Post-stack seismic data intelligent frequency expanding method

Abstract

The invention provides an intelligent frequency expanding method for post-stack seismic data, and relates to the field of seismic exploration and development, and the method comprises the following steps: (1) constructing a network: constructing a deep neural network through employing a coding and decoding architecture combining a convolutional neural network and a recurrent neural network; (2) generating a synthetic training data set which comprises input data and label data; and (3) training and prediction: training the deep neural network obtained in the step (1) by using the synthetic data set obtained in the step (2) to obtain a trained network model, inputting seismic data to be expanded into the trained network model to obtain frequency-expanded seismic data, and broadening the frequency band of the seismic data predicted by the trained network model by 10-15%. According to the post-stack seismic data intelligent frequency expanding method, the seismic data and the logging data are combined for frequency expanding, broadband information of the logging data is introduced into the seismic data, and the seismic data after frequency expanding is high in well-seismic correlation and high in reliability.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of seismic exploration and development, and particularly relates to an intelligent frequency extension method for post-stack seismic data. BACKGROUND

[0002] The vertical resolution of seismic data is determined by the bandwidth and dominant frequency of the data. Because the width of the source spectrum is limited, the seismic signal is also band-limited. During the propagation of seismic waves, the high-frequency components will be severely attenuated due to the absorption of the strata. Therefore, the conventional acquired seismic data lack effective high-frequency or low-frequency information and have insufficient resolution. The resolution of the conventional acquired seismic data is insufficient for the purpose of seismic interpretation. In general, people define the horizon smaller than one-quarter wavelength of the dominant frequency of the data as a thin layer. For such a horizon, the top and bottom of the reflected wave will be tuned, and therefore cannot be distinguished. In actual seismic data, because there are many thin reservoirs or small structures, it is usually necessary to improve the resolution of the seismic data. The conventional seismic data frequency extension method usually requires different assumptions to achieve better frequency extension effect. However, the actual seismic data of different work areas are quite different, and various noises and complex underground structures often make it difficult for the seismic data to meet the assumption conditions of the conventional frequency extension method, thereby affecting the frequency extension accuracy. In addition, most conventional frequency extension methods only use seismic data, which often introduces various artifacts, resulting in low accuracy of the frequency extension result. In recent years, deep learning algorithms have received widespread attention from geophysicists. Because the deep learning algorithm is a data-driven method, it has been applied to various geophysical problems that require generalization.

[0003] A Chinese patent with publication number CN116643310A discloses a seismic frequency extension method, device, electronic equipment and computer readable storage medium. The method in the invention includes: obtaining first intermediate frequency data and first full frequency data in seismic sample data; the first intermediate frequency data is used to represent the data in the first frequency range in the seismic sample data, and the first full frequency data is used to represent the data in the second frequency range in the seismic sample data; training a first neural network according to the first intermediate frequency data and the first full frequency data; performing seismic frequency extension on second intermediate frequency data according to the trained first neural network to obtain predicted second full frequency data. The invention proposes to use the full-band data of sample seismic data as training data for seismic frequency extension, to restore the low-frequency information and high-frequency information of seismic data at a lower cost, to improve the accuracy of full-waveform inversion, and to improve the resolution of seismic data.

[0004] Chinese Patent Publication No. CN113721294A discloses a complex-domain least-squares constrained spectral blueing upscaling method, proposing to design a broadband target spectrum to broaden the spectrum from the frequency domain. Specifically, it involves a complex-domain least-squares constrained spectral blueing upscaling method. The method includes: establishing an upscaling objective function based on the complex-domain least-squares method; solving the objective function to obtain a spectral blueing upscaling operator; multiplying the spectral blueing upscaling operator by the original seismic spectrum; and then performing an inverse Fourier transform to achieve upscaling processing of the seismic data. This method can effectively compensate for lost low-frequency information and enhance attenuated high-frequency information without changing the phase spectrum of the seismic data. Furthermore, by setting different constraint parameters, the distance between the spectrum of the upscaled seismic data and the broadband target spectrum can be adjusted to obtain upscaling results with different dominant frequencies and bandwidths.

[0005] For seismic data topology processing, the key issue lies in the reliability of the topology data. However, the methods mentioned above only use seismic data, which often introduces various artifacts and mismatches between well and seismic information. To address these issues, it is essential to find an intelligent topology method for post-stack seismic data that offers high well-seismic correlation and strong reliability. Summary of the Invention

[0006] This invention addresses the problems existing in the prior art by providing an intelligent frequency extension method for post-stack seismic data. The method in this invention combines well logging data and seismic data to create a dataset to train the network model, which has strong generalization ability and can be applied to data from work areas with different geological characteristics. This method overcomes the shortcomings of traditional methods that only use seismic data for frequency extension, and provides a way to introduce broadband information from well logging data into seismic data. It also has strong noise resistance and is suitable for various data scenarios.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A smart frequency topology method for post-stack seismic data includes the following steps:

[0009] (1) Building a deep neural network: A deep neural network is built using an encoder-decoder architecture that combines convolutional neural networks and recurrent neural networks;

[0010] (2) Generate synthetic training dataset: including input data and label data, the dataset is generated by the convolution of reflection coefficient and seismic wavelet data, noise is added to generate input data related to the characteristics of actual data; the label data for training is generated by the convolution of reflection coefficient and broadband B-spline wavelet data;

[0011] (3) Training and prediction: Use the synthetic dataset obtained in step (2) to train the deep neural network obtained in step (1) to obtain the trained network model. Input the earthquake data to be expanded into the trained network model to obtain the earthquake data with expanded frequency. The earthquake data frequency band predicted by the trained network model is widened by 10%-15%.

[0012] Preferably, the encoding and decoding architecture of the deep neural network is approximately symmetrical from left to right.

[0013] Preferably, the encoding part consists of two convolutional layers at each step, followed by a max pooling layer for downsampling; the decoding part includes upsampling layers, followed by two bidirectional convolutional LSTM (Long Short-Term Memory) networks for each upsampling layer.

[0014] This invention uses a two-layer bidirectional convolutional LSTM to solve the problem of LSTM's insufficient ability to capture the spatial characteristics of data.

[0015] Preferably, the upsampling layer uses the difference plus convolution method, and the max pooling layer is followed by batch normalization to accelerate the network convergence speed.

[0016] Preferably, step (2) specifically includes the following steps:

[0017] By analyzing the characteristics of actual data, wavelet information is extracted from seismic data to obtain seismic wavelet data;

[0018] Reflection coefficient generation: The logging data is converted to time depth through well-seismic calibration, and then the P-wave velocity and density curves are resampled to calculate the reflection coefficient;

[0019]

[0020] Where r is the reflection coefficient, ρ is the density, v is the velocity, and i = 1, 2, 3, ...;

[0021] A dataset is generated by convolution of reflection coefficients and seismic wavelet data, noise is added, and input data related to the characteristics of the actual data is generated; label data for training is generated by convolution of reflection coefficients and broadband B-spline wavelet data.

[0022] Preferably, the sampling rate of the P-wave velocity and density curves is the same as the sampling rate of the seismic data.

[0023] This invention uses broadband B-spline wavelets to generate training label data. Compared to the Ricker wavelet, the broadband B-spline wavelet has more adjustable parameters, allowing for arbitrary changes in the wavelet's dominant frequency and bandwidth, and a larger main-to-side lobe ratio. Moreover, compared to the Ormsby wavelet, the broadband B-spline wavelet produces fewer sidelobes with shorter durations.

[0024] Preferably, the actual data characteristics include the frequency content of the seismic wavelet, the frequency shape of the seismic wavelet, the content of the reflection coefficient, and the content of random noise.

[0025] Preferably, the noise is random noise with a white spectrum.

[0026] Preferably, the synthetic training dataset and the actual data are normalized, and the normalization formula is as follows:

[0027]

[0028] Where, x min and x max and are the minimum and maximum values ​​of the synthetic training dataset or the actual data, respectively. x is the original input synthetic training dataset or the original input actual data, and x′ is the normalized synthetic training dataset or the actual data.

[0029] Preferably, the training process in step (3) uses the gradient descent algorithm in the backpropagation algorithm to update and learn the deep neural network.

[0030] Preferably, the backpropagation algorithm is as follows: Initially, the initial weight values ​​of the deep neural network are random. After inputting the input data into the deep neural network, the output value of the deep neural network is calculated by combining the input data with the weight values ​​of the deep neural network. The error between the output value of the deep neural network and the training label data is calculated by the loss function. The error is backpropagated back to the deep neural network. The weight values ​​of the deep neural network are iteratively updated according to the gradient descent algorithm. The above process is repeated until the error meets the accuracy requirements.

[0031] Preferably, the loss function is calculated using the Symmetric Mean Absolute Percentage Error (SMAPE), and the formula is:

[0032]

[0033] Among them, X i Y is the output value of the deep neural network. i The training data consists of labeled data, where n is the number of labels and i ranges from 1 to n.

[0034] Preferably, the Adam optimizer algorithm is used when iteratively updating the weight values ​​of the deep neural network according to the gradient descent algorithm.

[0035] The Adam optimizer of this invention combines the advantages of SGDM and RMSProp optimization algorithms, and comprehensively considers the first-order momentum and second-order momentum of the gradient to calculate the update step size.

[0036] The formula for calculating the first-order momentum of an SGDM is:

[0037] mt =β1·m t-1 +(1-β1)·g t

[0038] Adding the second momentum V of RMSProp t The calculation formula is:

[0039]

[0040] Among them, g t Let β1 and β2 represent the gradient at time t, where t represents the total number of iterations. The coefficients β1 and β2 are the exponential decay rates. β1 controls the weight distribution (momentum and current gradient), and β2 controls the influence of the squared gradient. The squared gradient is weighted and averaged. Both β1 and β2 have a value of 0.999.

[0041] Both first-order and second-order momentum are calculated using exponential moving averages.

[0042] Initialize m0 = 0, V0 = 0. In the initial stage, the m obtained through iteration... t V t It will approach 0, through m t and V t This problem can be solved by performing deviation correction. and The corresponding correction values ​​are calculated using the following formulas:

[0043]

[0044] in, β1 raised to the power of t It is β² raised to the power of t.

[0045] The gradient descent algorithm has the following calculation formula:

[0046]

[0047] Where, θ t Let θ be the weight value at time t. t-1 Here are the weight values ​​at time t-1, α is the learning rate, and ε = 10. -8 , and These are the correction values ​​for first-order momentum and second-order momentum, respectively.

[0048] To prevent overfitting of data from causing poor prediction results, step (3) preferably includes equipping the deep neural network of step (1) with hyperparameters.

[0049] Preferably, the hyperparameters are set as shown in Table 1.

[0050] Table 1

[0051] Parameter Type Setting Convolution Kernel Size 3 Activation Function ReLU Optimizer Adam Regularization Batch Normalization Initial Learning Rate 0.002 Batch Size 32

[0052] The technical effects achieved by this invention are:

[0053] 1. This invention has been tested on model data and applied to actual data. The predicted seismic data volume wavelet length is compressed, sidelobe energy is suppressed, and the wavelet frequency band is significantly broadened. The phase axis of the processed seismic profile is significantly thinner, and the resolution of thin layers and small structures is improved. Spectral comparison shows that the seismic data upscaling processing method based on deep learning technology can predict information outside the original data frequency band, and the reliability of this method is verified by well-seismic comparison.

[0054] 2. This invention extracts features from the original data and combines them with well logging and seismic data to create a training dataset for training the network model. This enhances the generalization ability of the network model, making it applicable to work areas with different geological characteristics. The resolution of the frequency-spread seismic data is improved, the well-seismic correlation is high, and the signal-to-noise ratio is high, which helps to identify thin layers and small structures, thus having broad prospects for widespread application. Attached Figure Description

[0055] Figure 1 A flowchart for an intelligent frequency topology method for post-stack seismic data;

[0056] Figure 2 It is a deep neural network architecture;

[0057] Figure 3 The process flow for generating synthetic training datasets, training, and prediction steps;

[0058] Figure 4 This is a schematic diagram of the work area, where (a) is a plan view of the work area, (b) is a seismic profile of a certain survey line, and (c) is an example of some well logging curves;

[0059] Figure 5 The comparison is between wavelet and spectrum, where (a) is wavelet comparison and (b) is spectrum comparison;

[0060] Figure 6 The results are the frequency extension processing results, where (a) and (b) are the seismic profiles of the main survey line before and after frequency extension, respectively; and (c) and (d) are the seismic profiles of the connecting line before and after frequency extension, respectively. Detailed Implementation

[0061] Preferred embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While preferred 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 the present disclosure will be thorough and complete, and will fully convey the scope of the present disclosure to those skilled in the art.

[0062] like Figure 1 As shown, a smart frequency topology method for post-stack seismic data includes the following steps:

[0063] (1) Building a deep neural network: A deep neural network is built using an encoder-decoder architecture that combines convolutional neural networks and recurrent neural networks;

[0064] (2) Generate synthetic training dataset: including input data and label data, the dataset is generated by the convolution of reflection coefficient and seismic wavelet data, noise is added to generate input data related to the characteristics of actual data; the label data for training is generated by the convolution of reflection coefficient and broadband B-spline wavelet data;

[0065] (3) Training and prediction: Use the synthetic dataset obtained in step (2) to train the deep neural network obtained in step (1) to obtain the trained network model. Input the earthquake data to be expanded into the trained network model to obtain the earthquake data with expanded frequency. The earthquake data frequency band predicted by the trained network model is widened by 10%-15%.

[0066] like Figure 2 As shown, in one example, the encoding and decoding architecture of the deep neural network in step (1) is approximately symmetrical from left to right; the encoding part consists of two convolutional layers in each step, followed by a max pooling layer for downsampling; the decoding part includes upsampling layers, followed by two bidirectional convolutional LSTM layers; the upsampling layers use the difference plus convolution method, and the max pooling layers use batch normalization to accelerate network convergence;

[0067] like Figure 3 As shown, in one example, step (2) generates the synthetic training dataset and the labeled training data, and step (3) training and prediction includes:

[0068] Analyze the characteristics of actual data, extract wavelet information from seismic data to obtain seismic wavelet data; generate reflection coefficient: perform time-depth conversion on well logging data through well-seismic calibration, then resample the P-wave velocity and density curves to calculate the reflection coefficient;

[0069]

[0070] Where r is the reflection coefficient, ρ is the density, v is the velocity, and i = 1, 2, 3, ...;

[0071] A dataset is generated by convolution of reflection coefficients and seismic wavelet data, noise is added, and a synthetic training dataset related to the characteristics of the actual data is generated; labeled training data is generated by convolution of reflection coefficients and broadband B-spline wavelet data; the sampling rate of P-wave velocity and density curves is the same as the sampling rate of seismic data.

[0072] The synthetic dataset obtained in step (2) is used to train the deep neural network obtained in step (1) to obtain the trained network model. The seismic data to be overlaid is then input into the trained network model to obtain the overlaid seismic data.

[0073] In one example, the actual data characteristics include the frequency content of the seismic wavelet, the frequency shape of the seismic wavelet, the content of the reflection coefficient, and the content of random noise.

[0074] In one example, the noise is random noise with a white spectrum.

[0075] In one example, the synthetic training dataset and the actual data are normalized using the following formula:

[0076]

[0077] Where, x min and x max and are the minimum and maximum values ​​of the synthetic training dataset or the actual data, respectively. x is the original input synthetic training dataset or the original input actual data, and x′ is the normalized synthetic training dataset or the actual data.

[0078] In one example, the training process in step (3) uses the gradient descent algorithm in the backpropagation algorithm to update the deep neural network.

[0079] In one example, the backpropagation algorithm is as follows: Initially, the weights of the deep neural network are random. After inputting the synthetic training dataset into the deep neural network, the output value of the deep neural network is calculated by combining the weights with the output value. The error between the output value and the training label data is calculated using a loss function. The error is backpropagated back to the deep neural network. The weights of the deep neural network are iteratively updated according to the gradient descent algorithm. The above process is repeated until the error meets the accuracy requirements.

[0080] In one example, the loss function is calculated using the Symmetric Mean Absolute Percentage Error (SMAPE), as follows:

[0081]

[0082] Among them, Xi Y is the output value of the deep neural network. i The training data consists of labeled data, where n is the number of labels and i ranges from 1 to n.

[0083] In one example, the Adam optimizer algorithm is used when the weights of the deep neural network are iteratively updated according to the gradient descent algorithm.

[0084] The formula for calculating the first-order momentum of an SGDM is:

[0085] m t =β1·m t-1 +(1-β1)·g t

[0086] Adding the second momentum V of RMSProp t The calculation formula is:

[0087]

[0088] Among them, g t Let β1 and β2 represent the gradient at time t, where t represents the total number of iterations. The coefficients β1 and β2 are the exponential decay rates. β1 controls the weight distribution (momentum and current gradient), and β2 controls the influence of the squared gradient. The squared gradient is weighted and averaged. Both β1 and β2 have a value of 0.999.

[0089] Both first-order and second-order momentum are calculated using exponential moving averages.

[0090] and The corresponding correction values ​​are calculated using the following formulas:

[0091]

[0092] in, β1 raised to the power of t It is β² raised to the power of t.

[0093] The gradient descent algorithm has the following calculation formula:

[0094]

[0095] Where, θ t Let θ be the weight value at time t. t-1 Here are the weight values ​​at time t, α is the learning rate, and ε = 10. -8 , and These are the correction values ​​for first-order momentum and second-order momentum, respectively.

[0096] In one example, step (3) also includes equipping the deep neural network of step (1) with hyperparameters.

[0097] In one example, the hyperparameter settings are shown in Table 1.

[0098] Application Examples

[0099] This application selects a specific work area in a basin as an example, performs frequency conversion processing on the post-stack seismic data of this work area, and analyzes its frequency conversion effect. A schematic diagram of the work area is shown below. Figure 4 As shown.

[0100] The actual characteristics of the post-stack seismic data in this work area include 142 inline lines and 110 xline lines, with a time length of 2000 ms and a seismic data volume sampling time of 1 ms. The target layers are distributed between 750 ms and 1000 ms, with a total of three layers, from top to bottom: T1, T2, and T3. The entire work area includes 30 wells; data from 25 of these wells were selected for training, and data from the remaining wells were used for testing.

[0101] After performing low-frequency extension on all seismic traces in the entire work area, statistical wavelets were extracted from the entire seismic data volume before and after the extension and compared and analyzed (see...). Figure 5 (a)), from Figure 5 As can be seen in (a), the wavelet length is compressed after frequency extension, and the sidelobe energy is weaker.

[0102] Figure 5 (b) shows a comparison of the spectrum before and after the frequency expansion, from... Figure 4 (b) It can be seen that the bandwidth of the frequency-extended seismic profile is wider than that of the original input seismic profile, and the low-frequency information is also effectively protected.

[0103] One inline and one xline profile are as follows before and after frequency extension: Figure 6 As shown. From Figure 6 It can be seen that the visual resolution of each profile is significantly enhanced after the frequency upscaling process, the thin-layer recognition capability is improved, and the T1, T2 and T3 layers match well. At the same time, the lateral continuity of the phase axis is also well maintained.

[0104] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions made by those skilled in the art to the technical solution of the present invention do not depart from the essence and scope of the technical solution of the present invention.

Claims

1. A smart frequency topology method for post-stack seismic data, characterized in that: Includes the following steps: (1) Network construction: A deep neural network is constructed using an encoding and decoding architecture that combines convolutional neural networks and recurrent neural networks; (2) Generate synthetic training dataset: including input data and label data, the dataset is generated by the convolution of reflection coefficient and seismic wavelet data, noise is added to generate input data related to the characteristics of actual data; the label data for training is generated by the convolution of reflection coefficient and broadband B-spline wavelet data; (3) Training and prediction: Use the synthetic dataset obtained in step (2) to train the deep neural network obtained in step (1) to obtain the trained network model. Input the earthquake data to be expanded into the trained network model to obtain the earthquake data with expanded frequency. The earthquake data frequency band predicted by the trained network model is widened by 10%-15%.

2. The intelligent frequency extension method for post-stack seismic data according to claim 1, characterized in that: The encoding part in step (1) consists of two convolutional layers in each step, followed by a max pooling layer for downsampling; the decoding part includes upsampling layers, followed by two bidirectional convolutional LSTM layers in each upsampling layer.

3. The intelligent frequency topology method for post-stack seismic data according to claim 2, characterized in that: The upsampling layer uses the difference plus convolution method, and the max pooling layer is followed by batch normalization to accelerate the network convergence speed.

4. The intelligent frequency extension method for post-stack seismic data according to claim 1, characterized in that: Step (2) specifically includes the following steps: By analyzing the characteristics of actual data, wavelet information is extracted from seismic data to obtain seismic wavelet data; Reflection coefficient generation: The logging data is converted to time depth through well-seismic calibration, and then the P-wave velocity and density curves are resampled to calculate the reflection coefficient; Where r is the reflection coefficient, ρ is the density, v is the velocity, and i = 1, 2, 3, ...; A dataset is generated by convolution of reflection coefficients and seismic wavelet data, noise is added, and input data related to the characteristics of the actual data is generated; label data for training is generated by convolution of reflection coefficients and broadband B-spline wavelet data.

5. The intelligent frequency topology method for post-stack seismic data according to claim 4, characterized in that: The sampling rate for P-wave velocity and density curves is the same as that for seismic data.

6. The intelligent frequency topology method for post-stack seismic data according to claim 5, characterized in that: The actual data characteristics include the frequency content of the seismic wavelet, the frequency shape of the seismic wavelet, the content of the reflection coefficient, and the content of random noise.

7. The intelligent frequency extension method for post-stack seismic data according to claim 1, characterized in that: Opposition The training dataset and the actual data are normalized using the following formula: Where, x min and x max and are the minimum and maximum values ​​of the synthetic training dataset or the actual data, respectively. x is the original input synthetic training dataset or the original input actual data, and x′ is the normalized synthetic training dataset or the actual data.

8. The intelligent frequency extension method for post-stack seismic data according to claim 1, characterized in that: The training process in step (3) uses the gradient descent algorithm in the backpropagation algorithm to update the deep neural network. The backpropagation algorithm is as follows: initially, the initial weight values ​​of the deep neural network are random. After inputting the input data into the deep neural network, the output value of the deep neural network is calculated with the weight values ​​of the deep neural network. The error between the output value of the deep neural network and the training label data is calculated through the loss function. The error is backpropagated back to the deep neural network. The weight values ​​of the deep neural network are iteratively updated according to the gradient descent algorithm. The above process is repeated until the error meets the accuracy requirements.

9. The intelligent frequency extension method for post-stack seismic data according to claim 8, characterized in that: The loss function is calculated using the Symmetric Mean Absolute Percentage Error (SMAPE), and the formula is as follows: Among them, X i Y is the output value of the deep neural network. i The training data consists of labeled data, where n is the number of labels and i ranges from 1 to n.

10. The intelligent frequency extension method for post-stack seismic data according to claim 9, characterized in that: The Adam optimizer algorithm is used when iteratively updating the weights of a deep neural network using the gradient descent algorithm.

Images