Method of stripping strong reflection layer based on deep learning

Abstract

Disclosed herein is a method of stripping a strong reflection layer based on deep learning. The method establishes a direct mapping relationship between a strong reflection signal and seismic data of a target work area through a nonlinear mapping function of the deep neural network, and strips a strong reflection layer after the strong layer is accurately predicted. A mapping relationship between the seismic data containing the strong reflection layer and an event of the strong reflection layer is directedly found through training parameters. In addition, this method does not require an empirical parameter adjustment, and only needs to prepare a training sample that meets the actual conditions of the target work area according to the described rules.

Claims

1. A method of stripping a strong reflection layer based on deep learning, comprising: (1) making a training sample for seismic wavelet prediction; (2) training a mapping relationship between a seismic wavelet and a seismic record based on an alternate iterative deep neural network; (3) predicting a seismic wavelet w based on the mapping relationship between the seismic wavelet and the seismic record trained in step (2); (4) selecting seismic data below or above a strong reflection interface in a target work area, wherein the seismic data does not contain the strong reflection interface; and normalizing the seismic data trace by trace as a training sample of a weak reflection record S.sub.n for stripping the strong reflection layer; (5) counting a range K1-K2 of a ratio of a peak amplitude of the strong reflection layer in the target work area to a peak amplitude of a reflected wave except the strong reflection layer; (6) based on the seismic wavelet predicted in step 3, building a strong refection record S.sub.s only containing the strong reflection layer using a convolution algorithm; wherein a range of a reflection coefficient used in the convolution algorithm is set to the range K1-K2 obtained in step (5); and a position of a reflection interface changes randomly; (7) randomly choosing the strong reflection record S.sub.s and the weak reflection record S.sub.n followed by adding to obtain a seismic record S.sub.r containing strong reflection and weak reflection simultaneously; (8) weighting the S.sub.r, S.sub.s, and S.sub.n with the strong reflection layer as a center using Gaussian window G.sub.w followed by normalization to obtain GS.sub.r, GS.sub.s and GS.sub.n; (9) training a mapping relationship between the GS.sub.r and the GS.sub.s based on U-Net; (10) selecting seismic record S.sub.o containing the strong reflection layer in the target work area; recording a maximum value M.sub.s of each trace S.sub.oi; and taking a position t.sub.p of the maximum value M.sub.s of each trace as a position of the strong reflection layer; (11) weighting the seismic record S.sub.o with the position t.sub.p as a center using the Gaussian window G.sub.w to obtain a weighted seismic record; and normalizing seismic data of each trace by dividing by the maximum value M.sub.s to obtain seismic data of each trace GS.sub.oi; (12) predicting a strong reflection layer GD.sub.si of the seismic data of each trace GS.sub.oi based on the mapping relationship obtained in step (9); (13) subjecting the GD.sub.si to amplitude recovery according to a formula shown as follows: $\begin{array}{cc}{D}_{\mathrm{si}}\left(t\right)=\{\begin{array}{cc}0& G_w\left(t\right)\le 0.1\\ {\mathrm{GD}}_{\mathrm{si}}\left(t\right)*M_{\mathrm{si}}/G_w\left(t\right)& G_w\left(t\right)\ge 0.1\end{array}& \left(1\right)\end{array}$ wherein a subscript i is a trace number; and D.sub.si is a strong reflection layer after the amplitude recovery; (14) obtaining seismic data D.sub.ni after the strong reflection layer is stripped using a formula shown as follows:
D.sub.ni(t)=S.sub.oi(t)−D.sub.si(t)  (2); and (15) detecting, in the target work area, a weak reflection signal of an oil and gas reservoir with the seismic data D.sub.ni after the strong reflection layer is stripped, and performing reservoir prediction and oil-gas exploration in the target work area based on the detection.

2. The method of claim 1, wherein the Gaussian window G.sub.w used in step (8) is expressed as follows: $\begin{array}{cc}{G}_w\left(t\right)=e^{-{a\left(t-t_0\right)}^2}.& \left(3\right)\end{array}$

3. The method of claim 1, wherein an optimization function used in step (9) is shown as follows: $\begin{array}{cc}\underset{C_{\Theta }^{†}}{\min }{E}_{\mathrm{learn}}=\underset{C_{\Theta }^{†}}{\min }E\left[.Math.{\mathrm{GS}}_{\mathrm{si}}\left(t\right)-C_{\Theta }^{†}G{S}_{\mathrm{ri}}\left(t\right).Math.+.Math.{\mathrm{GS}}_{\mathrm{ri}}\left(t\right)-C_{\Theta }^{†}G{S}_{\mathrm{ri}}\left(t\right)-G{S}_{\mathrm{ni}}\left(t\right).Math.\right];& \left(4\right)\end{array}$ wherein a subscript i is a trace number; GS.sub.si is a seismic record only containing strong reflection in the training sample; GS.sub.ri is a seismic record containing the strong reflection; C.sub.Θ.sup.† is a mapping function determined by a parameter set Θ, and represents a mapping relationship between the seismic record containing the strong reflection and the seismic record only containing the strong reflection; and GS.sub.ni is a seismic record without the strong reflection.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) FIG. 1 is a flow chart of a method of stripping a strong reflection layer based on deep learning in accordance with an embodiment of the present disclosure;

(2) FIG. 2 schematically depicts a deep network used in training of a mapping relationship between a seismic wavelet and a seismic record in accordance with an embodiment of the present disclosure;

(3) FIG. 3 schematically depicts a U-Net deep network used in training of a mapping relationship between a seismic record with the strong reflection layer and a strong reflection record in accordance with an embodiment of the present disclosure;

(4) FIGS. 4A-4C schematically depict different seismic sections; where FIG. 4A: a seismic section with a strong reflection layer in accordance with an embodiment of the present disclosure; FIG. 4B: a seismic section with the strong reflection layer removed in accordance with an embodiment of the present disclosure; FIG. 4C: a seismic section only containing the strong reflection layer in accordance with an embodiment of the present disclosure; and

(5) FIGS. 5A-5D show horizontal slices of a three-dimensional seismic data; where FIG. 5A: a horizontal slice of a single-frequency (5 Hz) three-dimensional seismic data containing a strong reflection layer in accordance with an embodiment of the present disclosure; FIG. 5B: a horizontal slice of a single-frequency (80 Hz) three-dimensional seismic data containing a strong reflection layer in accordance with an embodiment of the present disclosure; FIG. 5C: a horizontal slice of the single-frequency (5 Hz) three-dimensional seismic data after the strong reflection layer is removed in accordance with an embodiment of the present disclosure; and FIG. 5D: a horizontal slice of a single-frequency (80 Hz) three-dimensional seismic data after the strong reflection layer is removed in accordance with an embodiment of the present disclosure.

DETAILED DESCRIPTION OF EMBODIMENTS

(6) The technical solutions of this disclosure will be described below clearly and completely with reference to the embodiments and the accompanying drawings. Obviously, presented herein is a part of embodiments of this disclosure, and other embodiments made by those skilled in the art based on the content disclosed herein without sparing any creative effort should fall within the scope of the present disclosure defined by the appended claims.

(7) It should be noted that the terms “first” and “second” used herein are used to distinguish similar objects, and not intended to describe a specific sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances so that the embodiments described herein can be implemented in a sequence other than those illustrated or described herein. In addition, the terms “including” and “comprising” and any variations thereof should be understood as a non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units are not limited to those clearly listed. Those steps or units may include other steps or units that are not clearly listed or are inherent to the process, method, product or device.

(8) The present disclosure will be further described with reference to the accompany drawings.

(9) 1. Stripping of a Strong Reflection Layer Based on Deep Learning

(10) A main process of the method provided herein is shown in FIG. 1, which is specifically described as follows.

(11) (1) Preparation of a Training Sample for Seismic Wavelet Prediction

(12) (a) Seismic Wavelet

(13) Ricker wavelet whose phase and dominant frequency change randomly. It is assumed that the dominant frequency of seismic data in a target work area is f.sub.0, a change interval of the dominant frequency of the Ricker wavelet is (1−0.3)*f.sub.0−(1+0.3)*f.sub.0. A phase change interval is −90°-90°.

(14) (b) Reflection Coefficient

(15) A velocity model and well-logging data acquired from the target survey are used to generate the reflection coefficient similar to a geological structure of the target work area.

(16) (2) Training of a Mapping Relationship Between the Seismic Wavelet and a Seismic Record Based on an Alternate Iterative Deep Neural Network

(17) The alternate iterative deep neural network is shown in FIG. 2. (a) Training of a mapping relationship between the seismic record and the reflection coefficient, in which an optimization function is shown as follows:

(18) $\begin{array}{cc}\underset{A_{\Theta }^{†}}{\min }{E}_{\mathrm{learn}}^{{\mathrm{initial}}_{-r}}=\underset{A_{\Theta }^{†}}{\min }E\left[{.Math.r\left(t\right)-A_{\Theta }^{†}s\left(t\right).Math.}_2^2+{\lambda }_1{.Math.{\hat{r}}_1\left(t\right).Math.}_1\right].& \left(5\right)\end{array}$ (b) Training of a mapping relationship between a least square solution of the seismic wavelet under a constraint of a least square model and a real seismic wavelet:

(19) $\begin{array}{cc}\underset{B_{\Theta }^{†}}{\min }{E}_{\mathrm{learn}}^w=\underset{B_{\Theta }^{†}}{\min }E\left[{.Math.w\left(t\right)-B_{\Theta }^{†}{Q}^{†}s\left(t\right).Math.}_2^2+{\lambda }_2{.Math.{\hat{w}}_1\left(t\right).Math.}_1\right]& \left(6\right)\end{array}$ in which s(t) is a seismic record in the training sample. A.sub.Θ.sup.† and B.sub.Θ.sup.† are mapping functions determined by the parameter set Θ, and require independent training in different networks. The A.sub.Θ.sup.† and B.sub.Θ.sup.† have no influence on each other. r(t) and w(t) are the reflection coefficients of the training sample and the seismic wavelet, respectively. Q.sup.†s(t) is equivalent to the least square solution of the seismic wavelet with known seismic record and reflection coefficient. {circumflex over (r)}.sub.1(t)=A.sub.Θ.sup.†s(t) is a reflection coefficient estimated by a first iterative network. ŵ.sub.1(t)=B.sub.Θ.sup.†Q.sup.†s(t) is a seismic wavelet estimated by a second iterative network. λ.sub.1 and λ.sub.2 are non-negative regularization parameters.

(20) (3) Preparation of a Training Sample for Prediction of Strong Reflection Layer

(21) (a) Wavelet

(22) The seismic record s(t) is input into the mapping functions A.sub.Θ.sup.† and B.sub.Θ.sup.† to estimate a seismic wavelet w.

(23) (b) Reflection Layer S.sub.si Representing a Strong Reflection Interface

(24) A time range of the strong reflection layer in the target work area is counted. The time range is extended by 4 times a wavelength as a time window range W, in which the strong reflection coefficient r.sub.i appears. A position of the strong reflection coefficient changes randomly within the time window W. Magnitude of the strong reflection coefficient is a range K1-K2 of a peak amplitude of the strong reflection layer in the target work area to a peak amplitude of a reflected wave except the strong reflection layer. The strong reflection wave is obtained according to a convolution equation shown as follows:
S.sub.si(t)=r.sub.i(t)*w(t)  (7)

(25) in which * is convolution. i is a trace number, i∈[1,N].

(26) (c) Record S.sub.ni without the Strong Reflection Layer

(27) Seismic data below or above a strong reflection interface in a target work area that does not contain the strong reflection interface is selected. An intercepted time window range is the W in step (b). The seismic data is normalized trace by trace.

(28) (d) Seismic Record S.sub.ri with a Strong Reflection Layer

(29) The S.sub.ni and S.sub.si are randomly extracted and added to obtain the S.sub.ri.

(30) (e) A Gaussian window with a time-domain window width of three wavelet lengths is used to weight the S.sub.ni, the S.sub.si and the S.sub.ri respectively to obtain GS.sub.ni, GS.sub.si and GS.sub.ri with a corresponding time t.sub.0 of the strong reflection layer as a center.

(31) The Gaussian window is shown as follows:

(32) $\begin{array}{cc}{G}_w\left(t\right)=e^{-{a\left(t-t_0\right)}^2}.& \left(3\right)\end{array}$

(33) (4) Training of a Mapping Relationship Between the GS.sub.ri and the GS.sub.si Based on U-Net

(34) A structure of the net is shown in FIG. 3. An optimization function used herein is shown as follows:

(35) $\begin{array}{cc}\underset{C_{\Theta }^{†}}{\min }{E}_{\mathrm{learn}}=\underset{C_{\Theta }^{†}}{\min }E\left[.Math.{\mathrm{GS}}_{\mathrm{si}}\left(t\right)-C_{\Theta }^{†}G{S}_{\mathrm{ri}}\left(t\right).Math.+.Math.{\mathrm{GS}}_{\mathrm{ri}}\left(t\right)-C_{\Theta }^{†}G{S}_{\mathrm{ri}}\left(t\right)-G{S}_{\mathrm{ni}}\left(t\right).Math.\right];& \left(4\right)\end{array}$

(36) wherein a subscript i is a trace number; GS.sub.si is a seismic record only containing strong reflection in the training sample; GS.sub.ri is a seismic record containing the strong reflection C.sub.Θ.sup.† is a mapping function determined by a parameter set Θ, representing a mapping relationship between the seismic record containing the strong reflection and the seismic record only containing the strong reflection; and GS.sub.ni is a seismic record without the strong reflection.

(37) (5) Data S.sub.o containing the strong reflection layer in the target work area is selected. A maximum value M.sub.si of each trace S.sub.oi is recorded, and a time position t.sub.p of the maximum value M.sub.s of each trace S.sub.oi is taken as the position of the strong reflection layer. Each seismic trace S.sub.oi is normalized by dividing by the maximum value M.sub.s.

(38) (6) A weighting is performed with the position t.sub.p as a center using the same Gaussian window as that in step (3) to obtain a weighted seismic record GS.sub.o.

(39) (7) The strong reflection layer GD.sub.si is predicted by inputting the actual seismic data GS.sub.oi of the target work area trace by trace using the mapping relationship trained in step (4):
GD.sub.si(t)=C.sub.Θ.sup.†GS.sub.oi(t)  (8).

(40) (8) The strong reflection record of each trace is subjected to amplitude recovery according to a formula shown as follows:

(41) $\begin{array}{cc}{D}_{\mathrm{si}}\left(t\right)=\{\begin{array}{cc}0& G_w\left(t\right)\le 0.1\\ {\mathrm{GD}}_{\mathrm{si}}\left(t\right)*M_{\mathrm{si}}/G_w\left(t\right)& G_w\left(t\right)\ge 0.1\end{array};& \left(1\right)\end{array}$

(42) in which a subscript i is a trace number. D.sub.si is the strong reflection layer after the amplitude is restored; and

(43) (14) Seismic data D.sub.ni after the strong reflection layer is stripped is obtained using a formula shown as follows:
D.sub.ni(t)=S.sub.oi(t)−D.sub.si(t)  (2).

(44) 2. Actual Seismic Data

(45) The method provided herein is applied to three-dimensional post-stack migration seismic data of a work area in the Ordos Basin. The data in this work area is affected by a strong reflection layer below the reservoir. FIG. 4A shows a typical two-dimensional cross section of the work area. It can be seen that the weak reflection signal of the strong reflection layer cannot be identified on the two-dimensional cross section. FIG. 4B is a seismic section after the strong reflection layer is stripped. It can be seen that after stripping, the sandstone reflection interface near the strong reflection layer is prominent, which is easy for reservoir prediction. The lateral energy change of the stripped strong shielding layer is consistent with the original strong reflection, and no stripping artifacts are generated (FIG. 4C). The stripping position is limited to a vicinity of the shielding layer, and the position information far away from the shielding layer is not affected.

(46) Single-frequency horizontal slices of 5 Hz and 80 Hz are performed on the cross sections before and after stripping the strong reflection layer, respectively. The single-frequency horizontal slices of 5 Hz and 80 Hz before stripping the strong reflection layer show that the two single-frequency slices have little difference in their structures due to the presence of the strong reflection layer (FIGS. 5A-5B). Whereas the single-frequency horizontal slices of 5 Hz and 80 Hz after stripping the strong reflection layer show that the two single-frequency horizontal slices have apparent difference in their structures. The low-frequency horizontal slice of 5 Hz shows rich low-frequency information of the geological structure, and mainly describes a rough outline of the geological structure (FIG. 5C). The high-frequency horizontal slice of 80 Hz mainly reflects detailed features of the geological structure (FIG. 5D).

(47) The above-mentioned are illustrative of the technical solutions of this disclosure, and not intended to limit the present disclosure. Modifications based on the spirit of this disclosure should fall within the scope of the present disclosure defined by the appended claims.