Pre-stack seismic inversion model design method based on combination of one-dimensional GRU and two-dimensional improved ASPP

Abstract

The invention discloses a pre-stack seismic inversion model design method based on a one-dimensional GRU combined with a two-dimensional improved ASPP, and the method comprises the steps: introducing a cavity space convolution pooling pyramid module into the field of pre-stack inversion, and enabling the improved ASPP to transmit the input into three parts: a convolution pooling layer, a pooling pyramid and a self-adaptive mean pooling layer; the output of the pooling pyramid and the output of the adaptive mean pooling layer are spliced, and the pooling pyramid and the output of the adaptive mean pooling layer are spliced with the convolution pooling layer after passing through a convolution block and a dropout module; a triple attention module is added in an inversion framework, features of different scales of seismic data are extracted, and low-frequency information of the seismic data is extracted by combining a recurrent neural network; in the network training stage, a multi-gather inversion method is adopted, and a network model is trained in combination with one-dimensional inversion and two-dimensional inversion. According to the pre-stack seismic inversion model design method of the invention, the inversion result is not only good in transverse continuity, but also accurate and clear in inversion of details, and especially, the density parameter inversion effect is obviously improved.

Description

Technical Field

[0001] The invention relates to the technical field of seismic inversion, and in particular to a pre-stack seismic inversion model design method based on one-dimensional GRU combined with two-dimensional improved ASPP. Background Art

[0002] Seismic inversion is the process of inferring the structure, morphology and material composition of the earth's interior based on various geophysical observation data, and quantitatively calculating various relevant geophysical parameters. Many problems in oil and gas seismic exploration can ultimately be attributed to seismic inversion problems. Therefore, the study of seismic inversion is of great significance in oil and gas seismic exploration and is an important part of underground energy exploration. Seismic inversion methods mainly include pre-stack inversion and post-stack inversion. Compared with post-stack inversion, pre-stack inversion uses pre-stack seismic data that contains richer and more accurate information. Therefore, pre-stack inversion has been widely used and has achieved many results.

[0003] Traditional seismic inversion methods are model-driven and can roughly invert the structure of the stratigraphic profile. However, due to acquisition limitations, potential measurement errors, noise and other factors, model-driven methods will inevitably produce errors. In order to obtain good inversion results, model-driven methods usually require a good initial model and a more accurate wavelet, but in practical applications, it is usually difficult to obtain a good initial model and an accurate wavelet.

[0004] Deep learning is a subset of machine learning. It is a data-driven method that can learn complex mapping relationships between input and output through a large amount of data. In recent years, deep learning has made breakthrough progress in image classification, object detection, image segmentation, image and video subtitles, speech recognition, machine translation and other fields. The success of deep learning in the fields of computer vision and natural language processing has made data-driven intelligent seismic inversion methods widely concerned by scholars. Deep learning methods do not require an initial model or estimate seismic wavelets. They continuously learn from a large amount of sample data and then fit complex nonlinear functions to achieve underground reservoir feature inversion. They have strong nonlinear mapping and generalization capabilities, so deep learning is very suitable for use in the field of seismic inversion. Applying deep learning to seismic inversion can obtain inversion results with higher accuracy and better effects.

[0005] Existing deep learning methods can already obtain good inversion results, but there are still problems such as insufficient logging data, poor lateral continuity of inversion results, insufficient feature extraction, and low inversion accuracy in details. Summary of the invention

[0006] In view of the above technical problems to be solved, the present invention provides a prestack seismic inversion model design method based on one-dimensional GRU combined with two-dimensional improved ASPP, which has high inversion result accuracy, good lateral continuity and sufficient detail inversion.

[0007] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0008] A prestack seismic inversion model design method based on one-dimensional GRU combined with two-dimensional improved ASPP, introduces the void space convolution pooling pyramid module into the prestack inversion framework, and improves ASPP to send input to three parts: convolution pooling layer, pooling pyramid and adaptive mean pooling layer; firstly, the outputs of the pooling pyramid and the adaptive mean pooling layer are spliced, and then passed through a convolution block and a dropout module, and then spliced ​​with the convolution pooling layer; a triple attention module is added to the inversion framework to extract features of seismic data at different scales, and a recurrent neural network is combined to extract low-frequency information of seismic data; in the network training stage, a multi-channel set inversion method is adopted to combine one-dimensional inversion and two-dimensional inversion to train the network model.

[0009] As a further improvement of the above technical solution:

[0010] Preferably, the design method specifically includes the following contents:

[0011] The overall network framework consists of an inversion network and a forward modeling network. The seismic data is input into the inversion network, and the S-wave velocity, P-wave velocity and density are output. Using a cyclic network structure, the S-wave velocity, P-wave velocity and density obtained by the inversion network are input into the forward modeling network to obtain the seismic data.

[0012] The inversion network includes a feature extraction module and a regression module. The feature extraction module is divided into a local feature extraction module and a sequence modeling module. During network training, the input of the local feature extraction module in the inversion network is two-dimensional seismic data, and the input of the sequence modeling module is one-dimensional seismic data.

[0013] The regression module maps data from the spatiotemporal feature sequence to shear wave velocity, longitudinal wave velocity and density; calculates the attribute loss by comparing the true shear wave velocity, longitudinal wave velocity, density and the inverted shear wave velocity, longitudinal wave velocity and density; and calculates the seismic data loss by comparing the true seismic data and the seismic data output by the forward modeling network.

[0014] Preferably, the convolutional pooling layer consists of a 1×5 convolution, a 3×3 pooling layer and two group normalization layers.

[0015] Preferably, a triple attention module is added before the output of the local feature extraction module, and the triple attention module consists of three branches. The first branch is the channel attention calculation branch, and the input feature is pooled, then 7×7 convolution, and finally the sigmoid activation function is used to generate the spatial attention weight; the second branch is the channel C and space W dimension interaction capture branch, the input dimensional feature is C×H×W, first undergoes a permutation operation to become an H×C×W dimensional feature, then a pooling operation is performed on the H dimension, and finally a permutation operation is performed to become a C×H×W dimensional feature; the third branch is the channel C and space H dimension interaction capture branch, the input feature is first permuted to become a W×H×C dimensional feature, and then pooled on the W dimension; after generating the attention weight, each branch averages and aggregates the outputs of the three branches to obtain the triple attention output.

[0016] Preferably, the regression module is composed of two deconvolution blocks, an LSTM layer and a fully connected layer, the deconvolution block is composed of a deconvolution layer and a group normalization layer, the group normalization groups the output of the convolution layer and normalizes each group using the learned mean and standard deviation; the deconvolution block has learnable kernel parameters for compensating for the resolution mismatch between seismic data and well logging data; the LSTM layer and the fully connected layer send the features extracted by the feature extraction module back to the target domain.

[0017] Preferably, the input of the regression module is the output of the feature extraction module, and the output is the inverted shear wave velocity, longitudinal wave velocity and density;

[0018] Calculate the mean square error between the well logging shear wave velocity, compressional wave velocity, density and the shear wave velocity, compressional wave velocity, density output by the regression module. Calculate the mean square error between the inverted data and the well logging data according to formula (9). Update the learnable parameters in the feature extraction module and the regression module, which are recorded as attribute loss:

[0019]

[0020] Among them, m i,t is the actual logging data, is the inverted data, L1 represents the attribute loss, and MSE is the mean square error function.

[0021] Preferably, the input of the forward modeling module is the output of the inversion network, and the output is seismic data. The forward modeling module is composed of four convolutional layers and two activation functions to complete the reconstruction task;

[0022] The mean square error between the input of the sequence modeling module and the output of the forward modeling module is calculated. The mean square error between the input seismic data and the seismic data obtained by the forward modeling network is calculated according to formula (10), which is recorded as seismic data loss:

[0023]

[0024] Among them, x i,t Well logging seismic data is input to the sequence modeling module. is the seismic data obtained by the forward modeling network, and L2 represents the loss of seismic data.

[0025] Preferably, the loss of the inversion process is composed of attribute loss and seismic data loss, and the total loss is shown in formula (11):

[0026] L=αL1+βL2 (7)

[0027] Among them, α and β are weight coefficients, which are used to control the proportion of attribute loss and seismic data loss in the total loss function. The loss function of the entire inversion network is shown in formula (12):

[0028]

[0029] The prestack seismic inversion model design method based on one-dimensional GRU combined with two-dimensional improved ASPP provided by the present invention has the following advantages compared with the prior art:

[0030] The prestack seismic inversion model design method based on one-dimensional GRU combined with two-dimensional improved ASPP of the present invention has an overall network structure composed of a forward network and an inversion network, wherein the inversion network is composed of a feature extraction module and a regression module, and in the inversion module, a multi-channel set inversion method is adopted to fully utilize the spatial information in the seismic data and improve the lateral continuity of the inversion result. At the same time, the present invention changes the connection order and mode of the three components of the ASPP module, so that it can fully extract the multi-scale features of the seismic data while paying attention to the long-range information in the seismic data, thereby improving the accuracy of the inversion result, and adding a Triplet Attention module to the improved ASPP module, which fully considers the spatial correlation and channel correlation of the seismic data and allows the network to better process the long-range relationship of the data. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the inversion network structure in the present invention.

[0032] FIG. 2( a ) is a structural diagram of the ASPP in the present invention.

[0033] FIG. 2( b ) is a structural diagram of IASPP in the present invention.

[0034] Figure 3 It is a structural diagram of Triplet Attention of the present invention.

[0035] Figure 4It is a structural diagram of the regression module of the present invention.

[0036] Figure 5 It is a structural diagram of the forward modeling module of the present invention.

[0037] Figure 6 (a) is a small-angle stacked seismic profile of the Marmousi2 model in the model data test.

[0038] Figure 6 (b) is the angle stacked seismic profile in the Marmousi2 model during the model data test.

[0039] Figure 6 (c) is the large-angle stacked seismic profile of the Marmousi2 model in the model data test.

[0040] Figure 7 (a) is the true shear wave velocity diagram of the Marmousi2 model in the model data test.

[0041] Figure 7 (b) is the true longitudinal wave velocity diagram of the Marmousi2 model in the model data test.

[0042] Figure 7 (c) is the true density profile of the Marmousi2 model in the model data test.

[0043] Figure 7 (d) is the initial shear wave velocity diagram of the Marmousi2 model in the model data test.

[0044] Figure 7 (e) is the initial longitudinal wave velocity diagram of the Marmousi2 model in the model data test.

[0045] Figure 7 (f) is the initial density profile of the Marmousi2 model in the model data test.

[0046] Figure 8 a(1) is a cross-sectional diagram of the shear wave velocity inversion results of the Marmousi2 model CNN method in the model data test.

[0047] Figure 8 a(2) is a cross-sectional diagram of the longitudinal wave velocity inversion results of the Marmousi2 model CNN method in the model data test.

[0048] Figure 8 a(3) is a profile diagram of the density inversion results of the Marmousi2 model CNN method in the model data test.

[0049] Figure 8b(1) is a cross-sectional diagram of the shear wave velocity inversion results of the Marmousi2 model TCN method in the model data test.

[0050] Figure 8 b(2) is a cross-sectional diagram of the longitudinal wave velocity inversion results of the Marmousi2 model TCN method in the model data test.

[0051] Figure 8 b(3) is the profile of density inversion results of the Marmousi2 model TCN method in the model data test.

[0052] Figure 8 c(1) is a cross-sectional diagram of the shear wave velocity inversion results of the Marmousi2 model CNN-GRU method in the model data test.

[0053] Figure 8 c(2) is a cross-sectional diagram of the longitudinal wave velocity inversion results of the Marmousi2 model CNN-GRU method in the model data test.

[0054] Figure 8 c(3) is a profile diagram of the density inversion results of the Marmousi2 model CNN-GRU method in the model data test.

[0055] Figure 8 d(1) is the profile of the shear wave velocity inversion results of the Marmousi2 model ASPP-GRU method in the model data test.

[0056] Figure 8 d(2) is the profile of the longitudinal wave velocity inversion results of the Marmousi2 model ASPP-GRU method in the model data test.

[0057] Figure 8 d(3) is the profile of the density inversion result of the ASPP-GRU method of the Marmousi2 model in the model data test.

[0058] Figure 8 e(1) is a cross-sectional diagram of the shear wave velocity inversion results of the Marmousi2 model IASPP-GRU method in the model data test.

[0059] Figure 8 e(2) is a cross-sectional diagram of the longitudinal wave velocity inversion results of the Marmousi2 model IASPP-GRU method in the model data test.

[0060] Figure 8 e(3) is a profile diagram of the density inversion results of the Marmousi2 model IASPP-GRU method in the model data test.

[0061] Figure 8f(1) is the profile of the shear wave velocity inversion results of the Marmousi2 model IASPP-GRU-init method in the model data test.

[0062] Figure 8 f(2) is the profile of the longitudinal wave velocity inversion results of the Marmousi2 model IASPP-GRU-init method in the model data test.

[0063] Figure 8 f(3) is the profile of the density inversion results of the Marmousi2 model IASPP-GRU-init method in the model data test.

[0064] Figure 8 g(1) is a cross-sectional diagram of the shear wave velocity inversion results of the Marmousi2 model IGIT method in the model data test.

[0065] Figure 8 g(2) is a cross-sectional diagram of the longitudinal wave velocity inversion results of the Marmousi2 model IGIT method in the model data test.

[0066] Figure 8 g(3) is a profile diagram of the density inversion results of the Marmousi2 model IGIT method in the model data test.

[0067] Fig. 9 a(1) is the absolute error profile of shear wave velocity inversion using the CNN method of the Marmousi2 model in the model data test.

[0068] Fig. 9 a(2) is the absolute error profile of the P-wave velocity inversion using the CNN method of the Marmousi2 model in the model data test.

[0069] Fig. 9 a(3) is the absolute error profile of density inversion using the Marmousi2 model CNN method in the model data test.

[0070] Fig. 9 b(1) is the absolute error profile of shear wave velocity inversion using the TCN method for the Marmousi2 model in the model data test.

[0071] Fig. 9 b(2) is the absolute error profile of the P-wave velocity inversion using the TCN method of the Marmousi2 model in the model data test.

[0072] Fig. 9 b(3) is the absolute error profile of density inversion using the TCN method of the Marmousi2 model in the model data test.

[0073] Fig. 9c(1) is the absolute error profile of shear wave velocity inversion using the CNN-GRU method of the Marmousi2 model in the model data test.

[0074] Fig. 9 c(2) is the absolute error profile of the longitudinal wave velocity inversion using the CNN-GRU method of the Marmousi2 model in the model data test.

[0075] Fig. 9 c(3) is the absolute error profile of density inversion of the Marmousi2 model CNN-GRU method in the model data test.

[0076] Fig. 9 d(1) is the absolute error profile of shear wave velocity inversion using the ASPP-GRU method of the Marmousi2 model in the model data test.

[0077] Fig. 9 d(2) is the absolute error profile of the longitudinal wave velocity inversion using the ASPP-GRU method of the Marmousi2 model in the model data test.

[0078] Fig. 9 d(3) is the absolute error profile of density inversion of the ASPP-GRU method of the Marmousi2 model in the model data test.

[0079] Fig. 9 e(1) is the absolute error profile of shear wave velocity inversion using the IASPP-GRU method of the Marmousi2 model in the model data test.

[0080] Fig. 9 e(2) is the absolute error profile of the P-wave velocity inversion using the IASPP-GRU method of the Marmousi2 model in the model data test.

[0081] Fig. 9 e(3) is the absolute error profile of density inversion of the Marmousi2 model IASPP-GRU method in the model data test.

[0082] Fig. 9 f(1) is the absolute error profile of shear wave velocity inversion using the IASPP-GRU-init method of the Marmousi2 model in the model data test.

[0083] Fig. 9 f(2) is the absolute error profile of the P-wave velocity inversion using the IASPP-GRU-init method of the Marmousi2 model in the model data test.

[0084] Fig. 9f(3) is the absolute error profile of density inversion of the Marmousi2 model IASPP-GRU-init method in the model data test.

[0085] Fig. 9 g(1) is the absolute error profile of shear wave velocity inversion using the IGIT method for the Marmousi2 model in the model data test.

[0086] Fig. 9 g(2) is the absolute error profile of the P-wave velocity inversion using the IGIT method for the Marmousi2 model in the model data test.

[0087] Fig. 9 g(3) is the absolute error profile of density inversion using the IGIT method for the Marmousi2 model in the model data test.

[0088] Fig.10 (a) is a comparison chart of the inversion results of the 200th seismic trace set of shear wave velocity using the Marmousi2 model CNN, TCN, CNN-GRU and the method of the present invention in the model data test.

[0089] Fig.10 (b) is a comparison chart of the inversion results of the 200th seismic trace set of longitudinal wave velocity using the Marmousi2 model CNN, TCN, CNN-GRU and the method of the present invention in the model data test.

[0090] Fig.10 (c) is a comparison chart of the inversion results of the 200th seismic trace set of the Marmousi2 model CNN, TCN, CNN-GRU and the method of the present invention in the model data test.

[0091] Fig.10 (d) is a comparison chart of the inversion results of the 200th seismic trace set of shear wave velocity using the Marmousi2 model ASPP-GRU, IASPP-GRU, IASPP-GRU-init and IGIT methods in the model data test.

[0092] Fig.10 (e) is a comparison chart of the inversion results of the 200th seismic trace set of longitudinal wave velocity using the Marmousi2 model ASPP-GRU, IASPP-GRU, IASPP-GRU-init and IGIT methods in the model data test.

[0093] Fig.10 (f) is a comparison chart of the inversion results of the 200th seismic trace set of the Marmousi2 model ASPP-GRU, IASPP-GRU, IASPP-GRU-init and IGIT method density in the model data test.

[0094] Fig.11 (a) is the small-angle stacked seismic profile of the SEAM model in the model data test.

[0095] Fig.11 (b) is the angle stacked seismic profile in the SEAM model during the model data test.

[0096] Fig.11 (c) is the large-angle stacked seismic profile of the SEAM model in the model data test.

[0097] Fig.12 (a) is the true shear wave velocity diagram of the SEAM model in the model data test.

[0098] Fig.12 (b) is the true longitudinal wave velocity diagram of the SEAM model in the model data test.

[0099] Fig.12 (c) is the true density profile of the SEAM model in the model data test.

[0100] Fig.12 (d) is the initial shear wave velocity diagram of the SEAM model in the model data test.

[0101] Fig.12 (e) is the initial longitudinal wave velocity diagram of the SEAM model in the model data test.

[0102] Fig.12 (f) is the initial density profile of the SEAM model in the model data test.

[0103] Fig.13 a(1) is a cross-sectional diagram of the shear wave velocity inversion results of the SEAM model CNN method in the model data test.

[0104] Fig.13 a(2) is a cross-sectional diagram of the longitudinal wave velocity inversion results of the SEAM model CNN method in the model data test.

[0105] Fig.13 a(3) is a profile diagram of the density inversion results of the SEAM model CNN method in the model data test.

[0106] Fig.13 b(1) is a cross-sectional diagram of the shear wave velocity inversion results of the SEAM model TCN method in the model data test.

[0107] Fig.13 b(2) is a cross-sectional diagram of the longitudinal wave velocity inversion results of the SEAM model TCN method in the model data test.

[0108] Fig.13 b(3) is the profile of the density inversion results of the SEAM model TCN method in the model data test.

[0109] Fig.13 c(1) is a cross-sectional diagram of the shear wave velocity inversion results using the CNN-GRU method of the SEAM model in the model data test.

[0110] Fig.13 c(2) is a cross-sectional diagram of the longitudinal wave velocity inversion results of the SEAM model CNN-GRU method in the model data test.

[0111] Fig.13 c(3) is a cross-sectional diagram of the density inversion results of the SEAM model CNN-GRU method in the model data test.

[0112] Fig.13 d(1) is the profile of the shear wave velocity inversion results of the SEAM model ASPP-GRU method in the model data test.

[0113] Fig.13 d(2) is the profile of the longitudinal wave velocity inversion results of the SEAM model ASPP-GRU method in the model data test.

[0114] Fig.13 d(3) is the profile of the density inversion result of the ASPP-GRU method of the SEAM model in the model data test.

[0115] Fig.13 e(1) is a cross-sectional diagram of the shear wave velocity inversion results of the SEAM model IASPP-GRU method in the model data test.

[0116] Fig.13 e(2) is a cross-sectional diagram of the longitudinal wave velocity inversion results of the SEAM model IASPP-GRU method in the model data test.

[0117] Fig.13 e(3) is a profile diagram of the density inversion results of the SEAM model IASPP-GRU method in the model data test.

[0118] Fig.13 f(1) is the profile of the shear wave velocity inversion results of the SEAM model IASPP-GRU-init method in the model data test.

[0119] Fig.13 f(2) is the profile of the longitudinal wave velocity inversion results of the SEAM model IASPP-GRU-init method in the model data test.

[0120] Fig.13 f(3) is the profile of the density inversion result of the IASPP-GRU-init method of the SEAM model in the model data test.

[0121] Fig.13g(1) is the profile of the shear wave velocity inversion results of the SEAM model IGIT method in the model data test.

[0122] Fig.13 g(2) is a cross-sectional diagram of the longitudinal wave velocity inversion results of the SEAM model IGIT method in the model data test.

[0123] Fig.13 g(3) is the profile of the density inversion results of the SEAM model IGIT method in the model data test.

[0124] Fig.14 a(1) is the absolute error profile of shear wave velocity inversion using the CNN method of the SEAM model in the model data test.

[0125] Fig.14 a(2) is the absolute error profile of the longitudinal wave velocity inversion using the CNN method of the SEAM model in the model data test.

[0126] Fig.14 a(3) is the absolute error profile of density inversion of the SEAM model CNN method in the model data test.

[0127] Fig.14 b(1) is the absolute error profile of shear wave velocity inversion using the TCN method of the SEAM model in the model data test.

[0128] Fig.14 b(2) is the absolute error profile of the longitudinal wave velocity inversion using the TCN method of the SEAM model in the model data test.

[0129] Fig.14 b(3) is the absolute error profile of density inversion using the TCN method of the SEAM model in the model data test.

[0130] Fig.14 c(1) is the absolute error profile of shear wave velocity inversion using the CNN-GRU method of the SEAM model in the model data test.

[0131] Fig.14 c(2) is the absolute error profile of the longitudinal wave velocity inversion using the CNN-GRU method of the SEAM model in the model data test.

[0132] Fig.14 c(3) is the absolute error profile of density inversion of the SEAM model CNN-GRU method in the model data test.

[0133] Fig.14 d(1) is the absolute error profile of shear wave velocity inversion using the ASPP-GRU method of the SEAM model in the model data test.

[0134] Fig.14d(2) is the absolute error profile of the longitudinal wave velocity inversion using the ASPP-GRU method of the SEAM model in the model data test.

[0135] Fig.14 d(3) is the absolute error profile of density inversion of the ASPP-GRU method of the SEAM model in the model data test.

[0136] Fig.14 e(1) is the absolute error profile of shear wave velocity inversion using the IASPP-GRU method of the SEAM model in the model data test.

[0137] Fig.14 e(2) is the absolute error profile of the longitudinal wave velocity inversion using the IASPP-GRU method of the SEAM model in the model data test.

[0138] Fig.14 e(3) is the absolute error profile of density inversion of the SEAM model IASPP-GRU method in the model data test.

[0139] Fig.14 f(1) is the absolute error profile of shear wave velocity inversion using the IASPP-GRU-init method of the SEAM model in the model data test.

[0140] Fig.14 f(2) is the absolute error profile of the longitudinal wave velocity inversion using the IASPP-GRU-init method of the SEAM model in the model data test.

[0141] Fig.14 f(3) is the absolute error profile of density inversion of the IASPP-GRU-init method of the SEAM model in the model data test.

[0142] Fig.14 g(1) is the absolute error profile of shear wave velocity inversion using the IGIT method of the SEAM model in the model data test.

[0143] Fig.14 g(2) is the absolute error profile of the longitudinal wave velocity inversion using the IGIT method of the SEAM model in the model data test.

[0144] Fig.14 g(3) is the absolute error profile of density inversion using the IGIT method of the SEAM model in the model data test.

[0145] Fig.15 (a) is a comparison chart of the inversion results of the 200th seismic trace set of shear wave velocity using the SEAM model CNN, TCN, CNN-GRU and the method of the present invention in the model data test.

[0146] Fig.15(b) is a comparison chart of the inversion results of the 200th seismic trace set of longitudinal wave velocity using the SEAM model CNN, TCN, CNN-GRU and the method of the present invention in the model data test.

[0147] Fig.15 (c) is a comparison chart of the inversion results of the 200th seismic trace set of the SEAM model CNN, TCN, CNN-GRU and the method of the present invention in the model data test.

[0148] Fig.15 (d) is a comparison chart of the inversion results of the 200th seismic trace set of shear wave velocity using the SEAM models ASPP-GRU, IASPP-GRU, IASPP-GRU-init and the IGIT method in the model data test.

[0149] Fig.15 (e) is a comparison chart of the inversion results of the 200th seismic trace set of longitudinal wave velocity using the SEAM models ASPP-GRU, IASPP-GRU, IASPP-GRU-init and the IGIT method in the model data test.

[0150] Fig.15 (f) is a comparison chart of the inversion results of the 200th seismic trace set of the SEAM model ASPP-GRU, IASPP-GRU, IASPP-GRU-init and the IGIT method in the model data test. DETAILED DESCRIPTION

[0151] The specific embodiments of the present invention are described in detail below. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0152] like Figures 1 to 15 As shown, the prestack seismic inversion model design method based on one-dimensional GRU combined with two-dimensional improved ASPP of the present invention introduces the void space convolution pooling pyramid module into the field of prestack inversion, and changes the connection order and mode of its three components, adds a triple attention module to extract the features of seismic data at different scales, and combines the recurrent neural network to extract the low-frequency information of seismic data; in the network training stage, the multi-channel set inversion method is adopted to combine the one-dimensional inversion and the two-dimensional inversion to train the network model.

[0153] The prestack seismic inversion model design method based on one-dimensional GRU combined with two-dimensional improved ASPP of the present invention specifically comprises the following steps:

[0154] Step S1, build inversion framework

[0155] The entire network framework consists of an inversion network and a forward modeling network. The seismic data is input into the inversion network, and the S-wave velocity, P-wave velocity and density are output. In order to improve the network's ability to learn the global information relevance of seismic data, a recurrent network structure is used to input the S-wave velocity, P-wave velocity and density obtained by the inversion network into the forward modeling network to obtain seismic data.

[0156] The inversion network consists of two parts, namely the feature extraction module and the regression module. The feature extraction module is divided into a local feature extraction module and a sequence modeling module. The structure of the inversion network is as follows: Figure 1 shown.

[0157] During network training, the input of the local feature extraction module in the inversion network is 2D seismic data, namely the seismic data of the wellside gather and the seismic data of the nearby 2k gathers centered on it. The local feature extraction module extracts the spatial and temporal features of the input seismic data through the improved ASPP module.

[0158] The input of the sequence modeling module is one-dimensional seismic data. The temporal characteristics are calculated according to the changes in time and the long-range correlation between seismic gathers is extracted.

[0159] The regression module maps the data from the spatiotemporal feature sequence to shear wave velocity, longitudinal wave velocity and density. The attribute loss is calculated by comparing the real shear wave velocity, longitudinal wave velocity and density with the inverted shear wave velocity, longitudinal wave velocity and density. At the same time, the seismic data loss is calculated by comparing the real seismic data with the seismic data output by the forward modeling network. The attribute loss and seismic data loss jointly guide the network to update the gradient.

[0160] Step S2, network model

[0161] S2-1, feature extraction module

[0162] The feature extraction module includes a local feature extraction module and a sequence modeling module. The local feature extraction module is composed of an improved ASPP module, and its output is the high-frequency components of shear wave velocity, longitudinal wave velocity and density.

[0163] The original ASPP module structure is shown in Figure 2(a), where the input is sent to three parts: convolutional pooling layer, pooling pyramid, and adaptive mean pooling layer. The convolutional pooling layer consists of a 1×1 pooling layer and an upsampling part. The pooling pyramid consists of three dilated convolution blocks, which are composed of 3×3 dilated convolutions, activation functions, and batch normalization. The dilation factor of the convolution can be customized to achieve free multi-scale feature extraction. The adaptive mean pooling layer can obtain an output of a specified size. By inputting the original size and the target size, the size of the convolution kernel and the step size of each movement are adaptively calculated. The outputs of the three parts are spliced ​​and then passed through the convolution block to obtain the output data.

[0164] In this embodiment, the original ASPP module is improved. The improved ASPP module is shown in FIG2(b) and is named IASPP (Improved Atrous Spatial Pyramid Pooling). The IASPP module feeds the input into three parts: a convolutional pooling layer, a pooling pyramid, and an adaptive mean pooling layer. The convolutional pooling layer no longer uses a 1×1 convolution block, but is composed of a 1×5 convolution, a 3×3 pooling layer, and two group normalization layers. Compared with the 1×1 convolution kernel, the 1×5 convolution kernel has a larger width, which can capture longer local information on the input data. And the 1×5 convolution kernel can reduce the number of parameters of the model to a certain extent, thereby reducing the computational and storage costs. The 1×5 convolution kernel can also extract more spatial features in the width direction of the input data, which can improve the lateral continuity of the inversion result. Finally, because the 1×5 convolution kernel has a larger receptive field, it can capture the information of the input data in a larger range and increase the nonlinear ability of the model. Therefore, in this embodiment, a 1×5 convolution kernel is used instead of a 1×1 convolution kernel.

[0165] Different from ASPP, IASPP in this embodiment first concatenates the outputs of the pooling pyramid and the adaptive mean pooling layer, and passes through a convolution block and a dropout layer. The dropout layer can effectively prevent overfitting. Then it is concatenated with the convolution pooling layer, which can not only make more full use of the original features of the seismic data, but also extract deep features. In addition, this embodiment adds a triplet attention module before the output of the local feature extraction module to suppress irrelevant information in the input data and highlight effective information. Its structure is as follows: Figure 3 shown.

[0166] Triplet Attention is an attention mechanism used to handle relationship modeling and feature extraction in sequence data. It consists of three branches. The first branch is the channel attention calculation branch, such as Figure 3In the middle blue part, the input features are pooled, then 7×7 convolution, and finally the sigmoid activation function is used to generate the spatial attention weight. The second branch is the channel C and spatial W dimension interaction capture branch, such as Figure 3 In the green part, the input dimension feature is C×H×W. It is first permuted to become H×C×W dimension feature. Then, it is pooled on the H dimension. The subsequent operations are similar to the first branch. Finally, it is permuted to become C×H×W dimension feature. The third branch is the channel C and space H dimension interactive capture branch, as shown in Figure 3 In the purple part, the input features are first permuted to become W×H×C dimensional features, and then pooled in the W dimension. The subsequent operations are similar to the second branch. After each branch generates the attention weight, the outputs of the three branches are averaged (Avg) to finally obtain the triple attention output.

[0167] This structure can integrate information from different dimensions through different permutation and arrangement operations to better capture the intrinsic characteristics of the data. At the same time, this method is computationally efficient and can be added as a module to the existing network architecture to enhance the network's understanding and processing capabilities of complex data structures.

[0168] Traditional attention mechanisms may have problems with poor information transmission when processing long sequences. Compared with traditional attention mechanisms, Triplet Attention has a more comprehensive modeling and can better model and process long-distance dependencies. Triplet Attention calculates the similarity between three elements and comprehensively considers the interactions between multiple elements. It can better extract important features in the sequence and obtain richer and more accurate feature representations. And when calculating the attention weight, it can clearly indicate the relationship between each element and other elements. This clear relationship representation makes Triplet Attention more interpretable.

[0169] The sequence modeling module is composed of three layers of bidirectional gate recurrent units (B-GRU). B-GRU has the function of modeling the input as sequential data, and has a simple structure and high computational efficiency. Therefore, in this embodiment, B-GRU is used to capture long-range dependencies, calculate time features based on time changes, and extract long-range correlations between seismic traces. Since the deep B-GRU can produce smooth output, the output of the sequence modeling module can be regarded as the low-frequency components of shear wave velocity, longitudinal wave velocity and density.

[0170] S2-2, regression module

[0171] The regression module is composed of two deconvolution blocks, LSTM layers and fully connected layers, such as Figure 4 shown.

[0172] The deconvolution block consists of a deconvolution layer and a group normalization layer. Group normalization groups the output of the convolution layer and normalizes each group using the learned mean and standard deviation, which can reduce covariate shift in the learned features and speed up learning. The deconvolution block has learnable kernel parameters to compensate for the resolution mismatch between seismic data and well logging data, and has been widely used in various fields such as semantic segmentation and image processing. The LSTM layer and the fully connected layer send the features extracted by the feature extraction module back to the target domain.

[0173] The input of the regression module is the output of the feature extraction module, and the output is the inverted shear wave velocity, compressional wave velocity and density. The mean square error between the well logging shear wave velocity, compressional wave velocity, density and the shear wave velocity, compressional wave velocity, density output by the regression module is calculated, that is, the mean square error between the inverted data and the well logging data is calculated according to formula (9) to update the learnable parameters in the feature extraction module and the regression module, which is recorded as attribute loss.

[0174]

[0175] Among them, m i,t is the actual logging data, is the inverted data, L1 represents the attribute loss, and MSE is the mean square error function.

[0176] S2-3, forward network

[0177] The input of the forward modeling module is the output of the inversion network, and the output is the seismic data, such as Figure 5 As shown in Figure 2, the forward modeling module consists of four convolutional layers and two activation functions to complete the reconstruction task.

[0178] The mean square error between the input of the sequence modeling module and the output of the forward modeling module is calculated, that is, the mean square error between the input seismic data and the seismic data obtained by the forward modeling network is calculated according to formula (10), which is recorded as seismic data loss.

[0179]

[0180] Among them, x i,t Well logging seismic data is input to the sequence modeling module. is the seismic data obtained by the forward modeling network, and L2 represents the loss of seismic data.

[0181] Step S3, loss function

[0182] The loss of the entire inversion process is composed of attribute loss and seismic data loss. The total loss is shown in formula (11).

[0183] L=αL1+βL2 (11)

[0184] Among them, α and β are weight coefficients, which are used to control the proportion of attribute loss and seismic data loss in the total loss function.

[0185] Combining equations (9), (10) and (11), we can get the loss function of the entire network as shown in equation (12).

[0186]

[0187] Experimental verification

[0188] In order to quantitatively evaluate the accuracy and reliability of the inversion results, the determination coefficient (R 2 ), Mean Squared Error (MSE) and Pearson's Correlation Coefficient (PCC) are used as evaluation criteria.

[0189] (1) The coefficient of determination is used to evaluate the goodness of fit between variables. Its value is between 0 and 1. The larger its value is, the greater the goodness of fit between the variables is, the higher the degree of explanation of the independent variable on the dependent variable is, and the higher the percentage of the change caused by the independent variable in the total change is, which can be expressed as shown in formula (13).

[0190]

[0191] where y i Represents actual data, represents the inverted data, Represents the average value of the actual data, and N is the total number of sampling points. The numerator represents the sum of the squares of the difference between the true value and the inversion result, similar to the mean square error. The denominator represents the sum of the squares of the difference between the true value and the mean, similar to the variance.

[0192] (2) The mean square error function is the average value of the sum of squares of the errors between the inverted data and the corresponding points of the real data. The smaller the value, the better the fit between the inverted data and the original data. It can be expressed as shown in formula (14).

[0193]

[0194] where y i Represents actual data, represents the data obtained by inversion, and N is the total number of sampling points.

[0195] (3) The Pearson correlation coefficient is a linear correlation coefficient and is the most commonly used correlation coefficient. It is used to reflect the degree of linear correlation between two variables. Its value is between -1 and 1. The larger the absolute value, the stronger the correlation. It can be expressed as shown in formula (15).

[0196]

[0197] where x i and i represent the real data and the inverted data respectively, and They represent the means of the corresponding data, and N is the total number of sampling points. When PCC is 0, it means that the two variables are unrelated, when it is between -1 and 0, it means that the two are negatively correlated, and when it is between 0 and 1, it means that the two are positively correlated.

[0198] Model data testing

[0199] 1) Marmousi2 model inversion experiment

[0200] The Marmousi2 model is a commonly used model in geophysics. It is used to simulate the structure and properties of underground strata and is widely used in the fields of seismic exploration and seismic imaging. The Marmousi2 model is a two-dimensional model. The model is constructed based on geological data from the Marmousi region in southern France and includes changes in multiple strata and media. These changes in strata and media can be used to simulate phenomena such as reflection, refraction, and scattering of seismic waves when they propagate underground. The Marmousi2 model is 17 km wide and 3.5 km deep. It has complex stratum structures and media changes, including different types of rocks, sediments, and faults. It is one of the important standard models for verification and performance evaluation of seismic exploration and seismic imaging algorithms.

[0201] The prestack seismic section of the Marmousi2 model is as follows Figure 6 As shown, Figure 6 (a) is the small angle seismic data stacked from 0° to 13°. Figure 6 (b) is the medium-angle seismic data stacked from 13° to 26°. Figure 6 (c) is the large-angle seismic data stacked from 26° to 40°. The number of seismic traces in the seismic profile is 3400, and the number of sampling points in each seismic trace is 300. The true shear wave velocity, longitudinal wave velocity and density profile are as follows: Figure 7(a), (b) and (c), the number of traces is 3400, and the number of sampling points in each trace is 1200. This embodiment uses a Gaussian matrix with a standard deviation of 0 and a size of 299×299 to smooth the true shear wave velocity, longitudinal wave velocity and density of the Marmousi2 model, and obtains the initial model used in this embodiment, which is shown as follows: Figure 7 As shown in (d), (e) and (f).

[0202] 12 traces are extracted from the shear wave velocity, compression wave velocity and density models as pseudo-well data, and the seismic data corresponding to the pseudo-well curve is used as the input of the sequence modeling module. For each pseudo-well curve, 2k+1 seismic traces are taken with the wellside traces as the center and k as the radius. In this embodiment, k is set to 2, and each pseudo-well curve corresponds to 5 seismic traces. The depth is 1200 sampling points, and the obtained seismic data is used as the input of the local feature extraction module.

[0203] The network was trained using seismic data and pseudo-well data. The iteration cycle was set to 700 and the batch size of each iteration cycle was 12. ADAM was selected as the optimizer and the learning rate was set to 0.005. In each training iteration, the weight coefficients in the loss function were set to α=1, β=0.2, the loss function was calculated and the network parameters were updated through network back propagation. The network training was implemented in the PyTorch framework, and GPU was used to accelerate the calculation. Finally, the trained network was used to perform shear wave velocity, longitudinal wave velocity and density inversion to obtain the inversion results.

[0204] In order to prove the effectiveness of the method proposed in this embodiment, the results of the inversion method of this embodiment are compared with the results of the commonly used deep learning inversion methods, including the conventional CNN network inversion method, the one-dimensional TCN-based inversion method and the CNN-GRU network-based semi-supervised inversion method. The above three networks and the method proposed in this embodiment use the same hyperparameters during training, and the experiment is carried out under the same training data and training conditions. The results of the CNN network inversion of shear wave velocity, longitudinal wave velocity and density are shown in Figure 2. Figure 8 a(1), a(2), and a(3) are shown in the figure. The corresponding absolute error profiles are shown in Fig. 9 a(1), a(2), and a(3). The results of inversion of shear wave velocity, longitudinal wave velocity, and density based on the one-dimensional TCN network are shown in Figure 8 b(1), b(2), and b(3) are shown in the figure. The corresponding absolute error profiles are shown in Fig. 9 b(1), b(2), and b(3). The results of inversion of shear wave velocity, longitudinal wave velocity, and density based on the CNN-GRU network are shown in Figure 2. Figure 8 c(1), c(2), and c(3) are shown in the figure. The corresponding absolute error profiles are shown in Fig. 9As shown in c(1), c(2), and c(3).

[0205] In order to verify the effectiveness of the improvements to the ASPP module, the initial model constraints, and the Triplet Attention module in this embodiment, the inversion effects before and after adding the above modules are compared. The method combining the ASPP module and GRU is denoted as ASPP-GRU. The inversion results of shear wave velocity, longitudinal wave velocity, and density are shown in Figure 2. Figure 8 d(1), d(2), and d(3) are shown in the figure. The corresponding absolute error profiles are shown in Fig. 9 d(1), d(2), and d(3). The method combining the IASPP module and the GRU module is denoted as IASPP-GRU. The inversion results of shear wave velocity, longitudinal wave velocity, and density are shown in Figure 8 e(1), e(2), and e(3), and the corresponding absolute error profiles are as follows Fig. 9 e(1), e(2), and e(3). On the basis of the above inversion network, an initial model is added to constrain the network. The shear wave velocity, longitudinal wave velocity, and density low-frequency initial model profiles are shown as follows: Figure 7 (d), (e), and (f) show that this method is denoted as IASPP-GRU-init. The inversion results of shear wave velocity, longitudinal wave velocity, and density are shown in Figure 2. Figure 8 f(1), f(2), and f(3) are shown in the figure. The corresponding absolute error profile is shown in the figure Fig. 9 f(1), f(2), and f(3) are shown in Figure 1. The final method proposed in this embodiment is obtained by adding the Triplet Attention module to the network model with the initial model added above, which is denoted as IGIT. The inversion results of shear wave velocity, longitudinal wave velocity, and density are shown in Figure 1. Figure 8 g(1), g(2), and g(3) are shown in the figure. The corresponding absolute error profiles are shown in the figure Fig. 9 As shown in g(1), g(2), and g(3).

[0206] from Figure 8 It can be seen that the inversion results of the method proposed in this embodiment are more similar to the real model. Compared with the other three methods used for comparative experiments, the inversion profile obtained by the method proposed in this embodiment has stronger horizontal continuity, weaker visible jitter in the horizontal and vertical directions, and more accurate inversion results in details. In particular, for the 500th to 750th seismic gathers, the inversion effect of sudden changes in the strata near the 220 sampling points in the depth direction is better. Fig. 9 is the absolute error profile. The more white parts are in the figure, the greater the error is. Fig. 9As shown, the main locations where errors occur are the interfaces of different strata and the locations of faults. Compared with other methods, the method of this embodiment can also invert the fold structure in the model well, with smaller errors, more continuous inversion results, more accurate inversion of the locations of faults, higher similarity with the actual stratum conditions, and lower errors at most locations than other inversion methods. This is due to the effective extraction of multi-scale features and the effective use of horizontal spatial correlation in this embodiment.

[0207] By comparing the inversion results and absolute error profiles of ASPP-GRU, IASPP-GRU, IASPP-GRU-init and the method of this embodiment, it can be seen that after changing the connection order and method of the three parts in ASPP and adding the initial model and triple attention module, the inversion accuracy is improved and the inversion details are clearer, which verifies the effectiveness of the improvements made to the network model in this embodiment.

[0208] In order to compare the details of the inversion results of different methods from a microscopic perspective, the representative 200th seismic trace gather is selected for inversion. The comparison diagrams of the single-channel inversion results of shear wave velocity, longitudinal wave velocity and density of the conventional CNN network inversion method, the one-dimensional TCN-based inversion method, the semi-supervised inversion method based on the CNN-GRU network and the method proposed in this embodiment are shown in Figure 2. Fig.10 (a), (b), and (c). The black line represents the real data, the red line represents the inversion result of the method proposed in this embodiment, and the line types corresponding to other methods are marked in the figure.

[0209] from Fig.10 It can be seen from (a), (b), and (c) that the conventional CNN network inversion results differ greatly from the true value, the inversion results of the inversion method based on one-dimensional TCN are relatively flat, and the details in the inversion model are not inverted. The semi-supervised inversion method based on the CNN-GRU network has a good inversion result, but the inversion effect is poor in the details and at a deeper depth. The difference between the inversion result of the method in this embodiment and the actual value is smaller than that of other methods, and the true value and the inversion result are almost the same.

[0210] The inversion results of the 200th gather of the method in this embodiment are compared with those of the three methods of ASPP-GRU, IASPP-GRU and IASPP-GRU-init. Fig.10 As shown in (d), (e), and (f), it can be seen from the figure that, compared with the other three methods, the method proposed in this embodiment has improved accuracy in most positions, especially at a deeper depth, the effect is more obvious, which further verifies the effectiveness of this embodiment in improving the network model.

[0211] In order to objectively and quantitatively evaluate the inversion results of each method, this example uses R 2, MSE and PCC are used as evaluation criteria, and Table 1 shows the evaluation indicators between the inversion results of shear wave velocity, longitudinal wave velocity and density of different methods and the real data. It can be seen from the table that the method proposed in this embodiment is superior to other methods in the three evaluation criteria and in the inversion results of the three parameters of shear wave velocity, longitudinal wave velocity and density, especially the inversion index of density is far superior to other methods, which verifies the effectiveness of the method of this embodiment.

[0212] Table 1 R between the inversion results of the Marmousi2 model and the actual data 2 , MSE and PCC

[0213]

[0214] 2) SEAM model inversion experiment

[0215] To further verify the performance of the method proposed in this embodiment, the SEAM model was used for experiments. The SEAM model is a three-dimensional seismic exploration model, and its transverse changes in shear wave velocity, longitudinal wave velocity and density are very drastic, and the inversion difficulty is higher than that of the Marmousi2 model. The SEAM model is constructed based on basic rock properties such as the volume of rock and sand. Its change trend follows the shale pore characteristics of the Gulf of Mexico, and can better simulate the actual geological conditions. It has been widely used in the field of seismic inversion.

[0216] The multi-angle pre-stack partial stack seismic profile of the SEAM model is as follows Fig.11 shown. Fig.11 (a) is the small angle seismic data stacked from 0° to 13°. Fig.11 (b) is the medium-angle seismic data stacked from 13° to 26°. Fig.11 (c) is the large-angle seismic data stacked from 26° to 40°. The number of seismic traces in each partial stacked section is 1751, and the number of sampling points in each seismic trace is 1501. The true shear wave velocity profile, longitudinal wave velocity profile and density profile are shown in Figure 2. Fig.12 (a), (b), and (c). The number of traces in the shear wave velocity, longitudinal wave velocity, and density profiles is also 1751, and the number of sampling points in each trace is 1501. This embodiment uses a Gaussian matrix with a standard deviation of 0 and a size of 251×251 to smooth the true shear wave velocity, longitudinal wave velocity, and density of the SEAM model, and obtains the initial model used in this embodiment, which is shown as follows: Fig.12(d), (e), and (f). 12 data are uniformly extracted from the shear wave velocity, compressional wave velocity, and density models as pseudo-well data. For each pseudo-well curve, 2k+1 seismic gathers with the wellside gather as the center and k as the radius are taken. K is also set to 2, so that each pseudo-well logging curve corresponds to 5 gathers of seismic data. The network is then trained using the same method as the Marmousi2 model, and the trained network is used for shear wave velocity, compressional wave velocity, and density inversion.

[0217] The inversion results are as follows Fig.13 As shown, Fig.13 a(1), a(2) and a(3) are the inversion results of shear wave velocity, longitudinal wave velocity and density by the conventional CNN network inversion method, respectively. The corresponding absolute error profiles are shown in Fig.14 shown in a(1), a(2) and a(3). Fig.13 b(1), b(2) and b(3) are the shear wave velocity, longitudinal wave velocity and density results based on the one-dimensional TCN network inversion, respectively. The corresponding absolute error profiles are shown in Fig.14 As shown in b(1), b(2) and b(3). Fig.13 c(1), c(2) and c(3) are the shear wave velocity, longitudinal wave velocity and density results inverted based on the CNN-GRU network, respectively. The corresponding absolute error profiles are shown in Fig.14 As shown in c(1), c(2) and c(3).

[0218] The inversion results of shear wave velocity, longitudinal wave velocity and density of the ASPP-GRU method are shown in Fig.13 d(1), d(2), and d(3) are shown in the figure. The corresponding absolute error profiles are shown in Fig.14 d(1), d(2), and d(3). The inversion results of shear wave velocity, longitudinal wave velocity, and density using the IASPP-GRU method are shown in Fig.13 e(1), e(2), and e(3), and the corresponding absolute error profiles are as follows Fig.14 e(1), e(2), and e(3). The profile of the initial model used by the IASPP-GRU-init method is shown in Fig.12 (d), (e) and (f), the inversion results of shear wave velocity, longitudinal wave velocity and density are shown in Fig.13 f(1), f(2), and f(3) are shown in the figure. The corresponding absolute error profile is shown in the figure Fig.14 The results of inverting the shear wave velocity, longitudinal wave velocity and density by the method proposed in this embodiment are shown as follows: Fig.13 g(1), g(2), and g(3) are shown in the figure. The corresponding absolute error profiles are shown in the figure Fig.14 As shown in g(1), g(2), and g(3).

[0219] from Fig.13and Fig.14 It can be seen that the inversion results of the method proposed in this embodiment are more similar to the real model, and the changes in the stratum structure can be clearly seen in the inversion result diagram, and the inversion of the details is also relatively clear, while the other three methods do not outline the stratum changes clearly enough and the details are not rich enough. In particular, the inversion effect of the complex stratum changes between the 700th and 1200th trace sets is significantly improved. Although the method proposed in this embodiment still has certain errors at the stratum boundary, it is better than other methods overall, which verifies the effectiveness of the method proposed in this embodiment.

[0220] By comparing the inversion results and absolute error profiles of ASPP-GRU, IASPP-GRU, IASPP-GRU-init and the method of this embodiment, it can be seen that after changing the connection order and method of the three parts in ASPP and adding the initial model and triple attention module, the inversion accuracy is improved, the inversion details are richer, and the effectiveness of the improvement made to the network model in this embodiment is further verified.

[0221] In order to compare the details of the inversion results of different methods from a microscopic perspective, the representative 200th seismic gather is selected for inversion. The comparison diagrams of the inversion of shear wave velocity, longitudinal wave velocity and density of the conventional CNN network inversion method, the inversion method based on the one-dimensional TCN network, the semi-supervised inversion method based on the CNN-GRU network and the method proposed in this embodiment are shown in Figure 2. Fig.15 (a), (b), and (c). The black line represents the real data, the red line represents the inversion data of the method proposed in this embodiment, and the line types corresponding to other methods are marked in the figure.

[0222] from Fig.15 It can be seen from (a), (b), and (c) that the conventional CNN network inversion results differ greatly from the true values. The inversion accuracy of the inversion method based on the one-dimensional TCN network has been improved, but the inversion effect is not good at the place where the formation changes drastically, and it is difficult to invert the detailed information. The semi-supervised inversion method based on the CNN-GRU network has a good inversion result, but the inversion effect is poor at the details and the formation boundary. The difference between the inversion result of the method in this embodiment and the actual value is smaller than that of other methods, and the inversion result is highly consistent with the true value.

[0223] The inversion results of the 200th gather of the method in this embodiment are compared with those of the three methods of ASPP-GRU, IASPP-GRU and IASPP-GRU-init. Fig.15 As shown in (d), (e), and (f), it can be seen from the figure that compared with the other three methods, the method proposed in this embodiment has improved accuracy in details, especially when the depth is deep, the inversion effect is obvious, which further proves the effectiveness of this embodiment in improving the network model.

[0224] In order to objectively and quantitatively evaluate the inversion results of each method, this example uses R 2 , MSE and PCC are used as evaluation criteria, and Table 2 shows the evaluation indicators between the inversion results of shear wave velocity, longitudinal wave velocity and density of different methods and the real data. It can be seen from the data in the comparison table that the method proposed in this embodiment is superior to other methods in the three evaluation criteria and in the inversion results of the three parameters of shear wave velocity, longitudinal wave velocity and density, which verifies the effectiveness of the method in this embodiment.

[0225] Table 2 R between SEAM model inversion results and actual data 2 , MSE and PCC

[0226]

[0227] This embodiment proposes a seismic pre-stack inversion method based on one-dimensional GRU combined with two-dimensional improved ASPP. The overall network structure consists of a forward network and an inversion network, wherein the inversion network consists of a feature extraction module and a regression module. In the inversion module, this embodiment adopts a multi-channel set inversion method to make full use of the spatial information in the seismic data and improve the lateral continuity of the inversion result. At the same time, this embodiment changes the connection order and mode of the three components of the ASPP module, so that it can fully extract the multi-scale features of the seismic data while paying attention to the long-range information in the seismic data, thereby improving the accuracy of the inversion result. And the Triplet Attention module is added to the improved ASPP module, which fully considers the spatial correlation and channel correlation of the seismic data and allows the network to better handle the long-distance relationship of the data.

[0228] The method proposed in this embodiment is applied to the Marmousi2 model and the SEAM model, and compared with other methods. The experimental results show that the method proposed in this embodiment has higher inversion accuracy, richer details of the inversion results, and better lateral continuity. A series of experiments on the Marmousi2 model and the SEAM model have demonstrated the effectiveness of the present embodiment in improving the network model. Experiments on different data have shown that the inversion results of the method proposed in this embodiment not only have better lateral continuity, but also better inversion effects in details, especially significantly improving the inversion effect of density parameters. The method proposed in this embodiment has made useful explorations for improving the effect of pre-stack seismic inversion.

[0229] The above implementation cases are only preferred embodiments of the present invention and are not intended to limit the present invention in any form. Although the present invention has been disclosed as above with preferred embodiments, they are not intended to limit the present invention. Therefore, any simple modification, equivalent changes and modifications made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the scope of protection of the technical solution of the present invention.

Claims

1. A prestack seismic inversion model design method based on one-dimensional GRU combined with two-dimensional improved ASPP, characterized in that: The atrous spatial convolutional pooling pyramid module is introduced into the prestack inversion framework, and the improved ASPP sends the input to three parts: convolutional pooling layer, pooling pyramid and adaptive mean pooling layer; the outputs of the pooling pyramid and the adaptive mean pooling layer are first spliced, passed through a convolution block and a dropout module, and then spliced ​​with the convolutional pooling layer; a triple attention module is added to the inversion framework to extract features of seismic data at different scales, and a recurrent neural network is used to extract low-frequency information of seismic data; in the network training stage, a multi-channel set inversion method is adopted, combining one-dimensional inversion and two-dimensional inversion to train the network model.

2. The prestack seismic inversion model design method based on one-dimensional GRU combined with two-dimensional improved ASPP according to claim 1 is characterized in that: The design method specifically includes the following contents: The overall network framework consists of an inversion network and a forward modeling network. The seismic data is input into the inversion network, and the S-wave velocity, P-wave velocity and density are output. Using a cyclic network structure, the S-wave velocity, P-wave velocity and density obtained by the inversion network are input into the forward modeling network to obtain the seismic data. The inversion network includes a feature extraction module and a regression module. The feature extraction module is divided into a local feature extraction module and a sequence modeling module. During network training, the input of the local feature extraction module in the inversion network is two-dimensional seismic data, and the input of the sequence modeling module is one-dimensional seismic data. The regression module maps data from the spatiotemporal feature sequence to shear wave velocity, longitudinal wave velocity and density; calculates the attribute loss by comparing the true shear wave velocity, longitudinal wave velocity, density and the inverted shear wave velocity, longitudinal wave velocity and density; and calculates the seismic data loss by comparing the true seismic data and the seismic data output by the forward modeling network.

3. The prestack seismic inversion model design method based on one-dimensional GRU combined with two-dimensional improved ASPP according to claim 2 is characterized in that: The convolutional pooling layer consists of a 1×5 convolution, a 3×3 pooling layer and two group normalization layers.

4. The prestack seismic inversion model design method based on one-dimensional GRU combined with two-dimensional improved ASPP according to claim 3 is characterized in that: A triple attention module is added before the output of the local feature extraction module. The triple attention module consists of three branches. The first branch is a channel attention calculation branch. The input feature is pooled, then 7×7 convolution, and finally a Sigmoid activation function is used to generate spatial attention weights. The second branch is the channel C and space W dimension interactive capture branch. The input dimension feature is C×H×W. It is first permuted to become H×C×W dimensional feature, then pooled on the H dimension, and finally permuted to become C×H×W dimensional feature. The third branch is the channel C and space H dimension interactive capture branch. The input feature is first permuted to become W×H×C dimensional feature, and then pooled on the W dimension. After each branch generates the attention weight, the outputs of the three branches are averaged and aggregated to obtain the triple attention output.

5. The prestack seismic inversion model design method based on one-dimensional GRU combined with two-dimensional improved ASPP according to claim 2, characterized in that: The regression module is composed of two deconvolution blocks, an LSTM layer and a fully connected layer. The deconvolution block is composed of a deconvolution layer and a group normalization layer. The group normalization groups the output of the convolution layer and normalizes each group using the learned mean and standard deviation. The deconvolution block has learnable kernel parameters for compensating for the resolution mismatch between seismic data and well logging data; the LSTM layer and the fully connected layer send the features extracted by the feature extraction module back to the target domain.

6. The prestack seismic inversion model design method based on one-dimensional GRU combined with two-dimensional improved ASPP according to claim 5, characterized in that: The input of the regression module is the output of the feature extraction module, and the output is the inverted shear wave velocity, longitudinal wave velocity and density; Calculate the mean square error between the well logging shear wave velocity, compressional wave velocity, density and the shear wave velocity, compressional wave velocity, density output by the regression module. Calculate the mean square error between the inverted data and the well logging data according to formula (9). Update the learnable parameters in the feature extraction module and the regression module, which are recorded as attribute loss: Among them, m i,t is the actual logging data, is the inverted data, L1 represents the attribute loss, and MSE is the mean square error function.

7. The prestack seismic inversion model design method based on one-dimensional GRU combined with two-dimensional improved ASPP according to claim 6, characterized in that: The input of the forward modeling module is the output of the inversion network, and the output is seismic data. The forward modeling module consists of four convolutional layers and two activation functions to complete the reconstruction task; The mean square error between the input of the sequence modeling module and the output of the forward modeling module is calculated. The mean square error between the input seismic data and the seismic data obtained by the forward modeling network is calculated according to formula (10), which is recorded as seismic data loss: Among them, x i,t Well logging seismic data is input to the sequence modeling module. is the seismic data obtained by the forward modeling network, and L2 represents the loss of seismic data.

8. The prestack seismic inversion model design method based on one-dimensional GRU combined with two-dimensional improved ASPP according to claim 7, characterized in that: The loss of the inversion process is composed of attribute loss and seismic data loss. The total loss is shown in formula (11): L=αL1+βL2(3) Among them, α and β are weight coefficients, which are used to control the proportion of attribute loss and seismic data loss in the total loss function. The loss function of the entire inversion network is shown in formula (12):

Images