Seismic data weak signal extraction method based on AI analysis

Abstract

The invention discloses a seismic data weak signal extraction method based on AI analysis, and belongs to the technical field of seismic exploration monitoring, and the method comprises the following steps: S1, carrying out the empirical mode decomposition of a collected seismic original signal containing noise, and obtaining a plurality of IMF components and a residual component; s2, calculating the permutation entropy of each IMF component; s3, screening an IMF component; s4, reconstructing an IMF component to obtain a seismic data weak signal; s5, expanding the number of samples of the seismic data weak signals; s6, constructing a seismic data weak signal extraction model, and training the seismic data weak signal extraction model based on the expanded seismic data weak signal until convergence; s7, extracting a seismic data weak signal based on a convergent seismic data weak signal extraction model; according to the method, the sample data volume of the seismic data weak signal can be expanded, and the training effect of the seismic data weak signal extraction model is improved, namely, the accuracy of the seismic data weak signal extraction model for extracting the seismic data weak signal is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of seismic exploration and monitoring, and specifically relates to a method for extracting weak signals from seismic data based on AI analysis. Background Art

[0002] Weak signals in seismic data refer to those effective signal data with low energy and small amplitude that are difficult to directly identify in the time domain, frequency domain or other characteristic domains during the seismic signal acquisition process.

[0003] Weak signals in seismic data often contain key information about deep geological structures, micro-earthquake activities, etc. Accurate extraction is of vital importance for geological structure research, earthquake disaster prediction, and oil and gas resource exploration.

[0004] With the continuous development of signal processing technology and artificial intelligence, a variety of seismic data weak signal extraction models have been applied to the field of seismic data weak signal extraction. However, due to the small number of samples of seismic data weak signals at present, the training volume of various seismic data weak signal extraction models is insufficient, resulting in poor accuracy of seismic data weak signal extraction of various seismic data weak signal extraction models.

[0005] In view of this, a weak signal extraction method for seismic data based on AI analysis is designed to solve the above problems. Summary of the Invention

[0006] To solve the problems raised in the above background technology, the present invention provides a method for extracting weak signals from seismic data based on AI analysis, which has the characteristics of being able to expand the sample data volume of weak signals from seismic data and improve the training effect of the weak signal extraction model of seismic data, that is, to improve the accuracy of the weak signal extraction model in extracting weak signals from seismic data.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for extracting weak signals from seismic data based on AI analysis, comprising the following steps:

[0008] S1: Perform empirical mode decomposition on the collected seismic original signal containing noise to obtain several IMF components and a residual component;

[0009] S2: Calculate the permutation entropy of each IMF component;

[0010] S3: Preset the permutation entropy threshold and filter the IMF components based on the permutation entropy threshold;

[0011] S4: reconstruct the filtered IMF components to obtain weak signals of seismic data;

[0012] S5: Based on the weak signal of seismic data, the number of samples of weak signal of seismic data is expanded through variational autoencoder;

[0013] S6: constructing a seismic data weak signal extraction model, and training the seismic data weak signal extraction model based on the expanded seismic data weak signal until convergence;

[0014] S7: Extract weak signals from seismic data based on the converged seismic data weak signal extraction model.

[0015] Furthermore, the specific steps of step S1 include:

[0016] Determine all local maximum and minimum points of the original seismic signal x(t) containing noise;

[0017] The upper envelope e is obtained by cubic spline interpolation. max (t) and the lower envelope e min (t);

[0018] Calculate the mean m1(t) of the upper and lower envelopes:

[0019]

[0020] Where: e max (t) represents the upper envelope; e min (t) represents the lower envelope;

[0021] Subtract the mean from the original signal to get h1(t):

[0022] h1(t)=x(t)-m1(t)

[0023] Where: x(t) represents the original seismic signal containing noise; m1(t) represents the mean of the upper and lower envelopes;

[0024] Determine whether h1(t) meets the two conditions of the IMF component. If so, h1(t) is the first IMF component c1(t). Otherwise, h1(t) is used as the new original signal.

[0025] Repeat the above process and get h after k screenings. k (t), so that it satisfies the IMF component condition, that is, c1(t)=h k (t);

[0026] Subtract the first IMF component c1(t) from the original signal x(t) to obtain the residual signal r(t);

[0027] Take r(t) as the new original signal and repeat the above decomposition process to obtain multiple IMF components in sequence until the remaining signal r n (t) It becomes a monotonic function or constant and cannot be decomposed into IMF components;

[0028] That is, decompose the original signal x(t) into a series of IMF components c i (t), i=1,2,…,n, and a residual component r n (t).

[0029] Furthermore, the specific steps of step S2 include:

[0030] Each IMF component c i (t) Reconstruct the phase space according to the embedding dimension m and time delay τ to obtain N-(m-1)τ m-dimensional vectors X i =(c i (t),c i+τ (t),…,c i+(m-1)τ (t)),i=1,2,…,N-(m-1)τ;

[0031] For each m-dimensional vector X i The elements in are sorted from small to large, and the sorted index series j1, j2, ..., j is obtained. m ;

[0032] Get n different permutations and combinations based on the sorted index sequence;

[0033] Count the probability p of each permutation combination appearing in all reconstruction directions n ,n=1,2,…,n;

[0034] Calculate the permutation entropy H according to the Shannon entropy formula p :

[0035]

[0036] Where: n represents the nth permutation and combination; p n Represents the probability of each permutation combination appearing in all reconstruction directions.

[0037] Furthermore, the specific steps of step S4 include:

[0038] For the selected IMF component c i (t) is superimposed to obtain the reconstructed signal x i (t):

[0039] x i (t)=∑ i∈S c i (t)

[0040] Where: c i (t) represents the IMF component; S represents the permutation entropy value H p Greater than the permutation entropy threshold H p,i The indexed set of IMF components.

[0041] Furthermore, the specific steps of step S5 include:

[0042] Variational Autoencoder Encoder: Reconstructs the input signal x i (t) is mapped to the latent space and outputs the mean vector μ and the logarithmic variance vector logσ 2 ;

[0043] Variational autoencoder decoder: Sample a vector z from the latent space, z = μ + ∈ × σ, ∈ represents the noise vector sampled from the standard normal distribution, and decode it into the reconstructed signal

[0044] The loss function in the above process consists of reconstruction loss and KL divergence loss:

[0045] L=L i +λL j

[0046] Where: L i Represents the reconstruction loss; λ represents the weight coefficient for balancing the two losses, which is usually adjusted through experiments; L j represents KL divergence loss;

[0047] The reconstruction loss is a measure of the reconstruction signal and reconstructed signal x i Differences in (t):

[0048]

[0049] Where: Represents the reconstructed signal x i (t) takes the expected value; x i (t) represents the reconstructed signal; represents the reconstruction signal;

[0050] The KL divergence loss constrains the distribution of the latent space to be close to the standard normal distribution, ensuring the continuity and sampleability of the latent space:

[0051]

[0052] Where: μ j represents the mean vector; represents the log-variance vector.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] The present invention first performs empirical mode decomposition on the collected noisy original seismic signal to obtain several IMF components, then calculates the permutation entropy for each IMF component, then screens the IMF components based on a preset permutation entropy threshold, and then reconstructs the screened IMF components, that is, extracts the weak signal of the seismic data, then performs sample expansion according to the extracted weak signal of the seismic data through a variational autoencoder, then trains a constructed weak signal extraction model of the seismic data based on the extracted weak signal of the seismic data and the expanded samples, and finally extracts the weak signal of the seismic data based on the trained weak signal extraction model of the seismic data. Compared with the existing technology, the present invention can expand the sample data amount of the weak signal of the seismic data, improve the training effect of the weak signal extraction model of the seismic data, that is, improve the accuracy of the weak signal extraction model of the seismic data in extracting the weak signal of the seismic data. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0056] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0057] The present invention provides the following technical solution: a method for extracting weak signals from seismic data based on AI analysis, comprising the following steps:

[0058] S1: Perform empirical mode decomposition on the collected seismic original signal containing noise to obtain several IMF components and a residual component;

[0059] S2: Calculate the permutation entropy of each IMF component;

[0060] S3: Preset the permutation entropy threshold and filter the IMF components based on the permutation entropy threshold;

[0061] S4: reconstruct the filtered IMF components to obtain weak signals of seismic data;

[0062] S5: Based on the weak signal of seismic data, the number of samples of weak signal of seismic data is expanded through variational autoencoder;

[0063] S6: constructing a seismic data weak signal extraction model, and training the seismic data weak signal extraction model based on the expanded seismic data weak signal until convergence;

[0064] S7: Extract weak signals from seismic data based on the converged seismic data weak signal extraction model.

[0065] Empirical mode decomposition is an adaptive signal decomposition method suitable for processing non-stationary and nonlinear signals. Its core idea is to decompose complex signals into multiple intrinsic mode function components with different time scale characteristics, namely IMF components. Each IMF component must meet two conditions: first, the number of extreme points and the number of zero crossings must be equal or differ by at most one over the entire data length; second, at any time, the mean of the upper envelope formed by the local maximum points and the lower envelope formed by the local minimum points is zero;

[0066] In seismic data, empirical mode decomposition can adaptively decompose seismic signals into IMF components of different frequencies. Each IMF component represents the characteristics of the seismic signal at different time scales, that is, decomposing complex seismic signals into multiple relatively simple components for subsequent analysis.

[0067] Specifically, the steps of step S1 include:

[0068] Determine all local maximum and minimum points of the original seismic signal x(t) containing noise;

[0069] The upper envelope e is obtained by cubic spline interpolation. max (t) and the lower envelope e min (t);

[0070] Calculate the mean m1(t) of the upper and lower envelopes:

[0071]

[0072] Where: e max (t) represents the upper envelope; e min (t) represents the lower envelope;

[0073] Subtract the mean from the original signal to get h1(t):

[0074] h1(t)=x(t)-m1(t)

[0075] Where: x(t) represents the original seismic signal containing noise; m1(t) represents the mean of the upper and lower envelopes;

[0076] Determine whether h1(t) meets the two conditions of the IMF component. If so, h1(t) is the first IMF component c1(t). Otherwise, h1(t) is used as the new original signal.

[0077] Repeat the above process and get h after k screenings. k (t), so that it satisfies the IMF component condition, that is, c1(t)=h k (t);

[0078] Subtract the first IMF component c1(t) from the original signal x(t) to obtain the residual signal r(t);

[0079] Take r(t) as the new original signal and repeat the above decomposition process to obtain multiple IMF components in sequence until the remaining signal r n (t) becomes a monotonic function or constant and cannot be decomposed into IMF components;

[0080] That is, decompose the original signal x(t) into a series of IMF components c i (t), i=1,2,…,n, and a residual component r n (t).

[0081] Permutation entropy is a method used to measure the complexity of time series. Its core idea is to measure the irregularity of the signal by analyzing the order of data points in the time series. By calculating the permutation entropy, the possibility of each IMF component containing a weak signal can be determined.

[0082] Generally speaking, the IMF component containing weak signals has a relatively high permutation entropy value due to the interference of weak signals, and the signal appears more complex and irregular.

[0083] Specifically, the specific steps of step S2 include:

[0084] Each IMF component c i (t) Reconstruct the phase space according to the embedding dimension m and time delay τ to obtain N-(m-1)τ m-dimensional vectors X i =(c i (t),c i+τ (t),…,c i+(m-1)τ (t)),i=1,2,…,N-(m-1)τ;

[0085] For each m-dimensional vector X i The elements in are sorted from small to large, and the sorted index series j1, j2, ..., j is obtained. m ;

[0086] Get n different permutations and combinations based on the sorted index sequence;

[0087] Count the probability p of each permutation combination appearing in all reconstruction directions n ,n=1,2,…,n;

[0088] Calculate the permutation entropy H according to the Shannon entropy formula p :

[0089]

[0090] Where: n represents the nth permutation and combination; p n Represents the probability of each permutation combination appearing in all reconstruction directions;

[0091] Permutation entropy H p The larger the value, the higher the complexity of the signal and the stronger the randomness. On the contrary, the permutation entropy value H p The smaller it is, the more regular the signal is and the stronger the certainty is.

[0092] Set a suitable permutation entropy threshold H p,i , filter out the permutation entropy value H p Greater than the permutation entropy threshold H p,i The IMF components that are screened out are considered to contain weak signals with a high probability;

[0093] Permutation entropy threshold H p,i The choice of can be adjusted according to the characteristics of actual seismic data and experience to ensure that components containing weak signals can be accurately screened out.

[0094] Specifically, the specific steps of step S3 include:

[0095] Preset permutation entropy threshold H p,i ;

[0096] Based on the permutation entropy threshold H p,i , for the IMF component c i (t) Screening is performed.

[0097] For the selected IMF component c i (t) is reconstructed to achieve the purpose of extracting weak signals from the original seismic signal x(t) containing noise, and the reconstructed signal obtained is the extracted weak signal.

[0098] Specifically, the specific steps of step S4 include:

[0099] For the selected IMF component c i (t) is superimposed to obtain the reconstructed signal x i (t):

[0100] x i (t)=∑ i∈S c i (t)

[0101] Where: c i (t) represents the IMF component; S represents the permutation entropy value H p Greater than the permutation entropy threshold H p,i The indexed set of IMF components.

[0102] The variational autoencoder encodes and decodes the signal, introduces random perturbations in the encoding space, and generates new samples;

[0103] Currently reconstructed signal x i (t), that is, the number of weak signal samples in seismic data is small, which is difficult to meet the sample size required for AI model training. Therefore, it is necessary to expand the sample size through variational autoencoders to train the AI ​​model.

[0104] Specifically, the specific steps of step S5 include:

[0105] Variational Autoencoder Encoder: Reconstructs the input signal x i (t) is mapped to the latent space and outputs the mean vector μ and the logarithmic variance vector logσ 2 ;

[0106] Variational autoencoder decoder: Sample a vector z from the latent space, z = μ + ∈ × σ, ∈ represents the noise vector sampled from the standard normal distribution, and decode it into the reconstructed signal

[0107] The loss function in the above process consists of reconstruction loss and KL divergence loss:

[0108] L=L i +λL j

[0109] Where: L i Represents the reconstruction loss; λ represents the weight coefficient for balancing the two losses, which is usually adjusted through experiments; L j represents KL divergence loss;

[0110] The reconstruction loss is a measure of the reconstruction signal and reconstructed signal x i Differences in (t):

[0111]

[0112] Where: Represents the reconstructed signal x i (t) takes the expected value; x i (t) represents the reconstructed signal; represents the reconstruction signal;

[0113] The KL divergence loss constrains the distribution of the latent space to be close to the standard normal distribution, ensuring the continuity and sampleability of the latent space:

[0114]

[0115] Where: μ j represents the mean vector; represents the log-variance vector.

[0116] With its powerful nonlinear mapping capabilities and adaptive learning mechanism, the AI ​​model can deeply explore weak effective signals hidden in complex background noise.

[0117] Specifically, the specific steps of step S6 include:

[0118] Construct a weak signal extraction model for seismic data;

[0119] The reconstructed signal x i (t) and the reconstructed signal Combine to form a data set;

[0120] The dataset is divided into a training set and a test set, with a ratio of 7:3;

[0121] The seismic data weak signal extraction model is trained by using the seismic data weak signals in the training set;

[0122] The seismic data weak signal extraction model is verified by the seismic data weak signal in the test set.

[0123] The verified seismic data weak signal extraction model has a high accuracy due to the sufficient training sample size. Therefore, the seismic data weak signal extraction model has a high accuracy in extracting seismic data weak signals.

[0124] Specifically, the specific steps of step S7 include:

[0125] Based on the converged seismic data weak signal extraction model, the weak signal of seismic data is extracted.

[0126] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for extracting weak signals from seismic data based on AI analysis, characterized in that: The following steps are involved: S1: Perform empirical mode decomposition on the collected seismic original signal containing noise to obtain several IMF components and a residual component; S2: Calculate the permutation entropy of each IMF component; S3: Preset the permutation entropy threshold and filter the IMF components based on the permutation entropy threshold; S4: reconstruct the filtered IMF components to obtain weak signals of seismic data; S5: Based on the weak signal of seismic data, the number of samples of weak signal of seismic data is expanded through variational autoencoder; S6: constructing a seismic data weak signal extraction model, and training the seismic data weak signal extraction model based on the expanded seismic data weak signal until convergence; S7: Extract weak signals from seismic data based on the converged seismic data weak signal extraction model.

2. The method for extracting weak signals from seismic data based on AI analysis according to claim 1, characterized in that: The specific steps of step S1 include: Determine all local maximum and minimum points of the original seismic signal x(t) containing noise; The upper envelope e is obtained by cubic spline interpolation. max (t) and the lower envelope e min (t); Calculate the mean m1(t) of the upper and lower envelopes: Where: e max (t) represents the upper envelope; e min (t) represents the lower envelope; Subtract the mean from the original signal to get h1(t): h1(t)=x(t)-m1(t) Where: x(t) represents the original seismic signal containing noise; m1(t) represents the mean of the upper and lower envelopes; Determine whether h1(t) meets the two conditions of the IMF component. If so, h1(t) is the first IMF component c1(t). Otherwise, h1(t) is used as the new original signal. Repeat the above process and get h after k screenings. k (t), so that it satisfies the IMF component condition, that is, c1(t)=h k (t); Subtract the first IMF component c1(t) from the original signal x(t) to obtain the residual signal r(t); Take r(t) as the new original signal and repeat the above decomposition process to obtain multiple IMF components in sequence until the remaining signal r n (t) becomes a monotonic function or constant and cannot be decomposed into IMF components; That is, decompose the original signal x(t) into a series of IMF components c i (t), i=1,2,…,n, and a residual component r n (t).

3. The method for extracting weak signals from seismic data based on AI analysis according to claim 2, characterized in that: The specific steps of step S2 include: Each IMF component c i (t) Reconstruct the phase space according to the embedding dimension m and time delay τ to obtain N-(m-1)τ m-dimensional vectors X i =(c i (t),c i+τ (t),…,c i+(m-1)τ (t)),i=1,2,…,N-(m-1)τ; For each m-dimensional vector X i The elements in are sorted from small to large, and the sorted index series j1, j2, ..., j is obtained. m ; Get n different permutations and combinations based on the sorted index sequence; Count the probability p of each permutation combination appearing in all reconstruction directions n ,n=1,2,…,n; Calculate the permutation entropy H according to the Shannon entropy formula p : Where: n represents the nth permutation and combination; p n Represents the probability of each permutation combination appearing in all reconstruction directions.

4. The method for extracting weak signals from seismic data based on AI analysis according to claim 3, characterized in that: The specific steps of step S4 include: For the selected IMF component c i (t) is superimposed to obtain the reconstructed signal x i (t): x i (t)=Σ i∈S c i (t) Where: c i (t) represents the IMF component; S represents the permutation entropy value H p Greater than the permutation entropy threshold H p,i The indexed set of IMF components.

5. The method for extracting weak signals from seismic data based on AI analysis according to claim 4, characterized in that: The specific steps of step S5 include: Variational Autoencoder Encoder: Reconstructs the input signal x i (t) is mapped to the latent space and outputs the mean vector μ and the logarithmic variance vector logσ 2 ; Variational autoencoder decoder: Sample a vector z from the latent space, z = μ + ∈ × σ, ∈ represents the noise vector sampled from the standard normal distribution, and decode it into the reconstructed signal The loss function in the above process consists of reconstruction loss and KL divergence loss: L=L i +λL j Where: L i Represents the reconstruction loss; λ represents the weight coefficient for balancing the two losses, which is usually adjusted through experiments; L j represents KL divergence loss; The reconstruction loss is a measure of the reconstruction signal and reconstructed signal x i Differences in (t): Where: Represents the reconstructed signal x i (t) takes the expected value; x i (t) represents the reconstructed signal; represents the reconstruction signal; The KL divergence loss constrains the distribution of the latent space to be close to the standard normal distribution, ensuring the continuity and sampleability of the latent space: Where: μ j represents the mean vector; represents the log-variance vector.

Images