this embodiment. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0023] The present invention will be further described below with reference to specific embodiments, but these are not intended to limit the scope of the invention.

[0024] This invention provides a five-dimensional interpolation method for seismic data based on deep learning, comprising the following steps: Step 1: Construct 5D convolutions using 3D convolutions and design the 5D convolutional layers. For example... Figure 1 As shown, in this step, the network total... Layer, number The 5D convolution operator for the layer is Size is , No. The input feature layer of the layer is Size is ,in Indicates batch dimension, The number of channels for the input feature. When , At that time, the corresponding 5D convolution operator size is After removing a single dimension, it is recorded as The corresponding 5D input feature layer size is After removing a single dimension, it is recorded as The output is Size is , Given the number of output channels, the 5D convolution process can be represented as:

[0025] Record No. The activation function of a layer is denoted as: Then the first A convolutional layer can be represented as ;No. A convolutional layer is represented as The output features of a 5D convolutional layer are denoted as... , To determine the number of output channels, The input is fed into a 5D convolutional layer, which extracts features in five dimensions and outputs a 7-dimensional tensor. .

[0026] Step 2: Construct a five-dimensional fully convolutional neural network model FCNN5D, which consists of multiple stacked 5D convolutional layers; In this embodiment, a five-dimensional fully convolutional neural network model, FCNN5D, is constructed. The input of FCNN5D is the missing 5D seismic data, and the output is the interpolation result. Figure 2 As shown, the FCNN5D in this embodiment consists of five stacked 5D convolutional layers, denoted as... ,but ,in" "" indicates the network parameters to be trained. Of course, in other embodiments, other numbers of 5D convolutional layers can be stacked to form FCNN5D.

[0027] Step 3: Randomly extract data from the simulation data The size is 5D data blocks as dataset One part of it is used as the training set. The other part serves as the verification set. ; In this step, the original complete data block is used as a label and normalized to the range [-1, 1]. The normalization method is to divide each complete data point by the maximum absolute value of the signal. Then, a portion (up to 80%) of the seismic traces in each data block is randomly selected and set to 0 to simulate random missing data. The normalized complete data block is denoted as […]. The missing data block is The complete and missing data are input into the network. The dataset is represented as follows: .

[0028] In this embodiment, the simulation data includes 51 flight paths, each with 96 seismic guns, 8 cables per gun, and 68 geophones per cable, with 625 time sampling points. The five dimensions of the data are time, source x and y coordinates, and geophone x and y coordinates. A 16-bit sample is randomly selected from the first 25 flight paths. 16 16 8 The training set contains 1000 data blocks, each denoted as 16. The validation set contains 100 data blocks.

[0029] Step 4: Input the training samples from Step 3 into the FCNN5D constructed in Step 2 for training, calculate the loss function, train the network model parameters by minimizing the loss function, and then verify the performance of FCNN5D on the validation dataset using the trained network parameters. Select the parameters corresponding to the best performance of FCNN5D on the validation set as the optimized FCNN5D parameters. The loss function is constructed as follows:

[0030] in, For the constructed FCNN5D, For network parameters, Represents the training dataset. For missing data blocks, This is a complete data block.

[0031] FCNN5D is trained using a training set. Batch training is employed, meaning a batch of sample pairs is input into FCNN5D for each iteration. Gradient descent algorithms (batch gradient descent, stochastic gradient descent, Adam algorithm, etc.) are used to minimize the loss function to train the FCNN5D parameters. The trained parameters are expressed as follows: ; In each iteration update of the FCNN5D parameters, the batch of samples used for training at the current iteration number is updated. The input to FCNN5D is obtained by randomly removing 80% of its seismic traces (setting missing traces to 0). FCNN5D was trained for 100 epochs, with an initial learning rate of 0.0001, cosine decay, and a batch size of 1. After each training epoch, the average signal-to-noise ratio (SNR) was calculated on the validation set. The formula for SNR is:

[0032] In the formula, Indicates complete data. This represents the FCNN5D interpolation result; A higher signal-to-noise ratio indicates a better interpolation effect. After multiple rounds of training, the parameters corresponding to the round with the largest average SNR on the validation set are selected as the optimal parameters for FCNN5D.

[0033] Step 5: Use the optimized FCNN5D from Step 4 to interpolate the data from other work areas; Let the optimized FCNN5D be... ,use For new missing data Perform interpolation, and denote the interpolation result as... ,but:

[0034] in, and All data are normalized. This is a sampling tensor, composed of the numbers "0" and "1", where "1" indicates that the seismic trace is not missing and "0" indicates that the seismic trace is missing. Its size is the same as that of the data block. same;" " represents a tensor consisting only of the digit "1", with a size of same;" " is the Hadamard product.

[0035] right Further inverse normalization yields the interpolated result of the original data.

[0036] Data is presented using an overlapping sliding window method. The data is divided into multiple overlapping windows, and interpolation is performed on the data in each window. Finally, the interpolation results are integrated together, and the overlapping areas are integrated using a weighted average method.

[0037] The above method was used to interpolate routes 26 to 41 in the simulation data. Figure 3 This displays complete data for flight paths 26-41 of the aforementioned narrow azimuth data, with a common shot point of (10, 46) and 625 time sampling points. Then, it uses... Interpolate the data for routes 26 to 41 of the above narrow azimuth data. Figure 4 A common shot gather from the interpolation results is shown here. The source coordinates are (10, 46), the detector coordinates are (3:6, 17:48), and the number of time sampling points is 400. Figure 4 (a) is the complete data. Figure 4 (b) shows the missing data for 80% of the seismic traces that are randomly missing.

[0038] To illustrate the effect of this embodiment, Figure 5 The interpolation results of this embodiment are compared with those of the UNet3D method, wherein... Figure 5 (a) and (b) correspond to the interpolation results of UNet3D and the interpolation results of the method in this embodiment, respectively, with signal-to-noise ratios of 12.08dB and 15.00dB, respectively. Figure 5 (c) and (d) correspond to the residuals of the UNet3D method and the method of this embodiment, respectively. Regarding the signal-to-noise ratio (SNR), the interpolation result of this embodiment has a higher SNR. Figure 5As can be seen from the rectangular areas in (a) and (b), in regions with continuous missing segments, the in-phase axis structure of the interpolation result in this embodiment is clearer and more continuous. From the residual... Figure 5 As can be seen from (c) and (d), there is less signal in the areas not reconstructed by the method in this embodiment, which indicates that this embodiment is a very effective five-dimensional interpolation method.

[0039] The above embodiments illustrate that this embodiment can effectively interpolate five-dimensional data, laying the foundation for subsequent data analysis.

[0040] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the content of this specification should be included within the protection scope of the present invention.

Claims

1. A five-dimensional interpolation method for seismic data based on deep learning, characterized in that, Includes the following steps: Step 1: Construct 5D convolution based on 3D convolution, and design 5D convolutional layers; Step 2: Construct a five-dimensional fully convolutional neural network model FCNN5D, which consists of multiple stacked 5D convolutional layers; the input of FCNN5D is the missing 5D seismic data, and the output is the interpolation result; Step 3: Randomly extract data from the simulation data with time, source x-axis, y-axis, and detector x-axis and y-axis respectively. The size is 5D data blocks as dataset One part of it is used as the training set. A portion of it serves as a verification set. ; Step 4: Input the training samples from Step 3 into the FCNN5D constructed in Step 2 for training, and construct a loss function. Train the network model parameters by minimizing the loss function, and then verify the performance of FCNN5D on the validation dataset using the trained network parameters. Select the parameters that correspond to the best performance of FCNN5D on the validation set as the optimized FCNN5D parameters. Step 5: Use the optimized FCNN5D from Step 4 to interpolate the data from other work areas; The specific method for step 1 is as follows: Record the network Layer, number The 5D convolution operator for the layer is Size is , No. The input feature layer of the layer is Size is ,in For batch dimension, The input is the number of channels, and the output is... Size is , Number of output channels; when , At that time, the size of the 5D convolution operator is After removing a single dimension, it is recorded as The corresponding 5D input feature layer size is After removing a single dimension, it is recorded as The 5D convolution process can be represented as: Record No. The activation function of a layer is denoted as: Then the first A convolutional layer is represented as , No. A convolutional layer is represented as .

2. The five-dimensional interpolation method for seismic data based on deep learning according to claim 1, characterized in that, The method for constructing dataset D in step 3 is as follows: The original complete data block is used as a label and normalized to the range [-1, 1]. Then, a portion of the seismic traces in each data block are randomly selected and set to 0 to simulate random missing data. The normalized complete data block is denoted as […]. The missing data block is The dataset is composed of both complete and missing data, represented as follows: .

3. The five-dimensional interpolation method for seismic data based on deep learning according to claim 2, characterized in that, The normalization method is to divide each complete data point by the maximum value of the absolute value of the signal.

4. The five-dimensional interpolation method for seismic data based on deep learning according to claim 1, characterized in that, In step 4, the loss function for: ； in, The square of the Frobenius norm. The number of samples in the training set. The network parameters to be trained, For missing data blocks, This is a complete data block.

5. The five-dimensional interpolation method for seismic data based on deep learning according to claim 1, characterized in that, In step 4, batch training is used when training FCNN5D. That is, a batch of sample pairs is input into FCNN5D in each iteration, and gradient descent-like algorithms are used to minimize the loss function to train the FCNN5D parameters. The trained parameters are: 。 6. The five-dimensional interpolation method for seismic data based on deep learning according to claim 5, characterized in that, Gradient descent algorithms include, but are not limited to, batch gradient descent, stochastic gradient descent, and Adam algorithm.

7. The five-dimensional interpolation method for seismic data based on deep learning according to claim 1, characterized in that, In step 4, the average signal-to-noise ratio (SNR) is used on the validation set to evaluate the performance of the trained FCNN5D parameters. The formula for calculating SNR is: in, Indicates complete data. This represents the FCNN5D interpolation result; Among them, a higher signal-to-noise ratio indicates a better interpolation effect. After multiple rounds of training, the parameters corresponding to the round with the largest average SNR on the validation set are selected as the optimal parameters for FCNN5D.

8. The five-dimensional interpolation method for seismic data based on deep learning according to claim 1, characterized in that, Let the optimized FCNN5D be... ,use Missing data for other work areas Perform interpolation, and record the interpolation result as... : in, and All data are normalized. For sampling tensors, Represents a tensor consisting only of the digit 1, with a size of ... same, Represents the Hadamard product; right Further inverse normalization is performed to obtain the interpolation result of the original data.

Images