Machine learning high-resolution seismic data processing method based on reflection structure characteristics

Abstract

The present invention discloses a machine learning high-resolution seismic data reconstruction method based on reflection structural features. By inputting seismic data, sonic logging curves and density logging curves are screened, and the logging data in the depth domain is converted into data in the time domain. First, acoustic impedance is calculated, and then the reflection coefficient is calculated. Next, a neural network training set is established, and using the original seismic data, the structural operator is calculated. Finally, the objective function of the machine learning high-resolution data processing system with reflection data structural features is established and solved, completing the machine learning high-resolution seismic data processing based on reflection structural features. The method of the present invention integrates the spatial features of seismic data into the machine learning high-resolution reconstruction process, enabling the prediction results to contain the lateral spatial information of the data, ultimately improving the stability and accuracy of the results. At the same time, the recovery effect of the thin layer structure of seismic data is enhanced, and the spatial distribution characteristics and reservoir characteristics of thin layer sand bodies can be depicted more precisely.

Description

Technical Field

[0001] The present invention belongs to the technical field of seismic data processing in oil and gas geophysical exploration, and particularly relates to a machine learning high-resolution seismic data processing method based on reflection structure characteristics. Background Art

[0002] The seismic exploration method is the main means for oil and gas exploration and development. This method is based on the artificial excitation of seismic waves propagating in the underground medium, and various processes are performed on the received reflected wave signals to explore the underground structure. With the growth of the demand for oil and natural gas resources, the objects of seismic exploration have become increasingly complex, developing from simple surface conditions and underground structures to complex ones, and from structural exploration to lithologic exploration. Therefore, improving the resolution of seismic data so that seismic data can distinguish thin layers and depict formation boundaries has become an urgent need in current seismic data processing.

[0003] In recent years, machine learning has achieved good applications in improving the resolution of seismic exploration. This method uses neural networks to extract low-frequency and high-frequency features from well logging data to broaden the spectral bandwidth of seismic data and improve the resolution of seismic exploration. Compared with existing methods, machine learning technology extracts high-frequency features from the training set and establishes complex non-linear relationships, and can better distinguish thin layers. However, because the well logging curves used to establish labels are one-dimensional sequences, most of the proposed machine learning methods are based on one-dimensional neural networks. One-dimensional networks are sensitive to noise in seismic data. When improving the resolution of multi-dimensional seismic data channel by channel, they often cannot maintain good lateral continuity. Especially when the well logging data is limited and the lithology changes rapidly, such neural networks cannot produce stable and accurate high-resolution results. Most of the conventional machine learning techniques for improving the resolution of seismic exploration are based on one-dimensional neural networks. One-dimensional networks are sensitive to noise in seismic data. When improving the resolution of multi-dimensional seismic data channel by channel, they often cannot maintain good lateral continuity and ignore the spatial features between seismic channels. Summary of the Invention

[0004] To solve the above technical problems, the present invention proposes a machine learning high-resolution seismic data processing method based on reflection structure characteristics, which integrates the spatial features of seismic data into the machine learning high-resolution reconstruction process, so that the prediction results contain the lateral spatial information of the data, and finally improves the stability and accuracy of the results.

[0005] The technical solution adopted by the present invention is as follows: A machine learning high-resolution seismic data processing method based on reflection structure characteristics, and the specific steps are as follows:

[0006] Step 1: Input seismic data, screen acoustic well logging curves and density well logging curves, convert the well logging data in the depth domain into time domain data, first calculate acoustic impedance, and then calculate reflection coefficients;

[0007] Step 2, establish a neural network training set;

[0008] Step 3, use the original seismic data to calculate the structural operator;

[0009] Step 4, establish and solve the objective function of the machine learning high-resolution data processing system with the characteristics of the reflection data structure, and complete the machine learning high-resolution seismic data processing based on the reflection structure characteristics.

[0010] Further, the specific steps of Step 1 are as follows:

[0011] Let the seismic data be represented by x(t), where t = 1, 2, … n, t represents the sampling time, n represents the number of sampling points, v(t) represents the acoustic logging curve, and ρ(t) represents the density logging curve. Then the acoustic impedance z(t) can be calculated as follows:

[0012] z(t) = ρ(t)v(t)

[0013] Then, using the acoustic impedance data, calculate the reflection coefficient r(t), where t = 1, 2, … n.

[0014] The expression of the reflection coefficient is:

[0015]

[0016] where r i-1 represents the reflection coefficient of the i - 1 layer, and z i-1 represents the acoustic impedance of the i - 1 layer.

[0017] Further, the specific steps of Step 2 are as follows:

[0018] The expression of the neural network training set is as follows:

[0019] x h (t) = r(t) * w h (t)

[0020] where * represents the convolution operation. After giving the high-frequency wavelet w h (t), after obtaining the high-frequency seismic data x h (t), extract the corresponding seismic trace x l (t) beside the well, and establish the training set X h and X l for neural network training, and generate a binary occlusion matrix M. This process can be described as:

[0021]

[0022] Among them, L represents the length of the seismic trace, and q represents the serial number of the seismic trace beside the well. Furthermore, a high-resolution processing system of machine learning is established L 1 :

[0023] L 1 = |X h - f θ (X l )|

[0024] Among them, f θ represents the machine learning neural network, and θ represents the neural network parameters.

[0025] On the basis of the existing first norm, the structural similarity (SSIM) is added to capture the complex structural features between data L SSIM :

[0026] L SSIM = [l(X h , X l )] α ·[c(X h , X l )] β ·[s(X h , X l )] γ

[0027]

[0028]

[0029]

[0030] Among them, μ and σ respectively represent the mean and standard deviation of the data, c 1 , c 2 and c 3 represent three preset constants to avoid the situation of instability due to too small denominator. α, β and γ respectively represent the weights corresponding to the l, c and s measurement values.

[0031] Furthermore, the specific steps of step 3 are as follows:

[0032] Calculate the expression of the construction operator h as follows:

[0033]

[0034] Among them, d(t, x) represents the original seismic data, nt represents the number of sampling points of the seismic data in the time direction, nx represents the number of seismic traces, and a and b respectively represent the lengths of the reflection structure operator in the time and space directions.

[0035] After obtaining h, convert it into the structural operator matrix H, that is:

[0036]

[0037] Among them, subscripts -1, -2, etc. represent the relative lengths of the reflection structure operator in the time and space directions.

[0038] Furthermore, the specific steps of step 4 are as follows:

[0039] Establish the objective function θ of the machine learning high-resolution data processing system with reflection data structure characteristics * :

[0040]

[0041] Among them, λ 1 represents the parameter controlling the strength of the L 1 constraint, λ 2 represents the parameter controlling the strength of the L SSIM constraint, λ 3 represents the adjustment factor for constructing the constraint, represents the two-dimensional convolution operator, X * represents the original seismic data matrix, · is the matrix element corresponding multiplication operator, M represents the binary occlusion matrix, which is 1 at the well logging positions and 0 at the other positions.

[0042] Use the ADMM algorithm to solve the objective function. After the network training is completed, save the weights, biases, and other training parameters of each neuron node in the network structure, then input the original seismic data into this network structure, and finally complete the machine learning high-resolution seismic data processing based on the reflection structure characteristics.

[0043] Advantages of the present invention: The present invention discloses a machine learning high-resolution seismic data reconstruction method based on reflection structure characteristics. By inputting seismic data, screening acoustic logging curves and density logging curves, converting the depth-domain logging data into time-domain data, first calculating acoustic impedance, then calculating reflection coefficients, establishing a neural network training set, and using the original seismic data to calculate the structure operator, finally establishing and solving the objective function of the machine learning high-resolution data processing system with reflection data structure characteristics to complete the machine learning high-resolution seismic data processing based on reflection structure characteristics. The method of the present invention integrates the spatial characteristics of seismic data into the machine learning high-resolution reconstruction process, making the prediction results contain the lateral spatial information of the data, ultimately improving the stability and accuracy of the results, while enhancing the restoration effect of the thin layer structure of seismic data and being able to more precisely depict the spatial distribution characteristics and reservoir characteristics of thin layer sand bodies. Brief Description of the Drawings

[0044] Figure 1Flow chart of a machine learning high-resolution seismic data processing method based on reflection structure features of the present invention.

[0045] Figure 2 Noisy synthetic seismic record diagram for experimenting with the method of the present invention in the embodiments of the present invention.

[0046] Figure 3 New seismic record diagram obtained by high-resolution processing using the method of the present invention in the embodiments of the present invention.

[0047] Figure 4 Post-stack seismic record diagram of Block A in a certain oilfield in the embodiments of the present invention.

[0048] Figure 5 Post-stack seismic record diagram of Block A in a certain oilfield after processing using the method of the present invention in the embodiments of the present invention. Detailed implementation manners

[0049] The method of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0050] As Figure 1 shown, in Embodiment 1, a flow chart of a machine learning high-resolution seismic data processing method based on reflection structure features of the present invention is as follows:

[0051] Step 1: Input seismic data, screen acoustic logging curves and density logging curves, convert logging data in the depth domain into data in the time domain, first calculate acoustic impedance, and then calculate the reflection coefficient;

[0052] Step 2: Establish a neural network training set;

[0053] Step 3: Use the original seismic data to calculate the structural operator;

[0054] Step 4: Establish and solve the objective function of a machine learning high-resolution data processing system with reflection data structure features to complete machine learning high-resolution seismic data processing based on reflection structure features.

[0055] In this embodiment, the specific steps of Step 1 are as follows:

[0056] Let x(t), t = 1, 2,... n represent seismic data, where t represents the sampling time, with the unit of millisecond, and n represents the number of sampling points. In the noisy seismic data generated in this embodiment, n = 350, t = 700 ms, the sampling interval is 2 ms, and there are 170 channels in total. Let v(t) represent the acoustic logging curve and ρ(t) represent the density logging curve, then the acoustic impedance z(t) can be calculated as:

[0057] z(t) = ρ(t) · v(t)

[0058] Using the wave impedance data again, the reflection coefficient r(t) is calculated, where t = 1, 2, … n.

[0059] The expression for the reflection coefficient is:

[0060]

[0061] where r i-1 represents the reflection coefficient of the (i - 1)-th layer, and z i-1 represents the acoustic impedance of the (i - 1)-th layer.

[0062] In this embodiment, step 2 is specifically as follows:

[0063] The expression for the neural network training set is as follows:

[0064] x h (t) = r(t) * w h (t)

[0065] where * represents the convolution operation. After the high-frequency wavelet w h (t) is given, the wavelet extracted in this embodiment is sampled at 81 points with a sampling interval of 2 ms to obtain the high-frequency seismic data x h (t). After that, the corresponding seismic trace x l (t) beside the well is extracted to establish the training set X h and X l for neural network training, and a binary occlusion matrix M is generated. This process can be described as:

[0066]

[0067] where L represents the length of the seismic trace, and q represents the serial number of the seismic trace beside the well. In this embodiment, L is 64 and q is 120. Furthermore, a high-resolution processing system L for machine learning is established: 1 :

[0068] L 1 = |X h - f θ (X l )|

[0069] where f θ represents the machine learning neural network, and θ represents the neural network parameters.

[0070] Based on the existing first norm, the structural similarity (SSIM) is added to capture the complex structural features between data L SSIM :

[0071] L SSIM = [l(X h , X l )] α ·[c(Xh , X l )] β · [s(X h , X l )] γ

[0072]

[0073]

[0074]

[0075] Among them, μ and σ respectively represent the mean and standard deviation of the data, and c 1 , c 2 and c 3 represent three preset constants to avoid the unstable situation of too small denominator, and α, β, and γ respectively represent the weights corresponding to the l, c, and s measurement values.

[0076] In this embodiment, the specific steps of step 3 are as follows:

[0077] Calculate the expression of the construction operator h as follows:

[0078]

[0079] Among them, d(t, x) represents the original seismic data, nt represents the number of sampling points of the seismic data in the time direction, nx represents the number of seismic channels, and a and b respectively represent the lengths of the reflection structure operator in the time and space directions. In this embodiment, a is 2 and b is 1.

[0080] After obtaining h, convert it into the structure operator matrix H, that is:

[0081]

[0082] Among them, subscripts -1, -2, etc. represent the relative lengths of the reflection structure operator in the time and space directions.

[0083] In this embodiment, the specific steps of step 4 are as follows:

[0084] Establish the objective function θ of the machine learning high-resolution data processing system with the characteristics of the reflection data structure * :

[0085]

[0086] Among them, λ 1 represents the parameter that controls the strength of the L 1 constraint, and λ 2 represents the parameter that controls the strength of the L SSIM constraint, and λ3 Denotes the adjustment factor for structural constraints Denotes a two-dimensional convolution operator, X * Denotes the original seismic data matrix, · is the operator for element-wise multiplication of matrices, M denotes the binary occlusion matrix, which is 1 at the well logging positions and 0 at the remaining positions.

[0087] In this embodiment, λ 1 = 0.6, λ 2 = λ 3 = 0.2, c 1 , c 2 and c 3 are 0.004, 0.036, 0.036 respectively, and α, β and γ are 0.0448, 0.2855, 0.3001 respectively.

[0088] The ADMM algorithm is used to solve the objective function. After the network training is completed, the weights, biases and other training parameters of each neuron node in the network structure are saved. Then the original seismic data is input into this network structure, and finally the high-resolution seismic data processing based on the reflection structure features is completed.

[0089] As Figure 2 , Figure 3 shown, Figure 2 is the synthetic seismic record with noise for testing the method of the present invention in the embodiment of the present invention, Figure 3 is the seismic record after high-resolution processing using the method of the present invention. Compared with Figure 2 , due to the addition of the reflection structure constraint, the spatial features of the machine learning results are protected and the accuracy of the processing results is improved.

[0090] As Figure 4 , Figure 5 shown, the present invention also provides Embodiment 2.

[0091] This embodiment is an application example of Block A of an oilfield. This exploration block is located in the east. In this area, the attenuation of seismic wave energy by the underground medium is very strong, resulting in very low resolution of the received seismic data. Moreover, as the depth increases, the resolution of the effective reflections in the deeper layers is even lower. Figure 4 is the stacked seismic record obtained from the field acquisition data of this block. It can be seen from this seismic record that there is a complex underground structure in the deeper layer of this block, and the resolution of the effective reflections is very low. Improving the resolution of the effective reflections under the background of such a complex structure to identify the target layer is the key and difficulty of the deconvolution processing. Figure 5This is the stacked seismic record of this block after being processed by the method of the present invention. Compared with before processing, it can be seen that, on the one hand, the machine learning high-resolution processing based on the reflection structure construction constraint effectively improves the vertical resolution of the seismic data, and the event axis becomes thinner; on the other hand, the method of the present invention also effectively protects the spatial continuity of the result, improves the seismic data processing accuracy, and provides high-quality data for subsequent seismic interpretation and seismic inversion.

[0092] In summary, different from directly using a one-dimensional neural network to reconstruct high-resolution seismic data, the method of the present invention not only integrates the spatial characteristics of seismic data into the machine learning high-resolution reconstruction process, but also adds a structural similarity (SSIM) loss function on the basis of the existing first norm as the loss function, and finally greatly enhances the stability and accuracy of the processing result.

[0093] Those of ordinary skill in the art will realize that the embodiments described herein are for helping readers understand the principles of the present invention, and it should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations that do not depart from the essence of the present invention according to these technical revelations disclosed by the present invention, and these deformations and combinations are still within the protection scope of the present invention.

Claims

1. A machine learning high-resolution seismic data processing method based on reflection structure features, the specific steps are as follows: Step 1: Input seismic data, screen acoustic logging curves and density logging curves, convert the logging data in the depth domain into data in the time domain, first calculate the acoustic impedance, and then calculate the reflection coefficient; Step 2: Establish a neural network training set; Step 3: Use the original seismic data to calculate the structural operator; Step 4: Establish and solve the objective function of a machine learning high-resolution data processing system with reflection data structure features to complete the machine learning high-resolution seismic data processing based on reflection structure features; The specific content of Step 4 is as follows: Establish the objective function θ of the machine learning high-resolution data processing system with the characteristics of the reflection data structure * : Among them, λ 1 represents the parameter λ for controlling the strength of constraint L 1 represents the parameter λ for controlling the strength of constraint L 2 represents the parameter λ for controlling the strength of constraint L SSIM represents the parameter λ for controlling the strength of constraint L 3 represents the adjustment factor for constructing the constraint represents the two-dimensional convolution operator, X * represents the original seismic data matrix, · is the operator for element-wise multiplication of matrices, M represents the binary occlusion matrix, which is 1 at the well logging positions and 0 at the other positions; The ADMM algorithm is used to solve the objective function. After the network training is completed, the weights, biases and other training parameters of each neuron node in the network structure are saved, and then the original seismic data is input into this network structure, and finally the machine learning high-resolution seismic data processing based on reflection structure features is completed.

2. A machine learning high-resolution seismic data processing method based on reflection structure features according to claim 1, characterized in that the specific content of Step 1 is as follows: Let x(t), t = 1, 2, … n represent the seismic data, t represents the sampling time, n represents the number of sampling points, v(t) represents the acoustic logging curve, and ρ(t) represents the density logging curve, then the acoustic impedance z(t) can be calculated as: z(t) = ρ(t) · v(t) Then, using the acoustic impedance data, the reflection coefficient r(t), t = 1, 2, … n is calculated; The expression of the reflection coefficient is: where r i-1 represents the reflection coefficient of the (i - 1)-th layer, and z i-1 represents the acoustic impedance of the (i - 1)-th layer.

3. A machine learning high-resolution seismic data processing method based on reflection structure features according to claim 1, characterized in that the specific content of Step 2 is as follows: The expression of the neural network training set is as follows: x h (t) = r(t) * w h (t) where, * represents the convolution operation. Given the high-frequency wavelet w h (t), the high-frequency seismic data x h (t) is obtained. After extracting the corresponding seismic trace x l (t) beside the well, the training set X h and X l are used for neural network training, and a binary occlusion matrix M is generated. This process is described as: where L represents the length of the seismic trace, and q represents the serial number of the seismic trace beside the well; furthermore, a high-resolution processing system L for machine learning is established 1 : L 1 = |X h - f θ (X l )| where f θ represents a machine learning neural network, and θ represents neural network parameters; Based on the existing first norm, the structural similarity SSIM is added to capture the complex structural features L between data SSIM : L SSIM = [l(X h , X l )] α · [c(X h , X l )] β · [s(X h , X l )] γ Among them, μ and σ respectively represent the mean and standard deviation of the data, and c 1 , c 2 and c 3 represent three preset constants to avoid the unstable situation caused by too small a denominator. α, β, and γ respectively represent the weights corresponding to the l, c, and s measurement values.

4. A machine learning high-resolution seismic data processing method based on reflection structure features according to claim 1, characterized in that the specific content of Step 3 is as follows: The expression for calculating the structural operator h is as follows: Among them, d(t, x) represents the original seismic data, nt represents the number of sampling points of the seismic data in the time direction, nx represents the number of seismic channels, and a and b respectively represent the lengths of the reflection structure operator in the time and space directions; After obtaining h, it is converted into a structural operator matrix H, that is: Among them, the subscripts -1, -2, etc. represent the relative lengths of the reflection structure operator in the time and space directions.

Images