Non-stationary seismic data attenuation compensation method and system based on spatio-temporal neural network

Abstract

The present invention discloses a non-stationary seismic data attenuation compensation method and system based on a spatio-temporal neural network. The estimated shallow seismic wavelet and the actual logging reflection coefficient sequence are synthesized into a seismic record trace at the logging position through a convolution model. Based on the non-stationary seismic record and the synthesized seismic record trace at the logging position, a data set considering the spatio-temporal characteristics of the seismic record is constructed. A spatio-temporal neural network structure is designed for non-stationary seismic data attenuation compensation. The constructed data set is used to train the established spatio-temporal neural network structure to determine the optimal parameter set of the spatio-temporal neural network. Based on the trained spatio-temporal neural network structure and the optimal parameter set, the non-stationary seismic trace is mapped into a stationary seismic trace, and the two-dimensional stationary seismic data estimated by the network is obtained. The present invention considers the spatio-temporal structural characteristics of seismic data, and realizes the attenuation compensation of non-stationary seismic data by training the spatio-temporal neural network, thereby improving the vertical resolution of deep seismic data.

Description

Technical Field

[0001] The present invention belongs to the technical field of seismic exploration, and particularly relates to a non-stationary seismic data attenuation compensation method and system based on a spatio-temporal neural network. Background Art

[0002] In order to clearly show the underground geological structure, lithological changes and fluid characteristics, it is particularly important to improve the vertical resolution of deep seismic data. Seismic deconvolution (SD) is a technique aimed at establishing a high-resolution reflectivity model of the underground medium using band-limited seismic data within an inversion framework.

[0003] For traditional seismic deconvolution methods described by convolution models, each stacked seismic trace can be regarded as the convolution of a seismic wavelet and a reflectivity. The establishment of this model is based on a set of assumptions, and three key assumptions are:

[0004] 1) The underground medium is composed of horizontal layers with invariant impedance;

[0005] 2) The incident angle of the seismic waves generated by the seismic source is 0 when hitting the boundary;

[0006] 3) The source wavelet does not change when moving underground.

[0007] However, in reality, the underground medium is complex, and the medium is usually heterogeneous and anisotropic, resulting in the invalidation of the first two assumptions. Due to the viscoelastic properties of the underground medium, it will also cause problems such as the decrease in the amplitude of seismic data with increasing depth, frequency dispersion and phase shift. In this case, the seismic wavelet should be considered non-stationary, which obviously violates the third assumption. Therefore, traditional seismic deconvolution methods based on convolution models do not work in complex situations.

[0008] Therefore, in a complex underground environment, the viscoelasticity of the underground medium seriously reduces the resolution of seismic data. Therefore, fully solving the viscoelasticity problem of the medium is one of the most important steps to improve the vertical resolution of deep seismic data.

[0009] There are many existing methods to improve the resolution of the underground medium, mainly including the following three: the inverse Q filtering method, the variable wavelet inversion method based on reflection seismic records, and the deep learning method based on the LSTM network. However, all three methods have a common drawback, that is, they cannot consider the spatial correlation of seismic data, which makes the spatial continuity of the compensated seismic data poor. Summary of the Invention

[0010] The technical problem to be solved by the present invention is to provide a non-stationary seismic data attenuation compensation method and system based on a spatio-temporal neural network for solving the technical problem of reducing the vertical resolution of mid-deep seismic data due to the viscoelasticity of underground media in view of the above-mentioned prior art.

[0011] The present invention adopts the following technical solutions:

[0012] The non-stationary seismic data attenuation compensation method based on a spatio-temporal neural network includes the following steps:

[0013] S1. Collect non-stationary post-stack reflected seismic data, and estimate shallow seismic wavelets by a statistical method; synthesize a stationary seismic record trace at the well logging position through the convolution model with the estimated shallow seismic wavelets and the actual well logging reflection coefficient sequence.

[0014] S2. Based on the non-stationary seismic record and the stationary synthetic seismic record trace obtained in step S1, construct a data set considering the spatio-temporal characteristics of the seismic record.

[0015] S3. Design a spatio-temporal neural network structure for non-stationary seismic data attenuation compensation.

[0016] S4. Based on the data set constructed in step S2, train the spatio-temporal neural network structure established in step S3 to determine the optimal parameter set of the spatio-temporal neural network.

[0017] S5. Based on the trained spatio-temporal neural network structure and the optimal parameter set obtained in step S4, map the non-stationary seismic trace to a stationary seismic trace to obtain the two-dimensional stationary seismic data estimated by the network.

[0018] Specifically, in step S3, constructing a data set considering the spatio-temporal characteristics of the seismic record is specifically as follows:

[0019] S301. Randomly generate N well points uniformly distributed within the spatial range of M*M.

[0020] S302. Randomly select several well points among them as the label points of a two-dimensional data according to the distance and angle.

[0021] S303. Randomly insert data points to generate a single interpolation sequence.

[0022] S304. According to the data point positions and the original synthetic seismic data obtained in steps S302 and S303, generate three-dimensional seismic data with attenuation and corresponding non-attenuated semi-label data.

[0023] Furthermore, before constructing the data set considering the spatio-temporal characteristics of the seismic record, the seismic data is preprocessed as follows:

[0024] During input, the seismic data values are uniformly magnified by 10 times, and during prediction, the prediction results are divided by 10; the input 3D attenuated seismic data is subjected to dimension expansion processing, expanding one dimension from the time dimension to form a two-dimensional spatio-temporal signal at a single time point.

[0025] Specifically, in step S4, the 3DCNN network and the ConvLSTM network are connected in a horizontal Concat connection manner to obtain a C3D_CLN spatio-temporal neural network structure.

[0026] Furthermore, the working principle of the ConvLSTM network is as follows:

[0027]

[0028]

[0029]

[0030]

[0031]

[0032] Among them, χ t is the input value at time t, C t is the state value at time t, H t is the hidden state value at time t, i t , f t , o t are intermediate values, W * and b * represent the corresponding convolution kernels and biases, '*' represents the convolution symbol, represents the Hadamard product, and 'σ' and 'tanh' represent non-linear functions.

[0033] Furthermore, the working principle of the 3DCNN network is as follows:

[0034]

[0035] Among them, represents the value of the (x, y, z) point of the jth feature map in the ith layer, 'tanh' represents the non-linear function, represents the value of the mth layer convolution kernel at the (p, q, r) point, b ij represents the bias.

[0036] Furthermore, the C3D_CLN spatio-temporal neural network structure uses Adam as the network optimizer, and the initial learning rate is 5e -4, the learning decay rate is 0.98, the number of Epochs is 60, the Loss function uses the mean square error loss with a self-constructed MASK matrix, the Batch_Size is 16, the convolutional kernel size of the 3DCNN layer is (5, 3, 3), the number of Filters is 8, the convolutional kernel size of the ConvLSTM layer is (3, 3), the number of filters is 8, and except for the last non-linear layer which is Linear, other non-linear layers are all Relu.

[0037] Furthermore, incorporate the Inception-v3 and two-stream network structures into the C3D_CLN spatio-temporal neural network structure to obtain the HST network. The TF1 data stream in the HST network is the input attenuated stacked data, and the TF2 data stream is the difference between the stacked data with attenuation at the previous and subsequent time points.

[0038] Specifically, the two-dimensional stationary seismic data estimated by the network is:

[0039]

[0040] where is the trained spatio-temporal neural network structure, represents the two-dimensional non-stationary seismic data composed of multiple observed non-stationary seismic traces.

[0041] In a second aspect, an embodiment of the present invention provides a non-stationary seismic data attenuation compensation system based on a spatio-temporal neural network, including:

[0042] An acquisition module that acquires non-stationary stacked reflection seismic data and estimates shallow seismic wavelets through statistical methods; synthesizes the estimated shallow seismic wavelets and the actual logging reflection coefficient sequence into a stationary seismic record trace at the logging position through a convolution model;

[0043] A data module that constructs a data set considering the spatio-temporal characteristics of the seismic record based on the non-stationary seismic record and the stationary synthetic seismic record trace at the logging position obtained by the acquisition module;

[0044] A design module that designs a spatio-temporal neural network structure for non-stationary seismic data attenuation compensation;

[0045] A training module that performs network training on the spatio-temporal neural network structure established by the design module based on the data set constructed by the data module to determine the optimal parameter set of the spatio-temporal neural network;

[0046] A compensation module that maps non-stationary seismic traces to stationary seismic traces based on the trained spatio-temporal neural network structure and the optimal parameter set obtained by the training module to obtain the two-dimensional stationary seismic data estimated by the network.

[0047] Compared with the prior art, the present invention has at least the following beneficial effects:

[0048] A non-stationary seismic data attenuation compensation method based on a spatio-temporal neural network combines the spatio-temporal relationship of seismic data to solve the problem of reduced vertical resolution of mid-deep seismic data. By comparing the performance of different networks in the compensation effect of post-stack seismic data with attenuation, a new spatio-temporal neural network C3D_CLN is constructed by combining 3DCNN and ConvLSTM networks. The spatio-temporal neural network structure HST is adopted, considering the spatio-temporal structure characteristics of seismic data. Through training the spatio-temporal neural network, the attenuation compensation of non-stationary seismic data is realized, thereby improving the vertical resolution of deep seismic data.

[0049] Furthermore, since there is a lack of effective two-dimensional training labels in the actual data, it is necessary to construct a training data set through random cross-well curves. Among them, the two-dimensional label data only has valid label data at the cross-well channels. This not only expands the amount of trainable data of the network, but also can more fully explore the underground spatio-temporal characteristics of seismic data.

[0050] Furthermore, since the order of magnitude of seismic data is different from the optimal value order of magnitude of network learning, in order to be more conducive to network training and not destroy the spatio-temporal relationship of seismic data, it is necessary to uniformly amplify the seismic data. The reason for time dimension expansion is to increase the amount of trainable data of the network and enhance the robustness of network learning.

[0051] Furthermore, according to the ablation experiment, the following conclusions can be obtained. The ConvLSTM network has strong spatio-temporal learning ability, but has a large dependence on the spatial position of labels and weak generalization ability; while the 3DCNN network has slightly worse learning effect than the ConvLSTM network, but this network has strong generalization ability and weak requirements for labels. Therefore, in order to combine the respective advantages of the two networks, improve the performance of the network and reduce the dependence of the network on the label position, the C3D_CLN spatio-temporal neural network is obtained by combining the two networks in a horizontal connection manner.

[0052] Furthermore, the key of this network is ConvLSTM. The ordinary LSTM network has been proven to be very powerful in dealing with time-correlated data, but it contains too much redundancy for spatial data. The ConvLSTM network is an extension based on the LSTM network. This network has a convolutional structure in both the input-to-state and state-to-state transitions, which enables the data flow to propagate in the network in a two-dimensional or three-dimensional form.

[0053] Furthermore, compared with the ConvLSTM network, the 3DCNN network can reduce the requirements of the network for the label position during the process of dealing with spatio-temporal data and improve the robustness of the network.

[0054] Furthermore, the C3D_CLN spatio-temporal neural network structure uses Adam as the network optimizer, which is more conducive to the decline of network loss. Since the two-dimensional label data only exists at random positions in several channels, the Loss function needs to use the mean square error loss with a self-constructed MASK matrix. Except for the last non-linear layer being Linear, other non-linear layers are all Relu because the Relu non-linear layer can accelerate the backpropagation of the network, and the seismic data output by the network has both positive and negative values, so the last layer needs to use a Linear layer.

[0055] Furthermore, the HST network is built on the basis of the C3D_CLN network structure, and its structure integrates the advantages of other spatio-temporal network structures, improving the network's ability to process spatio-temporal related data.

[0056] Furthermore, the trained network structure and its optimal parameters are applied to non-stationary seismic data compensation. The non-stationary seismic data after network compensation can be approximately regarded as stationary seismic data, ultimately solving the problem of reduced vertical resolution of mid-deep seismic data.

[0057] It can be understood that the beneficial effects of the second aspect can refer to the relevant descriptions in the first aspect above, and will not be elaborated here.

[0058] In summary, the present invention considers the spatio-temporal structural characteristics of seismic data, and realizes the attenuation compensation of non-stationary seismic data by training a spatio-temporal neural network, thereby improving the vertical resolution of deep seismic data.

[0059] Next, through the accompanying drawings and embodiments, the technical solutions of the present invention will be further described in detail. Description of the Drawings

[0060] Figure 1 It is an experimental result diagram of the ConvLSTM network on the three-dimensional data volume 1, where dark gray represents stationary single-channel seismic data, light gray represents non-stationary single-channel seismic data, and medium gray represents the data of non-stationary single-channel seismic data after network compensation. (a) is the effect of a single channel at the test set label position when the training label position is fixed, (b) is the effect of a single channel at the non-label position of the test set when the training label position is fixed, (c) is the effect of most channels of the test set when the training label position is random, and (d) is the effect of several channels of the test set when the training label position is random;

[0061] Figure 2Experimental result diagram of the 3DCNN network on the three-dimensional data volume 1, where dark gray represents stationary single-channel seismic data, light gray represents non-stationary single-channel seismic data, and medium gray represents the data of non-stationary single-channel seismic data after network compensation. (a) shows the effect of a single channel at the test set label position when the training label position is fixed. (b) shows the effect of a single channel at a non-label position in the test set when the training label position is fixed. (c) shows the effect of all channels in the test set when the training label position is random;

[0062] Figure 3 Experimental result diagram of the C3D_CLN network (a network combined only by 3DCNN and ConvLSTM) on the three-dimensional data volume 1 with fixed labels, where dark gray represents stationary single-channel seismic data, light gray represents non-stationary single-channel seismic data, and medium gray represents the data of non-stationary single-channel seismic data after network compensation;

[0063] Figure 4 Experimental result diagram of the training of the C3D_CLN network on the three-dimensional data volume 2. Among them, (a) shows the decline curves of the training loss (dark gray) and the validation loss (light gray) when the number of network training times is 60. (b) shows the compensation results of single-channel seismic data in the test set of the C3D_CLN network, non-attenuated single-channel seismic data (dark gray), single-channel seismic data with attenuation (light gray), and single-channel seismic data after compensation by the C3D_CLN network for single-channel seismic data with attenuation (medium gray);

[0064] Figure 5 Experimental result diagrams of the C3D_CLN and LSTM networks on the test set of the three-dimensional data volume 2 at Epoch = 60 under the same conditions. Among them, (a) shows the learning result of a certain two-dimensional data of the C3D_CLN network. (b) shows the learning result of a certain two-dimensional data of the LSTM network. (c) shows the difference between the learning compensation results of the two networks and the non-attenuated seismic data, the C3D_CLN network (left figure), the LSTM network (right figure);

[0065] Figure 6 Experimental result diagrams of the LSTM, C3D_CLN, and HST networks on the test set of the three-dimensional data volume 2 at Epoch = 200 under the same conditions;

[0066] Figure 7 Internal structure diagram of the ConvLSTM network;

[0067] Figure 8 Structure of the C3D_CLN network;

[0068] Figure 9 Structure of the HST network. Detailed implementation method

[0069] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0070] In the description of the present invention, it should be understood that the terms "include" and "comprise" indicate the presence of the described features, wholes, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or their combinations.

[0071] It should also be understood that the terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include the plural forms.

[0072] It should be further understood that the term "and / or" used in the specification of the present invention and the appended claims refers to any combination and all possible combinations of one or more of the related listed items, and includes these combinations. For example, A and / or B can represent three cases: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " in this article generally represents an "or" relationship between the contextually related objects.

[0073] It should be understood that although the terms first, second, third, etc. may be used in the embodiments of the present invention to describe preset ranges, etc., these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from each other. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be referred to as the second preset range, and similarly, the second preset range may also be referred to as the first preset range.

[0074] Depending on the context, the word "if" as used herein can be interpreted as "when" or "while" or "in response to determining" or "in response to detecting". Similarly, depending on the context, the phrase "if determined" or "if detected (stated condition or event)" can be interpreted as "when determined" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)".

[0075] Various structural schematic diagrams according to the disclosed embodiments of the present invention are shown in the accompanying drawings. These figures are not drawn to scale, where for the purpose of clear expression, some details are enlarged and some details may be omitted. The shapes of various regions and layers shown in the figures, as well as their relative sizes and positional relationships, are only exemplary. In practice, there may be deviations due to manufacturing tolerances or technical limitations, and those skilled in the art can design regions / layers with different shapes, sizes, and relative positions according to actual needs.

[0076] The present invention provides a non-stationary seismic data attenuation compensation method based on a spatio-temporal neural network. By comparing the performance of different networks in compensating the post-stack attenuated seismic data, a new spatio-temporal neural network, C3D_CLN, is constructed by combining 3DCNN and ConvLSTM networks. The spatio-temporal neural network structure, HST, is adopted, considering the spatio-temporal structural characteristics of seismic data. By training the spatio-temporal neural network, the attenuation compensation of non-stationary seismic data is realized, thereby improving the vertical resolution of deep seismic data.

[0077] A non-stationary seismic data attenuation compensation method based on a spatio-temporal neural network according to the present invention includes the following steps:

[0078] S1. Based on the collected non-stationary post-stack reflection seismic data d obs (t, n), estimate the shallow seismic wavelet by a statistical method, where t represents the time variable and n is the common depth point gather serial number; based on the estimated seismic wavelet and the actual well logging reflection coefficient sequence, synthesize the stationary seismic record trace at the well logging position through a convolution model, denoted as d cal (t, i), where t represents the time variable and i is the well logging serial number;

[0079] S2. Based on the non-stationary seismic record and the stationary synthetic seismic record trace at the well logging position, construct a data set that can consider the spatio-temporal characteristics of the seismic record for use in training and testing a deep neural network;

[0080] Since the stationary synthetic seismic record trace at the well logging position in the actual data only exists in the form of a single trace, it is very important to construct a two-and-a-half-dimensional label data set corresponding to the synthetic data. In order to make the constructed data better reflect the spatio-temporal characteristics of the original seismic data, the randomness of the label position and the construction data route is very important when constructing the data set. The construction method is as follows:

[0081] S201. Randomly generate N well points uniformly distributed within the space range of M*M;

[0082] S202. Randomly select several well points as the label points of a two-dimensional data according to certain rules (distance and angle). The distance and angle between two adjacent well points should meet certain conditions, that is, the distance between two adjacent well points should be greater than 50 and less than 100, and the included angle between three adjacent well points should be greater than 80°, so that the generated data is more reasonable.

[0083] S203. Randomly insert data points according to a certain probability to generate a single interpolation sequence. The so-called "according to a certain probability" means that the greater the distance between two points, the greater the probability of inserting a value in it, ensuring the randomness of the finally generated label positions and the continuity of the generated data.

[0084] S204. Generate three-dimensional seismic data with attenuation and corresponding non-attenuated semi-label data based on the data point positions obtained in steps S202 and S203 and the original synthetic seismic data.

[0085] In order to enable the neural network to better process seismic data, before inputting the training data into the neural network, the seismic data was preprocessed, including two aspects: data scaling and dimension expansion:

[0086] ① Since the values of the seismic data are small, they are uniformly magnified by 10 times when input, and when predicting, the prediction results are divided by 10.

[0087] ② In order to match the input requirements of the network and improve the performance of the network, the input three-dimensional seismic data with attenuation was subjected to dimension expansion processing, expanding one dimension from the time dimension to form a two-dimensional spatio-temporal signal of a single time point, that is, expanding the (N, T, S) data into (N, T, S, t, c) data, where N is the amount of training data, T is the time dimension, S is the space dimension, t is the expanded time dimension, and c is the number of channels set to 1.

[0088] S3. Design a suitable spatio-temporal neural network structure for non-stationary seismic data attenuation compensation, denoted as Γ θ , where θ represents the set of free parameters of the spatio-temporal neural network;

[0089] Since there are many existing spatio-temporal neural network design schemes, such as: 3DCNN, ConvLSTM network, etc., so the synthetic three-dimensional data volume 1 is used for comparison to select the most appropriate design scheme, where the size of the three-dimensional data volume 1 is (40, 1381, 30).

[0090] By Figure 1From the results, it can be seen that when the label position is fixed, in the test data, the ConvLSTM network performs very well in learning single-channel data with a fixed label position, while the learning effect of other channels is very poor. When the label position is random, the network performs well in most channels, and only performs poorly in a few channels. This indicates that the ConvLSTM network has a good learning effect, but has a strong dependence on the spatial position of the label and weak generalization performance.

[0091] From Figure 2 the results, it can be seen that whether the label position is fixed or random, in the test data, although the learning effect of the 3DCNN network is slightly worse than that of the ConvLSTM network, it is still acceptable. Moreover, the performance of the network when the label position is shuffled is not much different from that when the label position is not shuffled. All these can reflect the good generalization ability of the 3DCNN network.

[0092] In order to combine the respective advantages of the above two networks, improve the performance of the network and reduce the dependence of the network on the label position, after multiple experiments, the two networks are selected to be connected in a horizontal Concat connection method, that is, C3D_CLN. This network has a faster learning speed, better learning effect, and gets rid of the constraint of the label position on the network. As Figure 3 shown, the experimental results when the label is fixed on the three-dimensional data volume 1 show that even when the label position is fixed, the learning effect of the C3D_CLN network is better than that of the 3DCNN and ConvLSTM networks at most non-label positions.

[0093] The working principle of the C3D_CLN network: The C3D_CLN network is composed by integrating the respective advantages of the ConvLSTM network and the 3DCNN network. The ConvLSTM network is a spatio-temporal network proposed by Shi et al. based on LSTM. Among them, the connection between the input and each gate is replaced by a convolution of the fully connected layer of LSTM, and the connection between the states is also replaced by a convolution operation. The structure of the ConvLSTM is as Figure 7 shown, and the working principle of the ConvLSTM network is represented by Equation 2.

[0094]

[0095] Among them, χ t is the input value at time t, C t is the state value at time t, H t is the hidden state value at time t, i t , f t , o t are intermediate values, W * and b * represent the corresponding convolution kernels and biases, and '*' represents the convolution symbol. represents the Hadamard product, and 'σ' and 'tanh' represent non - linear functions.

[0096] The 3DCNN network was proposed by Ji et al. based on the 2DCNN network. The size of the convolutional kernel was replaced from two - dimensional to three - dimensional, and the working principle of the 3DCNN network is represented by Equation 3.

[0097]

[0098] Where, represents the value of the point (x, y, z) of the j - th feature map in the i - th layer, 'tanh' represents the non - linear function, represents the value of the convolutional kernel of the m - th layer at the point (p, q, r), and b ij represents the bias.

[0099] Network parameter design: Adam is used as the network optimizer, the initial learning rate is 5e - 4, the learning decay rate is 0.98, Epoch is 60, the Loss function uses the mean - square error loss with a self - constructed MASK matrix, the Batch_Size is 16, the size of the convolutional kernel in the 3DCNN layer is (5, 3, 3), Filter is 8, the size of the convolutional kernel in the ConvLSTM layer is (3, 3), filter is 8, and except for the last non - linear layer which is Linear, other non - linear layers are all Relu. The network structure is as Figure 8 shown.

[0100] In order to further improve the spatio - temporal compensation performance of the network, Inception - v3 and the two - stream network structure are incorporated into the C3D_CLN network to obtain the HST network, as Figure 9 shown. The TF1 data stream is the input post - stack data with attenuation, and the TF2 data stream is the difference between the post - stack data with attenuation at the previous and next time points. The reason for this design is that TF2 contains the change information of the time data, which can greatly improve the performance of the network. The learning results are as Figure 6 shown.

[0101] S4. Based on the dataset constructed in step S2 and the spatio - temporal neural network structure established in step S3, perform network training to determine the optimal parameter set θ * ;

[0102] S5. Based on the trained spatio - temporal neural network structure map the non - stationary seismic traces to stationary seismic traces to obtain the two - dimensional stationary seismic data estimated by the network

[0103]

[0104] Where, Represents two-dimensional non-stationary seismic data composed of multiple observed non-stationary seismic traces.

[0105] In order to reflect the generalization and compensation effects of the C3D_CLN network and the HST network, the synthetic 3D data volume 2 is used to test the network. The size of the 3D data volume 2 is (401, 150, 401), and under the same other conditions, a comparative experiment is carried out with other networks. The final result shows that the designed network has the best performance. Other networks include but are not limited to: 2D_CNN_Res, 3DCNN, LSTM, and ConvLSTM.

[0106] In another embodiment of the present invention, a non-stationary seismic data attenuation compensation system based on a spatio-temporal neural network is provided. This system can be used to implement the above-mentioned non-stationary seismic data attenuation compensation method based on a spatio-temporal neural network. Specifically, the non-stationary seismic data attenuation compensation system based on a spatio-temporal neural network includes an acquisition module, a data module, a design module, a training module, and a compensation module.

[0107] Among them, the acquisition module acquires non-stationary post-stack reflected seismic data and estimates the shallow seismic wavelet through a statistical method; synthesizes the stationary seismic record traces at the well logging positions by convolving the estimated shallow seismic wavelet and the actual well logging reflection coefficient sequence.

[0108] The data module constructs a data set considering the spatio-temporal characteristics of the seismic records based on the non-stationary seismic records and the stationary synthetic seismic record traces at the well logging positions obtained by the acquisition module.

[0109] The design module designs a spatio-temporal neural network structure for non-stationary seismic data attenuation compensation.

[0110] The training module performs network training on the spatio-temporal neural network structure established by the design module based on the data set constructed by the data module to determine the optimal parameter set of the spatio-temporal neural network.

[0111] The compensation module maps the non-stationary seismic traces to stationary seismic traces based on the trained spatio-temporal neural network structure and the optimal parameter set obtained by the training module, and obtains the two-dimensional stationary seismic data estimated by the network.

[0112] In another embodiment of the present invention, a terminal device is provided. The terminal device includes a processor and a memory. The memory is used to store a computer program, and the computer program includes program instructions. The processor is used to execute the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions to implement the corresponding method flow or corresponding function. The processor described in the embodiment of the present invention can be used for the operation of the non-stationary seismic data attenuation compensation method based on the spatio-temporal neural network, including:

[0113] Collect non-stationary post-stack reflected seismic data, and estimate the shallow seismic wavelet by statistical methods; Synthesize the stationary seismic record trace at the well logging position by convolving the estimated shallow seismic wavelet and the actual well logging reflection coefficient sequence; Based on the non-stationary seismic record and the obtained stationary synthetic seismic record trace at the well logging position, construct a data set considering the spatio-temporal characteristics of the seismic record; Design a spatio-temporal neural network structure for non-stationary seismic data attenuation compensation; Based on the constructed data set, train the established spatio-temporal neural network structure to determine the optimal parameter set of the spatio-temporal neural network; Based on the trained spatio-temporal neural network structure and the optimal parameter set, map the non-stationary seismic trace to a stationary seismic trace to obtain the two-dimensional stationary seismic data estimated by the network.

[0114] In another embodiment of the present invention, the present invention further provides a storage medium, specifically a computer-readable storage medium (Memory). The computer-readable storage medium is a memory device in a terminal device and is used to store programs and data. It can be understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and, of course, the extended storage medium supported by the terminal device. The computer-readable storage medium provides a storage space, and this storage space stores the operating system of the terminal. Moreover, in this storage space, one or more instructions suitable for being loaded and executed by the processor are also stored. These instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory (Non-Volatile Memory), such as at least one disk memory.

[0115] One or more instructions stored in the computer-readable storage medium can be loaded and executed by the processor to implement the corresponding steps of the method for non-stationary seismic data attenuation compensation based on the spatio-temporal neural network in the above embodiments; one or more instructions in the computer-readable storage medium are loaded and executed by the processor to perform the following steps:

[0116] Collect non-stationary post-stack reflected seismic data and estimate the shallow seismic wavelet through statistical methods; synthesize the estimated shallow seismic wavelet and the actual well logging reflection coefficient sequence through a convolution model to obtain a stationary seismic record trace at the well logging position; based on the non-stationary seismic record and the obtained stationary synthetic seismic record trace at the well logging position, construct a data set considering the spatio-temporal characteristics of the seismic record; design a spatio-temporal neural network structure for non-stationary seismic data attenuation compensation; perform network training on the established spatio-temporal neural network structure based on the constructed data set to determine the optimal parameter set of the spatio-temporal neural network; based on the trained spatio-temporal neural network structure and the optimal parameter set, map the non-stationary seismic trace to a stationary seismic trace to obtain the two-dimensional stationary seismic data estimated by the network.

[0117] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Generally, the components described and shown in the accompanying drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed present invention, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0118] Numerical simulation results

[0119] Synthetic seismogram data

[0120] The training data required by the network is randomly extracted from the 3D data volume 2, where the size of the 3D data volume 2 is (401, 150, 401), the number of label traces is 100, and the size of the extracted training data is (200, 150, 100), and each profile has 4 label traces.

[0121] Figure 4 Training experiment results of the C3D_CLN network on the 3D data volume 2. (a) The decline curves of the training loss (dark gray) and the validation loss (light gray) when the number of network training times is 60; (b) The single-trace seismic data compensation results of the C3D_CLN network test set, single-trace seismic data without attenuation (dark gray), single-trace seismic data with attenuation (light gray), and single-trace seismic data after compensation by the C3D_CLN network for single-trace seismic data with attenuation (medium gray).

[0122] In Figure 4 (a), the validation loss can decline and converge normally, indicating that the C3D_CLN network can learn the spatio-temporal characteristics of seismic data with attenuation normally; in Figure 4 (b), the single-trace seismic data without attenuation and the single-trace seismic data after compensation by the C3D_CLN network for single-trace seismic data with attenuation are almost completely fitted, and it can be seen that the C3D_CLN network has a very superior compensation effect on seismic data with attenuation.

[0123] Figure 5 Experimental results of the C3D_CLN and LSTM networks on the test set of the 3D data volume 2 under the same conditions. (a) The learning results of a certain 2D data of the C3D_CLN network; (b) The learning results of a certain 2D data of the LSTM network. (c) The difference between the learning results of the two networks and the seismic data without attenuation, the C3D_CLN network (left figure), the LSTM network (right figure).

[0124] In Figure 5 (a), the 2D seismic data after compensation by the C3D_CLN network for 2D seismic data with attenuation can almost perfectly reconstruct the seismic data without attenuation; in Figure 5 (b), the 2D seismic data after single-trace compensation by the LSTM network for single-trace seismic data with attenuation can also better compensate the attenuation of the seismic data; in order to better distinguish the differences between the two methods, the learning results of the two networks for the same 2D seismic data with attenuation and the 2D seismic data without attenuation are subtracted, and from Figure 5 (c), it can be clearly seen that the learning effect of the C3D_CLN network is better, making the spatio-temporal continuity of the learned seismic signal stronger.

[0125] Table 1 shows the comparison results of the peak signal-to-noise ratio between the C3D_CLN network and other neural networks:

[0126] Table 1

[0127] Network None 2DCNN 3DCNN LSTM ConvLSTM C3D_CLN PSNR / dB 25.83 36.29 36.97 38.89 41.90 42.64

[0128] Among them, "none" refers to the peak signal-to-noise ratio of non-attenuated seismic data and attenuated seismic data. After the attenuated seismic data is learned by the LSTM network, the peak signal-to-noise ratio is increased by 13.06 dB. After the attenuated seismic data is learned by the C3D_CLN network, the peak signal-to-noise ratio is increased by 16.80 dB. The C3D_CLN network is 3.74 dB higher than the LSTM network, and the improvement rate is about 28.6%. In addition, the C3D_CLN network also considers spatio-temporal correlation and increases the spatial continuity of the compensated seismic data.

[0129] In order to compare the performance improvement degrees of the HST network and other networks, under the conditions that the Epoch is set to 200 and the initial learning rate is 2e-3, the learning results are as follows Figure 6 As shown, the results indicate that the HST network has better learning and compensation capabilities for non-stationary seismic data.

[0130] Table 2 shows the comparison results of the peak signal-to-noise ratio between the HST network and other neural networks:

[0131] Table 2

[0132] Network None LSTM C3D_CLN HST PSNR / dB 25.83 40.82 44.27 46.01

[0133] Under the same experimental conditions, after the attenuated seismic data is learned by the LSTM network, the peak signal-to-noise ratio is increased by 14.99 dB. After the attenuated seismic data is learned by the C3D_CLN network, the peak signal-to-noise ratio is increased by 18.44 dB. After the attenuated seismic data is learned by the HST network, the peak signal-to-noise ratio is increased by 20.17 dB. The C3D_CLN network is 3.45 dB higher than the LSTM network, and the HST network is 5.18 dB higher than the LSTM network, and the improvement rate is about 34.6%.

[0134] In summary, for the non-stationary seismic data attenuation compensation method and system based on spatio-temporal neural network of the present invention, by comparing the performance of different networks in the compensation effect of post-stack seismic data with attenuation, a new spatio-temporal neural network, C3D_CLN, is constructed by combining 3DCNN and ConvLSTM. On this basis, another more reasonable spatio-temporal neural network, HST, is further proposed. From the experiments on synthetic seismic data, it can be clearly seen that this method takes into account the spatio-temporal structural characteristics of seismic data, and realizes the attenuation compensation of non-stationary seismic data by training the spatio-temporal neural network, thereby improving the vertical resolution of deep seismic data.

[0135] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the above-mentioned division of each functional unit and module is used as an example. In actual applications, the above-mentioned functions can be allocated to different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment can be integrated into a processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a software functional unit. In addition, the specific names of each functional unit and module are only for the convenience of mutual distinction and do not limit the protection scope of the present application. The specific working processes of the units and modules in the above system can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0136] In the above embodiments, the descriptions of the respective embodiments have their own emphases. For the parts not detailed or recorded in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0137] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or by a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0138] In the embodiments provided by the present invention, it should be understood that the disclosed device / terminal and method can be implemented in other ways. For example, the device / terminal embodiments described above are merely illustrative. For example, the division of the modules or units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling or direct coupling or communication connection between each other can be through some interfaces. The indirect coupling or communication connection of the device or unit can be in electrical, mechanical or other forms.

[0139] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0140] In addition, each functional unit in various embodiments of the present invention can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above integrated units can be implemented in the form of hardware or in the form of software functional units.

[0141] If the integrated module / unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, to implement all or part of the processes in the above method embodiments of the present invention, it can also be completed by a computer program instructing relevant hardware. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the content included in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0142] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or means for implementing the functions specified in one block or multiple blocks.

[0143] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means, and the instruction means implement the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or means for implementing the functions specified in one block or multiple blocks.

[0144] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or means for implementing the functions specified in one block or multiple blocks.

[0145] The above content is only for explaining the technical idea of the present invention and cannot be used to limit the protection scope of the present invention. Any modification made on the basis of the technical solution according to the technical idea proposed by the present invention falls within the protection scope of the claims of the present invention.

Claims

1. A non-stationary seismic data attenuation compensation method based on a spatio-temporal neural network, characterized in that, Including the following steps: S1. Collect non-stationary post-stack reflection seismic data, and estimate the shallow seismic wavelet by statistical methods; Synthesize the stationary seismic record traces at the logging positions by convolving the estimated shallow seismic wavelet with the actual logging reflection coefficient sequence; S2. Based on the non-stationary seismic records and the stationary synthetic seismic record traces at the logging positions obtained in step S1, construct a data set considering the spatio-temporal characteristics of the seismic records; S3. Design a spatio-temporal neural network structure for non-stationary seismic data attenuation compensation; S4. Based on the data set constructed in step S2, train the spatio-temporal neural network structure established in step S3 to determine the optimal parameter set of the spatio-temporal neural network; S5. Based on the trained spatio-temporal neural network structure and the optimal parameter set in step S4, map the non-stationary seismic traces to stationary seismic traces to obtain the two-dimensional stationary seismic data estimated by the network.

2. The non-stationary seismic data attenuation compensation method based on a spatio-temporal neural network according to claim 1, characterized in that, In step S3, the specific construction of the data set considering the spatio-temporal characteristics of the seismic records is as follows: S301. Randomly generate N well points uniformly distributed within the spatial range of M*M; S302. Randomly select several well points among them as the label points of a two-dimensional data according to distance and angle; S303. Randomly insert data points to generate a single interpolation sequence; S304. According to the data point positions and the original synthetic seismic data obtained in steps S302 and S303, generate three-dimensional seismic data with attenuation and corresponding non-attenuated semi-label data.

3. The non-stationary seismic data attenuation compensation method based on a spatio-temporal neural network according to claim 2, wherein Before constructing the data set considering the spatio-temporal characteristics of the seismic records, preprocess the seismic data as follows: When inputting, uniformly amplify the seismic data values by 10 times, and when predicting, divide the prediction result by 10; Perform dimension expansion processing on the input three-dimensional seismic data with attenuation, expand one dimension from the time dimension to form a two-dimensional spatio-temporal signal of a single time point.

4. The non-stationary seismic data attenuation compensation method based on a spatio-temporal neural network according to claim 1, wherein, In step S4, connect the 3DCNN network and the ConvLSTM network in a horizontal Concat connection manner to obtain the C3D_CLN spatio-temporal neural network structure.

5. The non-stationary seismic data attenuation compensation method based on a spatio-temporal neural network according to claim 4, wherein The working principle of the ConvLSTM network is as follows: Among them, χ t is the input value at time t, C t is the state value at time t, H t is the hidden state value at time t, i t , f t , o t are intermediate values, W * and b * represent the corresponding convolution kernel and bias, ‘*’ represents the convolution operator, represents the Hadamard product, and ‘σ’ and ‘tanh’ represent non - linear functions.

6. The non-stationary seismic data attenuation compensation method based on a spatio-temporal neural network according to claim 4, characterized in that The working principle of the 3DCNN network is as follows: Among them, represents the value of the point (x, y, z) of the j-th feature map in the i-th layer, and 'tanh' represents the non-linear function. represents the value of the convolutional kernel of the m-th layer at the point (p, q, r), and b ij represents the bias.

7. The non-stationary seismic data attenuation compensation method based on a spatio-temporal neural network according to claim 4, characterized in that The C3D_CLN spatio-temporal neural network structure uses Adam as the network optimizer, with an initial learning rate of 5e -4 , a learning decay rate of 0.98, an Epoch of 60, a self-constructed mean square error loss with a MASK matrix is used for the Loss function, a Batch_Size of 16, the convolutional kernel size of the 3DCNN layer is (5, 3, 3), the Filter is 8, the convolutional kernel size of the ConvLSTM layer is (3, 3), the filter is 8, and except for the last non-linear layer which is Linear, other non-linear layers are all Relu.

8. The non-stationary seismic data attenuation compensation method based on a spatio-temporal neural network according to claim 4, wherein, Integrate the Inception-v3 and two-stream network structures into the C3D_CLN spatio-temporal neural network structure to obtain the HST network. The TF1 data stream in the HST network is the input post-stack data with attenuation, and the TF2 data stream is the difference between the post-stack data with attenuation at the previous and next time points.

9. The non-stationary seismic data attenuation compensation method based on a spatio-temporal neural network according to claim 1, characterized in that Two-dimensional stationary seismic data estimated by network is as follows: Among them, is a trained spatio-temporal neural network structure, represents two-dimensional non-stationary seismic data composed of multiple observed non-stationary seismic traces.

10. A non-stationary seismic data attenuation compensation system based on a spatio-temporal neural network, characterized in that, Including: A collection module that collects non-stationary post-stack reflection seismic data and estimates the shallow seismic wavelet by statistical methods; Synthesize the stationary seismic record traces at the logging positions by convolving the estimated shallow seismic wavelet with the actual logging reflection coefficient sequence; A data module that constructs a data set considering the spatio-temporal characteristics of the seismic records based on the non-stationary seismic records and the stationary synthetic seismic record traces at the logging positions obtained by the collection module; A design module that designs a spatio-temporal neural network structure for non-stationary seismic data attenuation compensation; A training module that trains the spatio-temporal neural network structure established by the design module based on the data set constructed by the data module to determine the optimal parameter set of the spatio-temporal neural network; The compensation module, based on the trained spatio-temporal neural network structure and the optimal parameter set of the training module, maps non-stationary seismic traces into stationary seismic traces to obtain the two-dimensional stationary seismic data estimated by the network.

Images