Seismic wavelet amplitude spectrum estimation method based on deep learning with prior information

Abstract

The application provides a prior information integrated deep learning seismic wavelet amplitude spectrum estimation method, comprising the following steps: step 1, pre-processing original seismic data to obtain post-stack seismic data; step 2, smoothing pre-processing the post-stack seismic data; step 3, constructing a deep learning network structure; step 4, estimating the seismic wavelet amplitude spectrum of actual data to obtain a shallow seismic wavelet; and step 5, verifying the actual data to correct a deep seismic wavelet. The prior information integrated deep learning seismic wavelet amplitude spectrum estimation method fully utilizes the prior information that the seismic wavelet spectrum has smoothness, inputs the pre-processed seismic record amplitude spectrum into the deep learning network, accurately estimates the non-single-peak and single-peak wavelet amplitude spectrum without relying on any assumption premise, and thus realizes the fidelity high-resolution processing of seismic data.

Description

Technical Field

[0001] This invention relates to the field of geophysical exploration technology, and in particular to a deep learning seismic wavelet amplitude spectrum estimation method incorporating prior information. Background Technology

[0002] Oil and natural gas are strategic resources crucial to national economic development and national security. Further promoting the exploration and development of oil and gas resources in my country is a major national strategy. Seismic exploration is a primary method for oil and gas exploration. It involves observing and analyzing the propagation patterns of seismic waves generated by artificial earthquakes in the underground medium, and using various signal processing and inversion techniques to infer the geometric structure and properties of underground rock strata. The estimation of seismic wavelets is of great significance for processing seismic data. In the frequency domain, a seismic wavelet is equivalent to the product of its amplitude spectrum and phase spectrum. Therefore, the method for estimating seismic wavelets can be divided into two steps: the first step is to determine the amplitude spectrum, and the second step is to determine the phase spectrum.

[0003] Existing seismic wavelet estimation methods can be divided into two main categories: deterministic and statistical methods. Walden (1998) proposed a spectral analysis method, and Edgar and Van der Baan (2011) proposed a matched wavelet estimation method. In their methods, accurate seismic wavelet amplitude spectra accelerate the convergence speed of the algorithm. The other category is statistical methods, which typically assume that the reflection coefficient sequence has white-noise Gaussian properties. Among these, Van der Baan and Fomel (2009) proposed a non-stationary wavelet estimation method, Robinson and Treitel (2000) proposed a time-varying convolution model for seismic records, and Gao et al. (2017) proposed an adaptive partitioning method for non-stationary seismic records. In their methods, finding the equivalent wavelet amplitude spectrum in each time interval is crucial.

[0004] Assuming the reflection coefficient sequence follows a white noise distribution, researchers have proposed several methods for estimating wavelet amplitude spectra, such as the correlation function method (CF), logarithmic spectral averaging (LSA), spectral simulation (SS), and the compression mapping operator method. These techniques have the following drawbacks:

[0005] (1) Traditional spectral simulation methods assume that the wavelet amplitude spectrum satisfies a function polynomial, the parameters of which can be determined by least-squares fitting between the seismic record amplitude spectrum and the function's independent variables. In practical applications, it is difficult to select a suitable function form. In addition, a large amount of actual well logging data shows that the reflection coefficient sequence does not satisfy a random distribution, and its amplitude spectrum conforms to the blue characteristics, showing a trend of non-white distribution, which makes the above methods difficult to apply in practice;

[0006] (2) In reality, the shape of seismic wavelets is complex, and the above methods have clear restrictions on the shape of the wavelet amplitude spectrum, resulting in a single-peak smooth curve similar to Ricker wavelets, which cannot solve the problem of estimating the amplitude spectrum of non-single-peak wavelets.

[0007] Chinese patent application CN201811159314.0 discloses a deep learning-based seismic data interpolation method and system. The method includes: Step 1, setting a neural network structure, and decomposing seismic data stored by trace into acquisition point seismic data stored by sampling point according to the neural network structure; Step 2, setting a threshold for the number of iterations and error control parameters for the neural network; Step 3, inputting the acquisition point seismic data into the neural network structure according to the neural network structure, the threshold for the number of iterations, and the error control parameters, using the amplitude values ​​of the sampling points as the output verification data to obtain neural network interpolation parameters; Step 4, replacing the acquisition point seismic data with missing data according to the neural network interpolation parameters, repeating Step 3 to obtain the amplitude values ​​of the missing data, and thus obtaining reconstructed data. This invention uses a deep learning neural network to perform multi-dimensional interpolation processing on seismic data to achieve seismic data reconstruction.

[0008] Chinese patent application CN202111376321.8 discloses a deep learning-based method and system for tunnel seismic wave velocity inversion. This invention adds environmental information, geological prior information, and tunnel exploration and observation method information encoding to the deep learning seismic inversion method to help the algorithm obtain the corresponding prior knowledge. It also utilizes existing databases to perform preliminary network pre-training on the tunnel seismic wave velocity inversion network to assist the algorithm in matching the current tunnel engineering exploration task. Simultaneously, it establishes an inversion optimization method that updates the network synchronously with tunnel excavation. While the tunnel is being excavated and explored, the database is expanded and the algorithm is optimized as exploration progresses. The training method gradually becomes unsupervised, resulting in better generalization and more accurate inversion results.

[0009] Chinese patent application CN202111402338.6 relates to a deep learning-based microseismic event detection method based on variance fractals, comprising: synthesizing effective signals from seismic records using Ricker wavelets; extracting and constructing a dataset containing noise and signal sets; calculating the variance fractal dimension of each signal and noise sample; inputting the noise and signal sets with calculated variance fractal dimensions into a one-dimensional convolutional neural network consisting of convolutional layers, activation layers, dropout layers, and fully connected layers for training; after training, saving the model; calculating the overall variance fractal dimension of the data to be detected; obtaining multiple samples by sliding a window on the data with the variance fractal dimension to be detected; and inputting these samples into the saved model to obtain the detection results.

[0010] The existing technologies described above are significantly different from our invention and have failed to solve the technical problem we want to address. Therefore, we have invented a new deep learning method for estimating the amplitude spectrum of seismic wavelet by incorporating prior information. Summary of the Invention

[0011] The purpose of this invention is to provide a deep learning method for estimating seismic wavelet amplitude spectra by incorporating prior information, thereby achieving accurate estimation of non-single-peak and single-peak wavelet amplitude spectra and enabling high-resolution processing of seismic data.

[0012] The objective of this invention can be achieved through the following technical measures: a deep learning seismic wavelet amplitude spectrum estimation method incorporating prior information, which includes:

[0013] Step 1: Preprocess the raw seismic data to obtain post-stack seismic data;

[0014] Step 2: Perform smoothing preprocessing on the post-stack seismic data;

[0015] Step 3: Construct the deep learning network structure;

[0016] Step 4: Estimate the wavelet amplitude spectrum of the actual data to obtain the shallow wavelet.

[0017] Step 5: Verify with actual data to achieve correction of deep seismic wavelets.

[0018] The objective of this invention can also be achieved through the following technical measures:

[0019] In step 1, the raw seismic data is preprocessed to obtain post-stack seismic data: s(n)∈R N×1 , where R represents the real number field and N represents the number of sampling points for the seismic data.

[0020] In step 2, based on the prior information of the smoothness of the seismic wavelet amplitude spectrum, the observation data is preprocessed, including:

[0021] Step 2a: Calculate the amplitude spectrum of the earthquake record;

[0022] Step 2b involves using integral operations in the cumulative distribution function to smooth the irregular seismic record amplitude spectrum, thereby obtaining a relatively smooth seismic wavelet spectrum.

[0023] In step 2a, the seismic record can be represented in the time domain as follows:

[0024] s(n)=r(n)*w(n)+n(n) (1)

[0025] In the formula, s(n) represents the seismic record, w(n) represents the seismic wavelet, r(n) is the reflection coefficient sequence, * represents the convolution operation, and n(n) is random noise, where w(n), r(n), and n(n) all belong to L 2 (R);

[0026] Typically, the wavelet amplitude spectrum of a seismic event is obtained from the amplitude spectrum of a seismic record through nonlinear calculation. Therefore, by performing a Fourier transform on both sides of equation (1), the amplitude spectrum of the seismic record can be obtained as shown in equation (2).

[0027]

[0028] In the formula, S(w), W(w), R(w) and Let represent the spectra of s(n), w(n), r(n), and n(n), respectively.

[0029] In step 2b, the specific process is as shown in equations (3) and (4):

[0030]

[0031]

[0032] Formula (3) normalizes the earthquake record amplitude spectrum |S(w)| obtained from formula (2), and then performs the integral operation in formula (4) on the normalized result f0(w). F0(w) represents the cumulative distribution function of the normalized result f0(w), and this operator plays the role of low-pass filtering.

[0033] In step 3, a deep learning neural network (DNN) based on Long Short-Term Memory (LSTM) is constructed, consisting of LSTM units and fully connected layers. The input data is the amplitude spectrum of the seismic record. Data points within a time window are input into the LSTM layer. After the data is input into the LSTM layer, each unit of the LSTM processes the input data and outputs a vector. This vector is then input into a fully connected neural network, which outputs a data point of the fitted data for the current time window. Finally, these individual data points are combined to obtain the estimated wavelet amplitude spectrum.

[0034] In step 3, the objective function for optimizing the number of model parameters is shown in equation (5):

[0035]

[0036] Where W(w) represents the true seismic wavelet amplitude spectrum, The estimated wavelet amplitude spectrum is represented by n, which represents the number of samples. Secondly, the loss function is the MSE function, the gradient descent method is Adam, and the Dropout regularization method is used to prevent overfitting.

[0037] The process can be described as follows:

[0038]

[0039] Among them, g Θ (·) denotes the nonlinear fitting operator, which is determined by continuously optimizing the parameter set Θ during training through a deep convolutional network based on LSTM. F0(w) represents the preprocessed amplitude spectrum of the seismic record.

[0040] Step 3 also includes generating a dataset for training and testing the network. The sample seismic wavelets used in the model data are: Ricker wavelet, generalized seismic wavelet, controlled source wavelet, and bandpass wavelet. The first two wavelets are single-peaked wavelets, while the latter two are non-single-peaked wavelets, and the wavelet amplitude spectra differ significantly in shape. The reflection coefficient sequences in the samples are Gaussian white noise random number sequences, blue noise sequences, and random number sequences following an α-stable distribution, respectively. The seismic wavelets and reflection coefficient sequences are convolved according to the convolution model to obtain synthetic seismic records. The time sampling interval is 1ms, and Gaussian white noise with a signal-to-noise ratio of 20dB is added. A total of 8000 channels of synthetic seismic records are generated and divided into training set, validation set, and test set in a ratio of 6:1:1 for training the deep neural network.

[0041] In step 4, the wavelet amplitude spectrum of the actual data is estimated; the preprocessed earthquake record amplitude spectrum in step 2) is input into the deep learning network in step 3) to obtain the output wavelet amplitude spectrum.

[0042] In step 5, the seismic wavelet amplitude spectrum obtained by the deep learning network can yield a shallow seismic wavelet. Based on the shallow wavelet, the deep seismic wavelet is estimated, and the deep wavelet is corrected using the ratio between the deep and shallow waves.

[0043] This deep learning-based seismic wavelet amplitude spectrum estimation method incorporating prior information further includes, after step 5, training the deep learning network using synthetic seismic records and comparing the estimated wavelet amplitude spectrum with the results obtained from traditional spectral simulation methods; for ease of comparison, an error function Δ is defined. ASSW To measure the similarity between the estimated wavelet spectrum and the true wavelet spectrum:

[0044]

[0045] In the formula, ASSW estimatedThe calculated sub-wavelength spectrum, ASSW true Represents the true sub-wave spectrum.

[0046] This invention relates to a machine learning-based method for seismic wavelet extraction, and more particularly to a deep learning-based method for estimating the amplitude spectrum of seismic wavelets by incorporating prior information. This method is then applied to the processing of seismic data. Starting directly from seismic data, this data-driven approach utilizes a deep learning network structure to accurately estimate the amplitude spectra of both non-single-peak and single-peak wavelets, guided by prior information about the seismic wavelets, without relying on any assumptions.

[0047] This invention, based on the prior knowledge of the smoothness of seismic wavelet spectra, preprocesses observed seismic data and then inputs the smoothed seismic record amplitude spectrum into a deep neural network based on LSTM. Through the nonlinear fitting effect of the deep network, the seismic wavelet amplitude spectrum is obtained. Without requiring any assumptions about reflection coefficients, this invention can not only accurately estimate the amplitude spectra of single-peak and non-single-peak seismic wavelets but also avoids the computational errors caused by parameter estimation and effective frequency band estimation in various polynomial fitting methods, thus enabling better high-resolution processing.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] The method of this invention makes full use of the prior information that the seismic wavelet spectrum is smooth. By inputting the preprocessed seismic record amplitude spectrum into a deep learning network, it can accurately estimate the amplitude spectrum of non-single-peak and single-peak wavelet without relying on any assumptions, thereby achieving high-resolution processing of seismic data with high fidelity. Attached Figure Description

[0050] Figure 1 A flowchart illustrating a specific embodiment of the deep learning seismic wavelet amplitude spectrum estimation method incorporating prior information according to the present invention;

[0051] Figure 2 This is a diagram of the DNN-estimated wavelet amplitude spectrum structure in a specific embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram illustrating the estimation of the wavelet amplitude spectrum of the reflection coefficient sequence following a white noise distribution in a specific embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram illustrating the estimation of wavelet amplitude spectrum when the reflection coefficient sequence follows an α-stable distribution in a specific embodiment of the present invention;

[0054] Figure 5This is an estimation of the amplitude spectrum of a controllable source wavelet in a specific embodiment of the present invention. (a) Schematic diagram of a controllable source wavelet. Detailed Implementation

[0055] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, 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.

[0056] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, and / or combinations thereof.

[0057] This invention discloses a deep learning-based method for estimating seismic wavelet amplitude spectra by incorporating prior information. First, based on the prior information that seismic wavelet spectra are smooth, the observed seismic data is preprocessed. Then, the smoothed seismic record amplitude spectrum is input into a deep neural network based on Long Short-Term Memory (LSTM). Through the nonlinear fitting effect of the deep network, the seismic wavelet amplitude spectrum is obtained. Comparison with traditional spectral simulation methods reveals that this method, without requiring any reflection coefficient assumptions, can not only accurately estimate the amplitude spectra of unimodal and non-unimodal seismic wavelets, but also reduces the influence of human factors and avoids computational errors caused by parameter estimation and effective frequency band estimation in various polynomial fitting methods. It exhibits good noise resistance, as demonstrated by model examples and real-world data testing.

[0058] The following are several specific embodiments of the application of the present invention.

[0059] Example 1

[0060] In a specific embodiment 1 of the present invention, the deep learning seismic wavelet amplitude spectrum estimation method incorporating prior information includes the following steps:

[0061] Step 1: Preprocess the raw seismic data to obtain post-stack seismic data; s(n)∈R N×1 , where R represents the real number field and N represents the number of sampling points for the seismic data;

[0062] Step 2: Perform smoothing preprocessing on the post-stack seismic data; based on the prior information of the smoothness of the seismic wavelet amplitude spectrum, the observation data is preprocessed as follows:

[0063] First, earthquake records can be represented in the time domain as follows:

[0064] s(n)=r(n)*w(n)+n(n) (1)

[0065] In the formula, s(n) represents the seismic record, w(n) represents the seismic wavelet, r(n) is the reflection coefficient sequence, * represents the convolution operation, and n(n) is random noise, where w(n), r(n), and n(n) all belong to L 2 (R).

[0066] Typically, the wavelet amplitude spectrum of a seismic event is obtained from the amplitude spectrum of a seismic record through nonlinear calculation. Therefore, by performing a Fourier transform on both sides of equation (1), the amplitude spectrum of the seismic record can be obtained as shown in equation (2).

[0067]

[0068] In the formula, S(w), W(w), R(w) and Let represent the spectra of s(n), w(n), r(n), and n(n), respectively.

[0069] Secondly, by integrating the cumulative distribution function, the irregular amplitude spectrum of the seismic record is smoothed, thus obtaining a relatively smooth seismic wavelet spectrum. The specific process is shown in equations (3) and (4):

[0070]

[0071]

[0072] Formula (3) normalizes the earthquake record amplitude spectrum |S(w)| obtained from formula (2), and then performs the integral operation in formula (4) on the normalized result f0(w). F0(w) represents the cumulative distribution function of the normalized result f0(w), and this operator plays the role of low-pass filtering.

[0073] Step 3, Constructing the Deep Learning Network Structure; This invention constructs a deep learning neural network (DNN) based on Long Short-Term Memory (LSTM), composed of LSTM units and fully connected layer units, such as... Figure 2 As shown, this structure avoids problems such as gradient vanishing compared to traditional recurrent neural networks (RNNs), and is therefore suitable for problems where the data has temporal correlation.

[0074] The input data is the amplitude spectrum of seismic records. To fully utilize the temporal correlation of the data and obtain more robust results, we input data points within a time window into an LSTM layer. After the data is input into the LSTM layer, each unit of the LSTM processes the input data and outputs a vector. This vector is then input into a fully connected neural network, which outputs a data point of the fitted data for the current time window. Finally, these individual data points are combined to obtain the estimated wavelet amplitude spectrum.

[0075] The objective function for optimizing the number of model parameters is shown in equation (5):

[0076]

[0077] Where W(w) represents the true seismic wavelet amplitude spectrum, The estimated wavelet amplitude spectrum is represented by , and n represents the number of samples. Secondly, the MSE function is chosen as the loss function, the Adam method is used for gradient descent, and Dropout regularization is employed to prevent overfitting.

[0078] The process can be described as follows:

[0079]

[0080] Where gΘ(·) represents the nonlinear fitting operator, which is determined by continuously optimizing the parameter set Θ during training through a deep convolutional network based on LSTM, and F0(w) represents the preprocessed amplitude spectrum of the seismic record.

[0081] The constructed deep network is a data-driven approach, therefore requiring the generation of datasets for training and testing. The model data in this invention uses sample seismic wavelets of the following types: Ricker wavelet, generalized seismic wavelet, controlled-source wavelet, and bandpass wavelet. The first two wavelets are unimodal wavelets, while the latter two are non-unimodal wavelets, and their amplitude spectra differ significantly in morphology. The reflection coefficient sequences in the samples are obtained using Gaussian white noise random number sequences, blue noise sequences, and random number sequences following an α-stable distribution. The seismic wavelets and reflection coefficient sequences are convolved using a convolution model to obtain synthetic seismic records. The time sampling interval is 1 ms, and Gaussian white noise with a signal-to-noise ratio of 20 dB is added. A total of 8000 channels of synthetic seismic records are generated and divided into training, validation, and test sets in a 6:1:1 ratio for training the deep neural network.

[0082] Step 4: Estimate the wavelet amplitude spectrum of the actual data to obtain the shallow wavelet; input the preprocessed seismic record amplitude spectrum from Step 2) into the deep learning network in Step 3) to obtain the output wavelet amplitude spectrum. The wavelet amplitude spectrum obtained by the deep learning network yields a shallow wavelet. Based on the shallow wavelet, the deep wavelet is then estimated, and the ratio between the deep and shallow waveslet is used to correct the deep wavelet.

[0083] Example 2

[0084] In a specific embodiment 2 of this invention, the extraction of seismic wavelets plays a crucial role in seismic data processing. Seismic wavelets are determined by the product of their amplitude and phase spectra; therefore, estimating their amplitude spectra is a critical step. Commonly used seismic wavelet amplitude spectrum extraction methods have limitations regarding reflection coefficient types and wavelet morphologies, hindering their widespread application in practical seismic data processing. This invention proposes a deep learning-based seismic wavelet amplitude spectrum estimation method incorporating prior information, capable of simultaneously estimating the amplitude spectra of both non-single-peak and single-peak wavelets, thereby enabling better high-resolution processing.

[0085] like Figure 1 As shown, the deep learning seismic wavelet amplitude spectrum estimation method incorporating prior information includes the following steps:

[0086] Step 101: Preprocess the raw seismic data to obtain post-stack seismic data;

[0087] Step 102: Perform smoothing preprocessing on the post-stack seismic data;

[0088] Based on the prior information of the smoothness of the seismic wavelet amplitude spectrum, the observation data is preprocessed.

[0089] First, earthquake records can be represented in the time domain as follows:

[0090] s(n)=r(n)*w(n)+n(n) (1)

[0091] In the formula, s(n) represents the seismic record, w(n) represents the seismic wavelet, r(n) is the reflection coefficient sequence, * represents the convolution operation, and n(n) is random noise, where w(n), r(n), and n(n) all belong to L 2 (R).

[0092] Typically, the wavelet amplitude spectrum of a seismic event is obtained from the amplitude spectrum of a seismic record through nonlinear calculation. Therefore, by performing a Fourier transform on both sides of equation (1), the amplitude spectrum of the seismic record can be obtained as shown in equation (2).

[0093]

[0094] In the formula, S(w), W(w), R(w) and Let represent the spectra of s(n), w(n), r(n), and n(n), respectively.

[0095] Secondly, by integrating the cumulative distribution function, the irregular amplitude spectrum of the seismic record is smoothed, thus obtaining a relatively smooth seismic wavelet spectrum. The specific process is shown in equations (3) and (4):

[0096]

[0097]

[0098] Formula (3) normalizes the earthquake record amplitude spectrum |S(w)| obtained from formula (2), and then performs the integral operation in formula (4) on the normalized result f0(w). F0(w) represents the cumulative distribution function of the normalized result f0(w), and this operator plays the role of low-pass filtering.

[0099] Step 103, Construct the deep learning network structure:

[0100] This invention constructs a deep learning neural network (DNN) based on Long Short-Term Memory (LSTM), consisting of LSTM units and fully connected layer units, as follows: Figure 2 As shown, this structure avoids problems such as gradient vanishing compared to traditional recurrent neural networks (RNNs), and is therefore suitable for problems where the data has temporal correlation.

[0101] The input data is the amplitude spectrum of seismic records. To fully utilize the temporal correlation of the data and obtain more robust results, we input data points within a time window into an LSTM layer. After the data is input into the LSTM layer, each unit of the LSTM processes the input data and outputs a vector. This vector is then input into a fully connected neural network, which outputs a data point of the fitted data for the current time window. Finally, these individual data points are combined to obtain the estimated wavelet amplitude spectrum.

[0102] The objective function for optimizing the number of model parameters is shown in equation (5):

[0103]

[0104] Where W(w) represents the true seismic wavelet amplitude spectrum, The estimated wavelet amplitude spectrum is represented by , and n represents the number of samples. Secondly, the MSE function is chosen as the loss function, the Adam method is used for gradient descent, and Dropout regularization is employed to prevent overfitting.

[0105] The process can be described as follows:

[0106]

[0107] Among them, g Θ (·) denotes the nonlinear fitting operator, which is determined by continuously optimizing the parameter set Θ during training through a deep convolutional network based on LSTM. F0(w) represents the preprocessed amplitude spectrum of the seismic record.

[0108] The constructed deep network is a data-driven approach, therefore, a dataset needs to be generated for training and testing the network. The sample seismic wavelets used in this paper's model data are: Ricker wavelet, generalized seismic wavelet, controlled-source wavelet, and bandpass wavelet, with parameter settings shown in Table 1.

[0109] Table 1 Seismic wavelet parameter settings

[0110]

[0111]

[0112] The first two wavelets are single-peaked wavelets, while the latter two are non-single-peaked wavelets, and their amplitude spectra differ significantly in morphology. The reflection coefficient sequences in the samples are generated using Gaussian white noise random number sequences, blue noise sequences, and random number sequences following an α-stable distribution, respectively. The blue noise can be simulated using white noise passed through an ARMA(1,1) system, which is selected here as follows:

[0113]

[0114] Furthermore, the parameters in the α-stable distribution are set to α = 1.2, σ = 0.008, and β = μ = 0.

[0115] The seismic wavelet and reflection coefficient sequence were convolved according to the convolution model to obtain the synthetic seismic record. The time sampling interval was 1ms, and Gaussian white noise with a signal-to-noise ratio of 20dB was added. A total of 8000 channels of synthetic seismic record were obtained, which were divided into training set, validation set and test set in a ratio of 6:1:1 for training deep neural network.

[0116] Step 104: Input the preprocessed seismic record amplitude spectrum from step 2) into the deep learning network in step 3) to obtain the output seismic wavelet amplitude spectrum.

[0117] Step 105: Based on actual data verification, the seismic wavelet amplitude spectrum obtained by the deep learning network can yield a shallow seismic wavelet, thereby correcting the deep wavelet.

[0118] Example 3

[0119] In a specific embodiment 3 of the present invention, a deep learning network is trained using synthetic seismic records, and the estimated wavelet amplitude spectrum is compared with the results obtained by the conventional spectral simulation method (SS method). For ease of comparison, an error function is defined to measure the similarity between the estimated wavelet spectrum and the true wavelet spectrum:

[0120]

[0121] In the formula, ASSW estimated The calculated sub-wavelength spectrum, ASSW true Represents the true sub-wave spectrum.

[0122] First, the effectiveness of the invention is verified using seismic signals synthesized from reflection coefficients that follow a white noise distribution. Figure 3 As shown. Figure 3 The amplitude spectrum of the wavelet is estimated for the reflection coefficient sequence following a white noise distribution. (a) Zero-phase Ricker wavelet; (b) Reflection coefficients following a Gaussian white noise distribution; (c) Synthetic seismic record; (d) Seismic wavelet amplitude spectrum estimation results obtained by the DNN method; (e) True wavelet amplitude spectrum and wavelet amplitude spectra estimated by the DNN method and the SS method, respectively; (f) Error function value between the true and estimated wavelet spectra. Figure 3 (e) shows the wavelet amplitude spectra estimated by the DNN method and the SS method, respectively. Figure 3 (f) Plot their error functions. It can be observed that... Figure 3 (b) assuming the reflection coefficient satisfies a Gaussian white noise distribution Figure 3 (e) The wavelet amplitude spectra obtained by the SS method and the DNN method proposed in this paper both show high consistency with the true wavelet amplitude spectra, and Figure 3 The error function values ​​in (f) all vary within a relatively small amplitude range. This indicates that when the reflection coefficient sequence follows a Gaussian white noise distribution, both methods can extract accurate wavelet amplitude spectra from the seismic record amplitude spectrum relatively well. However, the SS method is highly sensitive to the effective frequency range and location of the seismic record spectrum, and the fitting parameters in the polynomial are difficult to determine when the true wavelet is unknown, thus significantly limiting the estimation of the wavelet amplitude spectrum.

[0123] Secondly Figure 4 The results of wavelet amplitude spectrum estimation for seismic signals synthesized from reflection coefficients that follow a non-white noise distribution are presented. Figure 4For wavelet amplitude spectrum estimation when the reflection coefficient sequence follows an α-stable distribution: (a) zero-phase Ricker wavelet; (b) reflection coefficients following an α-stable distribution; (c) synthetic seismic record; (d) seismic wavelet amplitude spectrum estimation results obtained by the DNN method; (e) true wavelet amplitude spectrum and wavelet amplitude spectra estimated by the DNN method and the SS method, respectively; (f) error function value between the true and estimated wavelet spectra.

[0124] Figure 4 The reflection coefficient in (b) follows an α-stable distribution. Figure 4 (e) represents the wavelet amplitude spectra estimated by the DNN method and the SS method, respectively. (Comparison) Figure 4 (e) and Figure 3 (e) It can be observed that the wavelet spectrum estimated by the DNN method maintains a high degree of consistency with the true wavelet spectrum, while the fitting effect of the SS method is significantly worse. Figure 3 The wavelet spectrum estimated by the method in (e) exhibits a large fitting deviation in the high-frequency band of the wavelet spectrum. This shows that when the reflection coefficient sequence follows a non-white noise distribution, the DNN method can accurately estimate the seismic wavelet amplitude spectrum, while traditional methods cannot.

[0125] also, Figure 5 The estimation results of the amplitude spectrum of the non-single-peak wavelet are given. Figure 5 The estimation of the amplitude spectrum of a controllable source wavelet. (a) Controllable source wavelet; (b) Estimation results of the amplitude spectrum of a non-single-peak seismic wavelet obtained by the DNN method; (c) The actual wavelet amplitude spectrum and the wavelet amplitude spectra estimated by the DNN method and the SS method, respectively; (d) The error function value between the actual and estimated wavelet spectra. Figure 5 (a) is a controllable source wavelet, observed... Figure 5 The estimation results of (b)(c) and Figure 5 The error function value in (d) reveals that the traditional SS method, based on a statistical method and a polynomial fitting function, has significant limitations on the waveform of seismic wavelets. It can only estimate the amplitude spectrum of single-peak wavelets similar to Ricker wavelets, while the estimation of non-single-peak wavelet spectra is distorted. Therefore, this type of method has significant limitations on the shape of seismic wavelets. In contrast, the DNN method proposed in this invention is based on deep learning. Through data-driven nonlinear relationships implemented in neural networks, it can learn wavelets of various shapes and has better generalization ability.

[0126] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0127] Except for the technical features described in the specification, all other technologies are known to those skilled in the art.

Claims

1. A deep learning-based seismic wavelet amplitude spectrum estimation method incorporating prior information, characterized in that... This deep learning-based seismic wavelet amplitude spectrum estimation method, which incorporates prior information, includes: Step 1: Preprocess the raw seismic data to obtain post-stack seismic data; Step 2: Perform smoothing preprocessing on the post-stack seismic data; Step 3: Construct the deep learning network structure; Step 4: Estimate the wavelet amplitude spectrum of the actual data to obtain the shallow wavelet. Step 5: Verify with actual data to achieve correction of deep seismic wavelets; In step 3, a deep learning neural network (DNN) based on Long Short-Term Memory (LSTM) was constructed, consisting of LSTM units and fully connected layers. The input data was the amplitude spectrum of the seismic record. Data points within a time window were input into the LSTM layer. After the data was input into the LSTM layer, each unit of the LSTM processed the input data and output a vector. This vector was then input into a fully connected neural network, which output a data point of the fitted data for the current time window. Finally, these individual data points were combined to obtain the estimated wavelet amplitude spectrum. In step 3, the objective function for optimizing the model parameters is shown in equation (5): (5) in, Indicates the first A real seismic wavelet amplitude spectrum Indicates the first The estimated wavelet amplitude spectrum, , m represents the number of samples; secondly, the loss function is the MSE function, the gradient descent uses the Adam method, and the Dropout regularization method is used to prevent overfitting. The process can be described as follows: (6) in, This represents a nonlinear fitting operator, which continuously optimizes the parameter set during training using a deep convolutional network based on LSTM. The decision, This represents the amplitude spectrum of the preprocessed seismic record; Step 3 also includes generating a dataset for training and testing the network. The sample seismic wavelets used in the model data are: Ricker wavelet, generalized seismic wavelet, controlled source wavelet, and bandpass wavelet. The first two wavelets are single-peaked wavelets, while the latter two are non-single-peaked wavelets, and the wavelet amplitude spectra differ significantly in shape. The reflection coefficient sequences in the samples are Gaussian white noise random number sequences, blue noise sequences, and random number sequences following an α-stable distribution, respectively. The seismic wavelets and reflection coefficient sequences are convolved according to the convolution model to obtain synthetic seismic records. The time sampling interval is 1ms, and Gaussian white noise with a signal-to-noise ratio of 20dB is added. A total of 8000 channels of synthetic seismic records are generated and divided into training set, validation set, and test set in a ratio of 6:1:1 for training the deep neural network.

2. The deep learning seismic wavelet amplitude spectrum estimation method incorporating prior information as described in claim 1, characterized in that, In step 1, the raw seismic data is preprocessed to obtain post-stack seismic data: , where R represents the real number field and N represents the number of sampling points for the seismic data.

3. The deep learning seismic wavelet amplitude spectrum estimation method incorporating prior information as described in claim 1, characterized in that, In step 2, based on the prior information of the smoothness of the seismic wavelet amplitude spectrum, the observation data is preprocessed, including: Step 2a: Calculate the amplitude spectrum of the earthquake record; Step 2b involves using integral operations in the cumulative distribution function to smooth the irregular seismic record amplitude spectrum, thereby obtaining a relatively smooth seismic wavelet spectrum.

4. The deep learning seismic wavelet amplitude spectrum estimation method incorporating prior information as described in claim 3, characterized in that, In step 2a, the seismic record is represented in the time domain as follows: The seismic wavelet amplitude spectrum is obtained from the seismic record amplitude spectrum through nonlinear calculation. Therefore, by performing a Fourier transform on both sides of equation (1), the seismic record amplitude spectrum can be obtained as shown in equation (2): s(n)=r(n)*w(n)+n(n) (1) （2） In the formula, Let w(n) represent the earthquake record, and w(n) represent the seismic wavelet. For the reflection coefficient sequence, This represents the convolution operation. It is random noise. , , and They represent ,w(n), and The spectrum, where n represents the number of sampling points.

5. The deep learning seismic wavelet amplitude spectrum estimation method incorporating prior information as described in claim 4, characterized in that, In step 2b, the specific process is as shown in equations (3) and (4): (3) (4) Formula (3) yields the earthquake record amplitude spectrum from formula (2). Perform normalization processing, and then analyze the normalization results. Perform the integral operation in formula (4).

6. The deep learning seismic wavelet amplitude spectrum estimation method incorporating prior information as described in claim 1, characterized in that, In step 4, the seismic wavelet amplitude spectrum is estimated from the actual data; the preprocessed seismic record amplitude spectrum from step 2 is input into the deep learning network in step 3 to obtain the output seismic wavelet amplitude spectrum.

7. The deep learning seismic wavelet amplitude spectrum estimation method incorporating prior information as described in claim 1, characterized in that, In step 5, the seismic wavelet amplitude spectrum obtained by the deep learning network can yield a shallow seismic wavelet. Based on the shallow wavelet, the deep seismic wavelet is estimated, and the deep wavelet is corrected using the ratio between the deep and shallow waves.

8. The deep learning seismic wavelet amplitude spectrum estimation method incorporating prior information as described in claim 7, characterized in that, This deep learning-based seismic wavelet amplitude spectrum estimation method incorporating prior information further includes, after step 5, training the deep learning network using synthetic seismic records and comparing the estimated wavelet amplitude spectrum with the results obtained from traditional spectral simulation methods; for ease of comparison, an error function Δ is defined. ASSW To measure the similarity between the estimated wavelet spectrum and the true wavelet spectrum: (8) In the formula, This represents the calculated sub-wavelength spectrum. Represents the true sub-wave spectrum.

Images