A noise reduction method for adaptive seismic leakage signal compensation

Abstract

The present invention discloses a denoising method for adaptive seismic leakage signal compensation. The present invention performs denoising by compensating for seismic leakage signals based on iterative autoencoder IDAE and nonlinear local signal cross-correlation. Specifically, a nonlinear local correlation layer is introduced to construct IDAE in a compensation iteration form, so that leakage signals can be extracted from removed noise and high-precision noise attenuation can be achieved. The nonlinear local correlation layer can be used to extract leakage signals and suppress noise in the extracted leakage signals. It can not only effectively extract leakage signals, but also avoid the assumption of signals and noise, thereby achieving high-precision extraction of leakage signals; the singleness of random noise suppression by DAE determines that it will not introduce new noise in the case of multiple iterative denoising. The method of the present invention uses IDEA in an iterative form to recover leakage signals from removed noise, thereby achieving high-precision denoising of seismic data.

Description

Technical Field

[0001] The invention belongs to the technical field of seismic data processing, and in particular relates to a noise reduction method for adaptive seismic leakage signal compensation. Background Art

[0002] Exploration seismic signals are velocity signals received by the geophone after artificially excited seismic waves. Due to the complex ground and underground environment and the observation errors of the receiving system, exploration seismic signals inevitably contain random noise. Therefore, random noise removal is a necessary step in the processing of exploration seismic data. When the noise becomes complex with the changes in the actual environment, although most denoising methods can effectively remove random noise, they will also remove the effective signals that are difficult to identify as interference, so that the removed noise contains useful signals, resulting in energy loss in the seismic signal. This phenomenon is called "signal leakage". These effective signals are directly lost during processing, and the geological details they contain also disappear at the same time. In addition, from the perspective of amplitude, the loss of signal energy is directly manifested as the attenuation of signal amplitude, which has a huge impact on oil and gas exploration technologies such as AVO that are closely related to amplitude. Therefore, how to recover the effective seismic signals in the removed noise and compensate them back to the original signal is very important for seismic data processing.

[0003] Seismic denoising methods can be divided into two categories based on physical assumptions and mathematical assumptions. Based on the biological assumptions of human brain visual imaging, deep learning technology (DL) has made great breakthroughs in computer vision and other fields by abstracting image features, especially in image denoising and other fields. It has achieved excellent results and has been widely used. Two-dimensional seismic data can be regarded as grayscale images in computer vision, so DL methods can be used to denoise seismic data. The simplest processing is to directly use the "noisy-clean" data pair as a data set to train the DL network to obtain a denoising network. This training method is called supervised learning; another training method is unsupervised learning, which uses a data-driven approach and does not require clean data as labels. The purpose of direct denoising can be achieved by using the network structure. Most of the networks used in unsupervised learning are in the form of "encoding-decoding", among which denoising autoencoders (DAE) are often used to remove random noise. However, even with the help of massive parameters and big data, when the noise is complex, the "signal leakage" in the DL denoising network still exists. Summary of the invention

[0004] In order to solve the above problems existing in the prior art and achieve high-precision attenuation of random noise, the present invention proposes a noise reduction method for adaptive seismic leakage signal compensation.

[0005] The specific technical solution of the present invention is: a noise reduction method for adaptive seismic leakage signal compensation, which specifically includes the following steps:

[0006] Step S1. Setting the hyperparameters of the denoising autoencoder DAE, and initializing the DAE according to the set hyperparameters;

[0007] Step S2. Use the noisy input S n Unsupervised training of DAE;

[0008] Step S3. Use the trained DAE to remove noise and obtain {S t , N t}, where S t represents the predicted signal, N t represents the noise of the prediction;

[0009] Step S4. Using {S t , N t Compensate the leakage signal in the nonlinear local cross-correlation layer to obtain S′ n :S′ n =FL(S t , S n ), where FL represents the nonlinear local image cross-correlation operator, (S n -S t ) represents the removed noise profile; R′ represents the nonlinear similarity matrix; Represents R′ and (S n -S t )’s Hadamard product; R is the similarity matrix; α and β are shape scaling coefficients, α is used to control the operator range, and β is used to define the domain range;

[0010] Step S5. If mean(R)>ε, then S′ n As the noisy input, repeat Step 3 and 4; otherwise, the final output S is obtained t+1 ; Among them, S t+1 =DAE t+1 (S′ n θ t+1 ), t represents the time, θ t and DAE t represents the network parameters and autoencoder network at time t, mean(·) represents the mean function, and ε is the similarity threshold.

[0011] Beneficial effects of the present invention: The present invention proposes a seismic leakage signal compensation denoising method based on iterative autoencoder IDAE and nonlinear local signal cross-correlation. Specifically, by introducing a nonlinear local correlation layer, IDAE is constructed in a compensated iterative form, so that it can extract the leakage signal from the removed noise and achieve high-precision noise attenuation. The nonlinear local correlation layer is used to extract the leakage signal and suppress the noise in the extracted leakage signal. This not only can effectively extract the leakage signal, but also avoids the assumption of signal and noise, thereby achieving high-precision extraction of the leakage signal on a physical basis; the singleness of DAE's suppression of random noise determines that it will not introduce new noise in the case of multiple iterative denoising. In summary, the IDEA of the present invention can realize the recovery of the leakage signal from the removed noise in an iterative form, thereby achieving high-precision denoising of seismic data. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a schematic diagram of the DAE network structure.

[0013] Figure 2 This is the algorithm flow chart of IDEA.

[0014] Figure 3 Schematic diagram of similarity matrix calculation.

[0015] Figure 4 is the graph of nonlinear correlation operators.

[0016] Figure 5 This is the data flow diagram of IDEA.

[0017] Figure 6 This is the original post-stack section 1.

[0018] Figure 7 (a) is the denoising result of profile 1 in the iteration; Figure 7 (b) Noise profile after profile 1 is removed during iteration

[0019] Figure 8 (a) is the final denoising result of the cross section using IDEA; Figure 8 (b) Profile using IDEA - final noise removed profile.

[0020] Fig. 9 for Figure 7 The results are relative to Figure 6 Schematic diagram of the resulting compensated leakage signal.

[0021] Fig.10 This is the original post-stack section 2.

[0022] Fig.11 (a) is the denoising result of profile 2 in the iteration; Fig.11 (b) Noise profile removed from profile 2 during iteration

[0023] Fig.12 (a) is the final denoising result of the cross section using IDEA; Fig.12 (b) Profile using IDEA - final noise removed profile.

[0024] Fig.13 for Fig.11 The results are relative to Fig.10 Schematic diagram of the resulting compensated leakage signal. DETAILED DESCRIPTION

[0025] The embodiments of the present invention are further described below in conjunction with the accompanying drawings.

[0026] In order to achieve high-precision attenuation of random noise, the present invention adopts the Iterative Denoising Autoencoder (IDAE) technology based on nonlinear local signal cross-correlation technology to compensate for leakage seismic signals. First, DAE is used to denoise the noisy seismic signal, and then the local signal cross-correlation technology is used to calculate the similarity between the estimated signal and the removed noise to obtain a similarity weight matrix. Then, the present invention uses the similarity weight matrix to construct a nonlinear mapping operator to enhance the similarity of the effective signal, suppress the similarity of the noise, and obtain a nonlinear similarity weight matrix. The obtained nonlinear local cross-correlation similarity matrix is ​​used to extract and compensate the leakage signal to obtain the seismic denoising profile after leakage signal compensation. The overall technical system of the present invention enables the autoencoder to run iteratively, so it is called iterative autoencoder (IDAE) based on nonlinear local signal cross-correlation technology. The iteration stops by judging whether the mean of the similarity matrix exceeds the similarity threshold. The present invention tests our leakage signal compensation method on actual data. From the experimental results, it is found that our method can well compensate for the effective signal of high-inclination leakage in the noise removal, and achieve high-precision denoising.

[0027] like Figure 1 As shown in the figure, DAE consists of an input layer, a hidden layer, and an output layer. In the DAE structure, the number of neurons in the hidden layer is less than the number of neurons in the input layer and the output layer, and the number of neurons in the input layer is equal to that in the output layer. This "shrink-amplify" symmetric structure is called the "encoding-decoding" structure. When noisy data is used as both network input and label, this structure enables DAE to learn the basic features of the input to reduce training errors. The DAE network can be formulated as:

[0028] S1=Decoder(Encoder(S n ;θ)) (1)

[0029] Among them, S1 and S ndenote the predicted output and noisy input respectively; Decoder(·) and Encodr(·) denote the decoder and encoder respectively; θ denotes the network parameter set of DAE.

[0030] Using gradient descent to train DAE requires calculating the error, while the unsupervised training method of DAE is to use noisy data S n As both network input and label, the network training process is expressed as:

[0031]

[0032] Wherein, DAE(·)=Decoder(Encoder(·)), L(·) represents the error function, and the mean square error (MSE) is generally used to calculate the distance between the two in the Euclidean space.

[0033] The specific parameter update process is completed by the gradient descent optimizer. In the training epoch i, the gradient calculation expression is as follows:

[0034]

[0035] θ i = =θ i-1 -α·g i , (4)

[0036] in,

[0037] 1) and Represent the i-th noisy data and predicted output respectively;

[0038] 2)θ i and θ i-1 Represent the network parameter sets for training epoch i and training epoch i-1 respectively;

[0039] 3) represents the network prediction output using the network parameter set at epoch i-1 and real noisy data Calculation error;

[0040] 4) represents the derivative of L with respect to θ;

[0041] 5) g i represents the error gradient in training epoch i; α represents the gradient descent step size coefficient.

[0042] The DL network calculates the error on the training data according to the error function (as shown in formula (2)), and then the gradient descent optimizer calculates the influence of the network parameter set on the network output according to the error (as shown in formula (3)), and finally optimizes the network parameter set according to the influence (as shown in formula (4)). Thus, a trained DL network is obtained.

[0043] The main purpose of the present invention is to compensate for the leakage signal in the noise during the denoising process. Based on this, the present invention introduces a nonlinear local cross-correlation layer (FL), combines DAE in the form of compensation iteration, and uses the nonlinear local signal cross-correlation technology as a leakage signal compensation operator, so that the entire system can compensate for the leakage signal of the removed noise. In IDAE, DAE is used for random noise attenuation, and nonlinear local correlation is used in the recursive layer between two iterations to generate a nonlinear similarity weight matrix to retrieve the signal leakage in the noise removal, and compensate it to the first DAE output as the input of the next DAE. The iteration is carried out until the stop judgment condition is reached. The present invention tests the compensation of leakage signals by IDEA in the denoising process on actual data. From the experimental results, it is found that the method of the present invention can well compensate for the leakage signal lost in denoising and improve the denoising accuracy.

[0044] like Figure 2 As shown in Figure 1, the DAE network is embedded into the compensation iterative structure as a denoising operator to obtain a denoising structure that can compensate for seismic leakage signals. The signal compensation process is as follows:

[0045] Assume that the real data consists of

[0046] S n =S true +N true (5)

[0047] Among them, S true is the real earthquake signal, N true is the real noise.

[0048] After DAE denoising, we get:

[0049] S n =S1+W1 (6)

[0050] Where S1=DAE(S n ) represents the true signal estimated by DAE, and N1 represents the noise removed by DAE.

[0051] From the above analysis, it can be seen that some leaked real signals still remain in N1. In the embodiment of the present invention, a nonlinear local signal cross-correlation method is used to extract the leakage signal. The local signal cross-correlation technology calculates the similarity between two signals by the coordinate transformation method, and the cross-correlation between the estimated signal S1 and the removed noise N1 is expressed as:

[0052]

[0053] Among them, r is the correlation coefficient, and its value range is

[01] ; and denote the means of S1 and N1, respectively; Nor(·) denotes linear normalization.

[0054] like Figure 3 As shown in Figure 1, by applying a sliding window, a similarity matrix R can be obtained, which describes the similarity between two seismic signals in terms of temporal and spatial morphology. According to the energy and temporal and spatial location information of the leakage signal, the cross-correlation technique can be used to extract the leakage signal from the removed noise.

[0055] According to the phenomenon that noise can also cause small positive similarity, this embodiment suppresses the positive correlation of noise through a nonlinear operator and enhances the positive correlation of leakage signal. The nonlinear similarity matrix R' calculated by the nonlinear operator is expressed as:

[0056]

[0057] Among them, R represents the similarity matrix; α and β are shape scaling coefficients, α is used to control the operator value range, and β is used to define the domain range.

[0058] In this embodiment, α=1, β=5 is used to control the range of the value and the domain to

[01] . Formula (8) can also be expressed as Figure 3 It can be seen that the similarity is less than When suppressed, greater than , where ε is the similarity threshold.

[0059] So an example of compensating the leakage signal by FL can be expressed as

[0060]

[0061] Among them, (S n -S1) represents the removed noise profile; R′ represents the nonlinear similarity matrix; Represents R′ and (S n -S1) Hadamard product; S′ n represents the data of the compensated leakage signal. Formula (9) shows the leakage signal extraction process in FL.

[0062] The nonlinear local image cross-correlation is used as the middle layer (FL) of the IDAE structure, so that IDAE can compensate for the leakage signal. The present invention sets a threshold ε to regard the similarity less than ε as a positive similarity contributed by noise and suppress it. In other words, when most of the similarity is less than the threshold ε, the contribution of the similarity is basically contributed by noise, and there is very little effective leakage signal. Therefore, the present invention sets that when the mean of the similarity matrix R is less than the threshold ε, the iteration stops, so a complete IDAE iteration process can be expressed as:

[0063] S t =DAE t (S n θ t ),

[0064] S′ n =FL(S t , S n )

[0065] ONLY IF mean(R)>ε,DOS t+1 =DAE t+1 (S′ n θ t+1 ), ELSE output = S t (10)

[0066] Among them, t represents the time (number of iterations); θ t and DAE t represents the network parameters and autoencoder network at time t; S t and S′ n represent the DAE output at time t and the data of the compensated leakage signal (i.e., the DAE input at time t+1), respectively; mean(·) represents the mean function; FL represents the nonlinear local image cross-correlation operator, as shown in Figure 4 shown.

[0067] The overall data flow is as follows Figure 5 As shown, the specific steps of this embodiment can be obtained as follows:

[0068] Step S1. Setting the hyperparameters of the denoising autoencoder DAE, and initializing the DAE according to the set hyperparameters;

[0069] Step S2. Use the noisy input S t Unsupervised training of DAE;

[0070] Step S3. Use the trained DAE to remove noise and obtain {S t , N t};

[0071] Step S4. Using {S t, N t Compensate the leakage signal in the nonlinear local cross-correlation layer to obtain S′ n ;

[0072] Step S5. If mean(R)>ε, then S′ n As the noisy input, repeat Step 3 and 4; otherwise, the final output S is obtained t+1 .

[0073] Using the above method Figure 6 The original post-stack section Fig.10 Processing for the original post-stack section 2.

[0074] Figure 7 (a) is the denoising result of profile 1 in the iteration; Figure 7 (b) Noise profile after profile 1 is removed during iteration; Figure 8 (a) is the final denoising result of the cross section using IDEA; Figure 8 (b) Profile using IDEA - final noise profile removed; Fig. 9 for Figure 7 The results are relative to Figure 6 Schematic diagram of the resulting compensated leakage signal.

[0075] Fig.11 (a) is the denoising result of profile 2 in the iteration; Fig.11 (b) Noise profile removed from profile 2 during iteration; Fig.12 (a) is the final denoising result of the cross section using IDEA; Fig.12 (b) Profile using IDEA - final noise profile removed; Fig.13 for Fig.11 The results are relative to Fig.10 Schematic diagram of the resulting compensated leakage signal.

[0076] It can be seen from the results in the figure that the actual denoising results of the method of the present invention are consistent with the theoretical assumptions, the compensation effect on high-angle seismic leakage signals is obvious, and the removed noise profile contains very little signal residue.

Claims

1. A noise reduction method for adaptive seismic leakage signal compensation, comprising the following steps: Step S1. Setting the hyperparameters of the denoising autoencoder DAE, and initializing the DAE according to the set hyperparameters; Step S2. Use the noisy input S n Unsupervised training of DAE; Step S3. Use the trained DAE to remove noise and obtain {S t , N t },in, S t represents the predicted signal, N t represents the noise of the prediction; Step S4. Using {S t , N t Compensate the leakage signal in the nonlinear local cross-correlation layer to obtain S′ n :S′ n =FL(S t , S n ), where FL represents the nonlinear local image cross-correlation operator, (S n -S t ) represents the removed noise profile; R′ represents the nonlinear similarity matrix; Represents R′ and (S n -S t )’s Hadamard product; R is the similarity matrix; α and β are shape scaling coefficients, α is used to control the operator range, and β is used to define the domain range; Step S5. If mean(R)>ε, then S′ n As the noisy input, repeat steps S3 and 4; otherwise, the final output S t+1 ; Among them, S t+1 =DAE t+1 (S′ n θ t+1 ), t represents the time, θ t and DAE t represents the network parameters and autoencoder network at time t, mean(·) represents the mean function, and ε is the similarity threshold.

Images