Seismic intelligent high resolution processing method and system integrated with physical information

Abstract

This invention provides a high-resolution intelligent seismic processing method and system incorporating physical information. The processing method includes: acquiring stacked one-dimensional seismic data; preprocessing the one-dimensional seismic data to obtain a training set and estimating the initial non-attenuated wavelet; constructing a deep network model based on the absorption and attenuation characteristics of the one-dimensional seismic data; learning an attenuation compensation operator; and applying the attenuation compensation operator to non-stationary seismic records to obtain stationary seismic records. This method does not require known Q-values ​​and fully utilizes existing well logging data and non-attenuated wavelets, enabling the acquisition of more stable compensation results.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of seismic exploration, and particularly relates to a seismic intelligent high-resolution processing method and system integrated with physical information. BACKGROUND

[0002] Based on this model, methods for improving the resolution of non-stationary seismic records can be broadly categorized into two types. The first type uses piecewise deconvolution to improve the resolution of non-stationary seismic records. The second type is equivalent to applying the inverse process of attenuation to the seismic record, transforming the non-stationary seismic record into a stationary one, thereby improving the resolution and laying the foundation for subsequent deconvolution methods. The first type of method first divides the non-stationary seismic record into several stationary segments in the time domain using time windows, applies deconvolution to each segment separately, and then synthesizes the segments with improved resolution into a complete signal (Wang, 1969; Van der Ban, 2008). For example, Van der Ban (2008) proposed a statistical method for achieving maximum kurtosis by constant phase rotation, which can extract time-varying wavelets for frequency extension of seismic records. However, this type of method uses a fixed-length window and cannot adapt to the non-stationarity of the seismic signal. Margrave et al. (2001) proposed the Gabor transform to convert non-stationary signals to the time-frequency domain, which can directly extract the time spectrum of the attenuated wavelet and reflection coefficient. However, this method assumes the subsurface medium is a homogeneous viscoelastic medium, ignoring the stratification of the subsurface medium, and is only applicable to single Q values. Gao Jinghuai et al. (2009) constructed an adaptive non-uniform Gabor frame based on the characteristics of seismic records, broadening the frequency band in the time-frequency domain. This method is applicable to time-varying Q values. Hale (1981) proposed an inverse Q-filtering method using a series expansion. However, inverse Q-filtering suffers from amplitude compensation instability. Therefore, Wang (2002) started from the perspective of wavefield extension, assuming the strata are layered models with Q being constant for each layer, derived the formula for inverse Q-filtering, and stabilized the amplitude compensation operator, achieving stable and effective inverse Q-filtering. Wang (2006) further extended the above method to be applicable to cases where Q varies with depth or travel time, making it closer to the actual situation of the strata than the stratification method. However, in the presence of environmental noise, Wang (2002, 2006)'s stable inverse Q-filtering method undercompensates for high-frequency components in order to keep high-frequency compensation stable. Besides the two types of inverse Q-filtering methods mentioned above, many researchers have used inversion-based methods for inverse Q-filtering. Zhang et al. (2007) and Wang et al. (2014) transformed the seismic record compensation problem into an inverse problem and introduced Bayesian theory for solution. Wang Benfeng et al. (2014), based on Wang's (2002, 2006) Q-forward modeling formula, used the inversion idea to achieve stable and efficient seismic record attenuation compensation. Zhang et al. (2017) introduced synchronous compression transform and achieved attenuation compensation by solving the inverse problem of L1 norm as a regularization term in the time-frequency domain. Chen et al. (2019, 2021b), given the known Q value, directly inverted the reflection coefficient from non-stationary seismic records based on a non-stationary convolution model, and the convolution of the reflection coefficient with the seismic wavelet yielded a stationary seismic record.The second type of method requires known Q-values ​​as prior information, but estimating Q-values ​​is difficult, and the accuracy of the estimation affects the compensation effect. Margrave et al. (2011) extended the Gabor deconvolution method to time-varying Q-values ​​by replacing them with equivalent Q-values. Wang et al. (2014) improved the accuracy of the adaptive partitioning method for representing non-stationary seismic records by allowing the analysis window to adapt to the frequency domain characteristics of non-stationary seismic records. The above methods do not require known Q-values ​​and treat the signal as piecewise stationary.

[0003] The existing technology has the following drawbacks:

[0004] 1) The piecewise deconvolution method requires the seismic record to be piecewise stationary, and the way it is piecewise divided affects the accuracy of the representation of non-stationary seismic records.

[0005] 2) The method of applying the inverse process of attenuation to non-stationary seismic records requires known Q value, which is difficult to estimate, and the method has poor stability. Summary of the Invention

[0006] In view of the above problems, the present invention is proposed to provide an intelligent high-resolution earthquake processing method and system that incorporates physical information to overcome or at least partially solve the above problems.

[0007] According to one aspect of the present invention, an intelligent high-resolution earthquake processing method incorporating physical information is provided, the processing method comprising:

[0008] Acquire stacked one-dimensional seismic data, preprocess the one-dimensional seismic data to obtain a training set, and estimate the initial non-attenuated wavelet;

[0009] Based on the absorption and attenuation characteristics of the one-dimensional seismic data, a deep network model is constructed.

[0010] Learning attenuation compensation operator;

[0011] The attenuation compensation operator is applied to the non-stationary seismic record to obtain a stationary seismic record.

[0012] Optionally, the acquisition of stacked one-dimensional seismic data, the preprocessing of the one-dimensional seismic data to obtain a training set, and the estimation of the initial non-attenuated wavelet specifically include:

[0013] Obtain a set of one-dimensional seismic data after stacking, denoted as s non Based on existing well logging data, the non-attenuated wavelet w is estimated from shallow seismic data, and the elastic wave impedance z in the well logging data is used. According to the Robison model, the well logging data and the non-attenuated wavelet w are convolved to obtain a stationary seismic record.

[0014]

[0015] s sta (t)=w(t)*r(t) (2)

[0016] Where t represents time, r represents the reflection coefficient sequence calculated from impedance, δ represents the unit impulse function, and s sta This represents a set of one-dimensional stationary seismic records. The seismic records synthesized from well logging and their corresponding well-side traces are treated as a label pair, and these label pairs are divided into training datasets. and validation dataset

[0017] Optionally, the step of building a deep network model based on the absorption and attenuation characteristics of the one-dimensional seismic data specifically includes:

[0018] Based on the absorption and attenuation characteristics of earthquake records, a deep network model is built using gated recurrent units, convolution, and pooling.

[0019] Optionally, the step of building a deep network model using gated recurrent units, convolution, and pooling based on the absorption and attenuation characteristics of seismic records specifically includes:

[0020] A deep network is built, and its nonlinear mapping operator is learned to compensate for attenuation in nonstationary seismic records. To represent a deep network, non-stationary seismic records and known estimated wavelets are simultaneously fed into the deep network as input:

[0021]

[0022] Deep networks learn operators Where Θ represents the set of parameters in the deep network, and A represents the operator of the forward modeling process of absorption attenuation. The pseudo-inverse operator of A represents the inverse process of recovering attenuated seismic records.

[0023] Optionally, the learning attenuation compensation operator specifically includes: performing semi-supervised training on the deep network model based on the principle of autoencoders to learn the optimal attenuation compensation operator.

[0024] Optionally, the step of performing semi-supervised training on the deep network model based on the autoencoder principle to learn the optimal attenuation compensation operator specifically includes:

[0025] Design autoencoders and decoders;

[0026] The entire autoencoder section is equivalent to a cascaded connection of two deep networks. It has two input channels, one for the known wavelet w and the other for non-stationary seismic records. The output is a one-dimensional non-stationary seismic record. Labeled pairs with well logging data are removed, and the remaining unlabeled seismic observation data is denoted as s.non The wavelet is input into the entire autoencoder along with the encoder for constraint. Therefore, the loss function of the entire semi-supervised training method is divided into three parts: encoder, decoder, and autoencoder, and is written as:

[0027]

[0028] The loss function of the encoder is represented by labeled non-stationary seismic records as input, and the corresponding inversion target is known stationary seismic records.

[0029] This represents the loss function of the decoder, which is the exact reverse of the encoder's input and output.

[0030] The loss function of the autoencoder is represented by the following: the overall input and output of the autoencoder are guaranteed to be consistent, and unlabeled data is used to construct the loss function.

[0031] The loss functions are then integrated to obtain the final loss function, where α, β, and γ are constants:

[0032]

[0033] Ideally, the optimal set of network parameters is obtained:

[0034]

[0035] Optionally, the design of the autoencoder and decoder specifically includes:

[0036] The autoencoder is designed with 2 input channels, which are known wavelet w and non-stationary seismic records, respectively. Therefore, the size of the input data is set to N×2, where N represents the length of a seismic record in the time dimension and 2 represents the number of input channels. The encoder output target is the stationary seismic record of the corresponding well logging synthesis, with a size of N×1, which is the stationary seismic record in the training label pair.

[0037] The deep network of the decoder has one input channel, and each channel contains a known wavelet w.

[0038] The input is a stationary seismic record, but the parameters of these two parts are not shared. The input is a stationary seismic record, and the decoder outputs a stationary seismic record synthesized from the corresponding well log, with a size of N×1, which is the stationary seismic record in the training label pair.

[0039] Optionally, applying the attenuation compensation operator to the non-stationary seismic record to obtain a stationary seismic record specifically includes:

[0040] The attenuation compensation operator obtained from training Applying this to the validation dataset, it calculates the similarity between the compensated results and the labels based on the Pearson coefficient:

[0041]

[0042] σ1 and σ2 represent y1 and y2 respectively. The standard deviation is given by cov, which represents the calculation of covariance, specifically expressed as cov(X,Y)=E[(XE(X))(YE(Y)).

[0043] The similarity coefficient is used to determine whether it meets the requirements. If it does, it means that the method is feasible. The attenuation compensation operator is applied to the entire dataset to obtain high-quality, high-resolution seismic data.

[0044]

[0045] This invention also provides an intelligent high-resolution earthquake processing system incorporating physical information, applying the aforementioned intelligent high-resolution earthquake processing method incorporating physical information. The processing system includes:

[0046] The seismic data acquisition module is used to collect stacked one-dimensional seismic data, preprocess the one-dimensional seismic data to obtain a training set, and estimate the initial non-attenuated wavelet.

[0047] A deep network construction module is used to build a deep network model based on the absorption and attenuation characteristics of the one-dimensional seismic data.

[0048] A semi-supervised training module is used to learn the decay compensation operator;

[0049] The attenuation compensation module is used to apply the attenuation compensation operator to the non-stationary seismic record to obtain a stationary seismic record.

[0050] Optionally, the deep network construction module specifically includes:

[0051] Deep network structure building unit, used to build deep network structures suitable for attenuation compensation problems;

[0052] The attenuation compensation operator calculation unit is used to calculate the attenuation compensation operator.

[0053] This invention provides a high-resolution intelligent seismic processing method and system incorporating physical information. The processing method includes: acquiring stacked one-dimensional seismic data; preprocessing the one-dimensional seismic data to obtain a training set and estimating the initial non-attenuated wavelet; constructing a deep network model based on the absorption and attenuation characteristics of the one-dimensional seismic data; learning an attenuation compensation operator; and applying the attenuation compensation operator to non-stationary seismic records to obtain stationary seismic records. This method does not require known Q-values ​​and fully utilizes existing well logging data and non-attenuated wavelets, enabling the acquisition of more stable compensation results.

[0054] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 A flowchart of an intelligent high-resolution earthquake processing method incorporating physical information, provided in an embodiment of the present invention;

[0057] Figure 2 This is a schematic diagram of the deep network structure used in this invention;

[0058] Figure 3 This is a schematic diagram of the training process of the intelligent processing method of the present invention. Detailed Implementation

[0059] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0060] The terms "comprising" and "having," and any variations thereof, in the specification, embodiments, claims, and drawings of this invention are intended to cover non-exclusive inclusion, such as including a series of steps or units.

[0061] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0062] Example 1

[0063] The purpose of this invention is to provide an intelligent high-resolution earthquake processing method and system that incorporates physical information. This method can effectively solve the technical problem in the prior art that requires known or estimated Q values ​​or the classification of non-stationary earthquake records to improve the accuracy of earthquake record representation. This method can obtain high-quality, high-resolution earthquake data.

[0064] To achieve the above objectives, this invention discloses an intelligent high-resolution seismic processing method and system incorporating physical information. Addressing the absorption and attenuation phenomenon in post-stack seismic data caused by the viscoelasticity of the medium within the strata, this method transforms seismic absorption and attenuation compensation into resolving the nonlinear mapping between non-stationary and stationary seismic data. It utilizes a deep network to learn the nonlinear operator for attenuation compensation and estimates the unattenuated seismic wavelet. This wavelet, along with the non-stationary seismic data, is used as input to the deep network. A nonlinear compensation operator, represented by network parameters, is then semi-supervised and applied to attenuated seismic records to obtain stationary, high-resolution seismic data.

[0065] The intelligent high-resolution earthquake processing method incorporating physical information includes the following steps:

[0066] 1) Obtain the stacked one-dimensional observation seismic record, preprocess the one-dimensional seismic data to obtain the training set and estimate the initial non-attenuated wavelet;

[0067] 2) Based on the absorption and attenuation characteristics of seismic records, a deep network model is built using gated recurrent units, convolution, and pooling.

[0068] 3) Based on the principle of autoencoders, attenuation compensation operators are learned in a semi-supervised manner;

[0069] 4) Apply the attenuation compensation operator to the non-stationary seismic record to obtain a stationary seismic record.

[0070] Step 1) The specific operation is as follows:

[0071] Obtain a set of one-dimensional seismic data after stacking, denoted as s non The wavelet estimated from shallow seismic data based on existing well logging data is denoted as w and is considered unattenuated. The elastic wave impedance in the well logging data is denoted as z. According to the Robison model, the well logging data and the unattenuated wavelet w are convolved to obtain a stationary seismic record.

[0072]

[0073] s sta (t)=w(t)*r(t) (2)

[0074] Where t represents time, r represents the reflection coefficient sequence calculated from impedance, δ represents the unit impulse function, and s sta This represents a set of one-dimensional stationary seismic records. The synthesized seismic records from well logging and their corresponding well-side traces are treated as a label pair. The label pairs are then divided into training datasets as needed. and validation dataset

[0075] Step 2) The specific operation is as follows:

[0076] A deep network is constructed that uses a learned nonlinear mapping operator to achieve attenuation compensation for nonstationary seismic records. To represent this network, non-stationary seismic records and known estimated wavelets are simultaneously fed into the depth network as input:

[0077]

[0078] Operators are obtained through deep network learning. Where Θ represents the set of parameters in the deep network, and A represents the operator of the forward modeling process of absorption attenuation. This represents the inverse process of recovering attenuated seismic records, i.e., the pseudo-inverse operator of A.

[0079] Step 3) The specific operation is as follows: The specific training process is as follows: Figure 3 As shown.

[0080] An autoencoder is designed with two input channels: one for a known wavelet w and the other for non-stationary seismic records. The input data size is set to N×2, where N represents the length of a seismic record in the time dimension and 2 represents the number of input channels. The encoder output is a stationary seismic record synthesized from the corresponding well logs, with a size of N×1, representing the stationary seismic record in the training label pair. The decoder's deep network structure is identical to the encoder except for having only one input channel, but the parameters are not shared. It takes stationary seismic records as input and outputs the same stationary seismic record synthesized from the corresponding well logs, with a size of N×1, representing the stationary seismic record in the training label pair. The entire autoencoder is essentially a "concatenation" of the two deep networks, with two input channels: one for a known wavelet w and the other for non-stationary seismic records, and outputting a one-dimensional non-stationary seismic record. Excluding the label pairs with well log data, the remaining unlabeled seismic observation data is denoted as s. non The wavelet is input into the entire autoencoder along with the encoder for constraint. Therefore, the loss function of the entire semi-supervised training method is divided into three parts: encoder, decoder, and autoencoder, and is written as:

[0081]

[0082] This represents the loss function of the encoder part, which takes labeled non-stationary seismic records as input and the corresponding inversion target is known stationary seismic records.

[0083] This represents the loss function of the decoder, which is the exact reverse of the encoder's input and output. The loss function of the autoencoder is defined such that the overall input and output of the autoencoder are consistent. Therefore, we use unlabeled data to construct this part of the loss function.

[0084] Integrating the above loss functions, we obtain the final loss function (where α, β, and γ are constants):

[0085]

[0086] Ideally, the optimal set of network parameters is obtained:

[0087]

[0088] In step 4), the attenuation compensation operator obtained from training is... Applying this to the validation dataset, it calculates the similarity between the compensation result of the validation dataset and its labels based on the Pearson coefficient:

[0089]

[0090] σ1 and σ2 represent y1 and y2 respectively. The standard deviation is given, and cov represents the calculated covariance, specifically expressed as cov(X,Y)=E[(XE(X))(YE(Y)). The calculated similarity coefficient is used to determine if it meets the requirements. If it does, the method is feasible. The calculated attenuation compensation operator is then applied to the entire dataset to obtain high-quality, high-resolution seismic data.

[0091]

[0092] The present invention also discloses an intelligent high-resolution earthquake processing system that integrates physical information, comprising:

[0093] The seismic data acquisition module is used to preprocess post-stack seismic data to obtain datasets and non-attenuated wavelets, as well as training and validation datasets.

[0094] The deep network building module is used to construct deep network solution structures suitable for attenuation compensation problems, and to compute attenuation compensation operators.

[0095] The semi-supervised training module is used to perform semi-supervised training on deep network training models, optimize the network, and obtain the best decay compensation operator.

[0096] The attenuation compensation module uses the optimal compensation operator to test the method on the validation dataset and compensates for all datasets to obtain the processed seismic data y.

[0097] Example 2

[0098] The present invention discloses an intelligent high-resolution seismic processing technology incorporating physical information. Addressing the absorption and attenuation phenomenon in post-stack seismic data caused by the viscoelasticity of the medium within the strata, this method transforms seismic absorption and attenuation compensation into resolving the nonlinear mapping between non-stationary and stationary seismic data. It utilizes a deep network to learn the nonlinear operator for attenuation compensation and estimates the unattenuated seismic wavelet. This wavelet, along with the non-stationary seismic data, is used as input to the deep network. A nonlinear compensation operator, represented by network parameters, is then semi-supervised and applied to attenuated seismic records to obtain stationary, high-resolution seismic data.

[0099] like Figure 1 As shown, the specific steps include:

[0100] Obtain a set of one-dimensional seismic data after stacking, denoted as s non And preprocess the post-stack seismic data:

[0101] Obtain a set of one-dimensional seismic data after stacking, denoted as s non The non-attenuated wavelet estimated from shallow seismic data based on existing well logging data is denoted as w, and the elastic wave impedance in the well logging data is denoted as z. According to the Robison model, the well logging data and the non-attenuated wavelet w are convolved to obtain a stationary seismic record:

[0102]

[0103] s sta (t)=w(t)*r(t) (11)

[0104] Where t represents time, r represents the reflection coefficient sequence calculated from impedance, δ represents the unit impulse function, and s sta This represents a one-dimensional stationary seismic record. The seismic record synthesized from well logging and its corresponding well-side trace are treated as a label pair. The label pairs are then divided into training datasets as needed. and validation dataset

[0105] Based on the absorption and attenuation characteristics of seismic records, a deep network model is constructed using gated recurrent units, convolution, and pooling: A schematic diagram of the deep network model is shown below. Figure 2 As shown.

[0106] A deep network is constructed that uses a learned nonlinear mapping operator to achieve attenuation compensation for nonstationary seismic records. To represent this network, non-stationary seismic records and known estimated wavelets are simultaneously fed into the depth network as input:

[0107]

[0108] Operators are obtained through deep network learning. Where Θ represents the set of parameters in the deep network, and A represents the operator of the forward modeling process of absorption attenuation. This represents the inverse process of recovering attenuated seismic records, i.e., the pseudo-inverse operator of A.

[0109] Based on the principle of autoencoders, the attenuation compensation operator is learned in a semi-supervised manner:

[0110] An autoencoder is designed with two input channels: one for a known wavelet w and the other for non-stationary seismic records. The input data size is set to N×2, where N represents the length of a seismic record in the time dimension and 2 represents the number of input channels. The encoder output is a stationary seismic record synthesized from the corresponding well logs, with a size of N×1, representing the stationary seismic record in the training label pair. The decoder's deep network structure is identical to the encoder except for having only one input channel, but the parameters are not shared. It takes stationary seismic records as input and outputs the same stationary seismic record synthesized from the corresponding well logs, with a size of N×1, representing the stationary seismic record in the training label pair. The entire autoencoder is essentially a "concatenation" of the two deep networks, with two input channels: one for a known wavelet w and the other for non-stationary seismic records, and outputting a one-dimensional non-stationary seismic record. Excluding the label pairs with well log data, the remaining unlabeled seismic observation data is denoted as s. non The wavelet is input into the entire autoencoder along with the encoder for constraint. Therefore, the loss function of the entire semi-supervised training method is divided into three parts: encoder, decoder, and autoencoder, and is written as:

[0111]

[0112] This represents the loss function of the encoder part, which takes labeled non-stationary seismic records as input and the corresponding inversion target is known stationary seismic records.

[0113] This represents the loss function of the decoder, which is the exact reverse of the encoder's input and output. This represents the loss function of the autoencoder. As long as the overall input and output of the autoencoder are consistent, unlabeled data is used to construct this part of the loss function.

[0114] Integrating the above loss functions, we obtain the final loss function (where α, β, and γ are constants):

[0115]

[0116] Ideally, the optimal set of network parameters is obtained:

[0117]

[0118] Applying the attenuation compensation operator to non-stationary seismic records yields stationary seismic records:

[0119] The attenuation compensation operator obtained from training Applying this to the validation dataset, it calculates the similarity between the compensation result of the validation dataset and its labels based on the Pearson coefficient:

[0120]

[0121] σ1 and σ2 represent y1 and y2 respectively. The standard deviation is given, and cov represents the calculated covariance, specifically expressed as cov(X,Y)=E[(XE(X))(YE(Y)). The calculated similarity coefficient is used to determine if it meets the requirements. If it does, the method is feasible. The calculated attenuation compensation operator is then applied to the entire dataset to obtain high-quality, high-resolution seismic data.

[0122]

[0123] Beneficial Effects: The intelligent high-resolution seismic processing method incorporating physical information disclosed in this invention addresses the absorption and attenuation phenomenon in post-stack seismic data caused by the viscoelasticity of the medium in the formation. This method transforms seismic absorption and attenuation compensation into resolving the nonlinear mapping between non-stationary and stationary seismic data. It utilizes a deep network to learn the nonlinear operator for attenuation compensation and estimates the seismic non-attenuated wavelet. This wavelet, along with the non-stationary seismic data, is used as input to the deep network. A nonlinear compensation operator, represented by network parameters, is obtained through semi-supervised learning and applied to attenuated seismic records to obtain stationary, high-resolution seismic data. Compared to non-stationary deconvolution methods, the proposed method does not require consideration of the representation accuracy of non-stationary seismic records based on the partitioning method. Compared to inverse Q-filtering methods, the proposed method does not require known Q-values ​​and fully utilizes existing well logging data and the non-attenuated wavelet, resulting in more stable compensation results.

[0124] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A high-resolution intelligent earthquake processing method incorporating physical information, characterized in that, The processing method includes: Acquire stacked one-dimensional seismic data, preprocess the one-dimensional seismic data to obtain a training set, and estimate the initial non-attenuated wavelet; Based on the absorption and attenuation characteristics of the one-dimensional seismic data, a deep network model is constructed. Learning attenuation compensation operator; The attenuation compensation operator is applied to the non-stationary seismic record to obtain a stationary seismic record.

2. The intelligent high-resolution earthquake processing method incorporating physical information according to claim 1, characterized in that, The acquisition of stacked one-dimensional seismic data, the preprocessing of the one-dimensional seismic data to obtain a training set, and the estimation of the initial non-attenuated wavelet specifically include: Obtain a set of one-dimensional seismic data after stacking, denoted as s non Based on existing well logging data, the non-attenuated wavelet w is estimated from shallow seismic data, and the elastic wave impedance z in the well logging data is used. According to the Robison model, the well logging data and the non-attenuated wavelet w are convolved to obtain a stationary seismic record. s sta (t)=w(t)*r(t) (2) Where t represents time, r represents the reflection coefficient sequence calculated from impedance, δ represents the unit impulse function, and s sta This represents a set of one-dimensional stationary seismic records. The seismic records synthesized from well logging and their corresponding well-side traces are treated as a label pair, and these label pairs are divided into training datasets. and validation dataset 3. The intelligent high-resolution earthquake processing method incorporating physical information according to claim 1, characterized in that, The construction of the deep network model based on the absorption and attenuation characteristics of the one-dimensional seismic data specifically includes: Based on the absorption and attenuation characteristics of earthquake records, a deep network model is built using gated recurrent units, convolution, and pooling.

4. The intelligent high-resolution earthquake processing method incorporating physical information according to claim 3, characterized in that, The method of building a deep network model based on the absorption and attenuation characteristics of seismic records using gated recurrent units, convolution, and pooling specifically includes: A deep network is built, and its nonlinear mapping operator is learned to compensate for attenuation in nonstationary seismic records. To represent a deep network, non-stationary seismic records and known estimated wavelets are simultaneously fed into the deep network as input: Deep networks learn operators Where Θ represents the set of parameters in the deep network, and A represents the operator of the forward modeling process of absorption attenuation. The pseudo-inverse operator of A represents the inverse process of recovering attenuated seismic records.

5. The intelligent high-resolution earthquake processing method incorporating physical information according to claim 1, characterized in that, The learning attenuation compensation operator specifically includes: based on the principle of autoencoders, performing semi-supervised training on the deep network model to learn the optimal attenuation compensation operator.

6. The intelligent high-resolution earthquake processing method incorporating physical information according to claim 5, characterized in that, The process of semi-supervised training of a deep network model based on the autoencoder principle to learn the optimal attenuation compensation operator specifically includes: Design autoencoders and decoders; The entire autoencoder section is equivalent to a cascaded connection of two deep networks. It has two input channels, one for the known wavelet w and the other for non-stationary seismic records. The output is a one-dimensional non-stationary seismic record. Labeled pairs with well logging data are removed, and the remaining unlabeled seismic observation data is denoted as s. non The wavelet is input into the entire autoencoder along with the encoder for constraint. Therefore, the loss function of the entire semi-supervised training method is divided into three parts: encoder, decoder, and autoencoder, and is written as: The loss function of the encoder is represented by labeled non-stationary seismic records as input, and the corresponding inversion target is known stationary seismic records. This represents the loss function of the decoder, which is the exact reverse of the encoder's input and output. The loss function of the autoencoder is represented by the following: the overall input and output of the autoencoder are guaranteed to be consistent, and unlabeled data is used to construct the loss function. The loss functions are then integrated to obtain the final loss function, where α, β, and γ are constants: Ideally, the optimal set of network parameters is obtained:

7. The intelligent high-resolution earthquake processing method incorporating physical information according to claim 6, characterized in that, The designed autoencoder and decoder specifically include: The autoencoder is designed with 2 input channels, which are known wavelet w and non-stationary seismic records, respectively. Therefore, the size of the input data is set to N×2, where N represents the length of a seismic record in the time dimension and 2 represents the number of input channels. The encoder output target is the stationary seismic record of the corresponding well logging synthesis, with a size of N×1, which is the stationary seismic record in the training label pair. The deep network of the decoder has one input channel, which consists of a known wavelet w and non-stationary seismic records. However, the parameters of these two parts are not shared. The input is a stationary seismic record, and the output target of the decoder is the corresponding well-logged synthesized stationary seismic record with a size of N×1, which is the stationary seismic record in the training label pair.

8. The intelligent high-resolution earthquake processing method incorporating physical information according to claim 1, characterized in that, The step of applying the attenuation compensation operator to the non-stationary seismic record to obtain a stationary seismic record specifically includes: The attenuation compensation operator obtained from training Applying this to the validation dataset, it calculates the similarity between the compensated results and the labels based on the Pearson coefficient: σ1 and σ2 represent y1 and y2 respectively. The standard deviation is given by cov, which represents the calculation of covariance, specifically expressed as cov(X,Y)=E[(XE(X))(YE(Y)). The similarity coefficient is used to determine whether it meets the requirements. If it does, it means that the method is feasible. The attenuation compensation operator is applied to the entire dataset to obtain high-quality, high-resolution seismic data.

9. A high-resolution intelligent earthquake processing system incorporating physical information, employing the high-resolution intelligent earthquake processing method incorporating physical information as described in any one of claims 1-8, characterized in that, The processing system includes: The seismic data acquisition module is used to collect stacked one-dimensional seismic data, preprocess the one-dimensional seismic data to obtain a training set, and estimate the initial non-attenuated wavelet. A deep network construction module is used to build a deep network model based on the absorption and attenuation characteristics of the one-dimensional seismic data. A semi-supervised training module is used to learn the decay compensation operator; The attenuation compensation module is used to apply the attenuation compensation operator to the non-stationary seismic record to obtain a stationary seismic record.

10. The intelligent high-resolution earthquake processing system incorporating physical information according to claim 9, characterized in that, The deep network construction module specifically includes: Deep network structure building unit, used to build deep network structures suitable for attenuation compensation problems; The attenuation compensation operator calculation unit is used to calculate the attenuation compensation operator.

Images