A seismic wave attenuation estimation method and estimation system based on deep learning
By introducing deep learning neural network model in sparse time-frequency transformation, the problems of long calculation time and inefficient parameter selection of traditional methods are solved, and efficient seismic attenuation estimation is achieved.
Patent Information
- Application Number
- CN202211208146.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-09-30
- Publication Date
- 2025-05-16
- Estimated Expiration
- 2042-09-30
AI Technical Summary
The traditional sparse time-frequency transformation method requires a lot of calculation time in earthquake attenuation estimation, and manual parameters are inefficient in experiments, which affects resolution.
A deep learning neural network model was introduced, a sparse short-time Fourier transform model was constructed based on the Unet network, and an unsupervised sparse time-frequency domain adaptive transformation model was constructed through transfer learning methods, reducing manual intervention and calculation time.
The efficiency of time-frequency analysis is improved, the calculation time and parameter selection problems of sparse time-frequency transformation are overcome, and the practicality and reliability of earthquake attenuation estimation are enhanced.
Smart Images
Figure CN115932953B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of geophysical exploration, and in particular relates to a seismic attenuation estimation method and estimation system based on deep learning. Background Art
[0002] When seismic waves propagate in the strata, there will be energy attenuation, which is affected and restricted by many factors. These factors include: frequency, pressure, temperature, saturation, strain amplitude, and rock characteristics. In the process of studying the absorption and attenuation characteristics of strata, it is very helpful to understand the influence of these factors for the study of attenuation problems.
[0003] Time-frequency analysis can obtain the localized characteristics of signals and has been widely used in describing seismic signal processing and interpretation. Traditional time-frequency transforms can be simply divided into linear time-frequency transforms and nonlinear time-frequency transforms. Linear time-frequency transforms include continuous wavelet transforms, S transforms, short-time Fourier transforms, etc. Since this type of time-frequency transform method satisfies the uncertainty principle, its time-frequency resolution is limited, resulting in its inability to accurately characterize the localized properties of seismic signals. Nonlinear time-frequency transforms include Wigner-Ville distribution, etc. Although this type of time-frequency transform method has a high time-frequency resolution, it is interfered by cross terms, resulting in its inability to accurately analyze and describe time-frequency characteristics.
[0004] In order to improve the resolution of time-frequency transform, scholars have proposed many improved time-frequency transform methods, including synchronous squeezing transform, synchronous extraction transform, sparse time-frequency transform, etc. Among them, sparse time-frequency transform expresses the solution of time-frequency coefficients as an inverse problem, which is solved by optimizing the iterative algorithm. This method can improve the resolution of the time-frequency transform method. At present, sparse time-frequency transform has been widely used in scenarios such as earthquake attenuation estimation and earthquake complex orbit analysis, and has high practical significance. However, the traditional sparse time-frequency spectrum analysis method has the following problems: First, the sparse time-frequency spectrum analysis method requires a lot of computing time; second, the selection of inappropriate regularization parameters will affect the resolution of the sparse time-frequency spectrum analysis method. Summary of the invention
[0005] The present invention aims to solve at least one of the above technical problems to a certain extent or at least provide a useful commercial option. To this end, one object of the present invention is to propose a seismic wave attenuation estimation method based on deep learning. The seismic wave attenuation estimation method introduces a deep learning neural network model based on the calculation of sparse time-frequency transform to avoid multiple inefficient manual parameter selection experiments and a large amount of calculation time, overcome the shortcomings of sparse time-frequency transform in the prior art, and further apply it to actual seismic data to perform seismic attenuation estimation, thereby enhancing the practicality and reliability of this method. Another object of the present invention is to provide a seismic wave attenuation estimation system based on deep learning.
[0006] The seismic wave attenuation estimation method based on deep learning according to the present invention comprises the following steps:
[0007] Use deep learning algorithms to build a sparse short-time Fourier transform model based on the Unet network;
[0008] Calculating the sparse time-frequency spectrum of synthetic seismic data as labels based on the sparse short-time Fourier transform model to form a synthetic data set;
[0009] Using the synthetic data set to train the sparse short-time Fourier transform model to obtain a trained sparse short-time Fourier transform model, and verifying the training effect of the sparse short-time Fourier transform model;
[0010] Using a transfer learning method, constructing an unsupervised sparse time-frequency domain adaptive transformation model based on the trained sparse short-time Fourier transform model;
[0011] Using unlabeled actual earthquake data to train the unsupervised sparse time-frequency domain adaptive transformation model to obtain a trained unsupervised sparse time-frequency domain adaptive transformation model, and verifying the second prediction effect;
[0012] The three-dimensional post-stack seismic data are applied to the trained unsupervised sparse time-frequency domain adaptive transform model to estimate the seismic attenuation amplitude.
[0013] The deep learning-based seismic wave attenuation estimation method of the present invention introduces a deep learning neural network model on the basis of sparse time-frequency transform calculation to avoid multiple inefficient manual parameter selection experiments and a large amount of calculation time, overcomes the shortcomings of sparse time-frequency transform in the prior art, and further applies it to actual seismic data to perform seismic attenuation estimation, thereby enhancing the practicability and reliability of this method.
[0014] In addition, the seismic wave attenuation estimation method based on deep learning according to the present invention may also have the following technical features:
[0015] The sparse short-time Fourier transform model includes a mapping structure, a feature extraction structure and an upsampling structure. The mapping structure includes a 1x3 convolution operation for converting one-dimensional seismic data into a two-dimensional feature map. The feature extraction structure includes a repeated 3x3 convolution and a 2x2 maximum pooling operation unit. The upsampling structure includes a repeated 3x3 convolution operation unit, a 2x2 convolution operation unit and a 1x1 convolution operation unit.
[0016] Before using the sparse time-frequency spectrum of the synthetic seismic data as a label, the following steps are also included:
[0017] The synthetic seismic data is normalized.
[0018] The normalization processing performed on the synthetic seismic data includes at least one of maximum normalization and mean variance normalization.
[0019] The method of training the sparse short-time Fourier transform model by using the synthetic data set to obtain a trained sparse short-time Fourier transform model specifically comprises the following steps:
[0020] The Adam optimization algorithm is used as the optimizer of the short-time Fourier transform model, and the mean square error loss function is selected to measure the difference between the learning results of the sparse short-time Fourier transform model and its corresponding labels.
[0021] The method of using the transfer learning method to construct an unsupervised sparse time-frequency domain adaptive transformation model based on the trained sparse short-time Fourier transform model specifically includes the following steps:
[0022] The unsupervised sparse time-frequency domain adaptive transformation model includes a source domain and a target domain, and the similarity between the source domain and the target domain is calculated by cosine distance.
[0023] The method of applying the three-dimensional post-stack seismic data to the trained unsupervised sparse time-frequency domain adaptive transformation model to estimate the seismic attenuation amplitude specifically comprises the following steps:
[0024] Using high-frequency three-dimensional post-stack seismic data and low-frequency three-dimensional post-stack seismic data to transform the trained unsupervised sparse time-frequency domain adaptive transformation model to obtain high-frequency seismic amplitude and low-frequency seismic amplitude respectively;
[0025] The difference between the high-frequency earthquake amplitude and the low-frequency earthquake amplitude is calculated, and the intensity of earthquake attenuation is estimated according to the difference.
[0026] The present invention also provides a seismic wave attenuation estimation system based on deep learning, and the seismic wave attenuation estimation system is implemented by any of the above-mentioned seismic wave attenuation estimation methods.
[0027] The present invention has the following beneficial effects:
[0028] The present invention introduces a deep learning network model to calculate the sparse time-frequency spectrum of seismic data, avoiding a large number of manual parameter selection experiments in the traditional calculation process and improving the efficiency of time-frequency analysis. At the same time, it uses the relevant knowledge of transfer learning to break through the problem of the difficulty in obtaining actual seismic data labels, and successfully applies it to earthquake attenuation estimation, which has high practical significance.
[0029] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:
[0031] Figure 1 A schematic diagram of the network structure of a sparse time-frequency transform Unet (STFNet) model according to an embodiment of the present invention;
[0032] Figure 2 The figure is a schematic diagram of the effect of the STFNet model on the test set according to an embodiment of the present invention, wherein: Figure 2 (a) to (b) represent two schematic diagrams of synthetic data randomly selected from the test set. Figure 2 (c)~(d) represent the schematic diagrams of sparse time-frequency spectrum predicted by the STFNet network;
[0033] Figure 3 A schematic diagram of an unsupervised sparse time-frequency domain adaptive transform (USTFDA) model according to an embodiment of the present invention;
[0034] Figure 4 FIG. 1 is a schematic diagram of an actual seismic data training set according to an embodiment of the present invention, wherein: Figure 4 (a) to (b) represent two actual seismic orbit diagrams randomly extracted from the actual data set. Figure 4 (c)~(d) represent the schematic diagrams of the sparse time-frequency spectrum predicted by the USTFDA network;
[0035] Figure 5 A schematic diagram of frequency slices calculated by the USTFDA model according to an embodiment of the present invention, wherein Figure 5 (a) and (b) are schematic diagrams of frequency slices of 20 Hz and 45 Hz predicted by the USTFDA model;
[0036] Figure 6 It is a schematic diagram of the attenuation results calculated at 20 Hz and 45 Hz obtained by using the USTFDA model according to an embodiment of the present invention. Specific implementation methods
[0037] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.
[0038] Figure 1 A schematic diagram of the network structure of a sparse time-frequency transform Unet (STFNet) model according to an embodiment of the present invention; Figure 2 The figure is a schematic diagram of the effect of the STFNet model on the test set according to an embodiment of the present invention, wherein: Figure 2 (a) to (b) represent two schematic diagrams of synthetic data randomly selected from the test set. Figure 2 (c)~(d) represent the schematic diagrams of sparse time-frequency spectrum predicted by the STFNet network; Figure 3 A schematic diagram of an unsupervised sparse time-frequency domain adaptive transform (USTFDA) model according to an embodiment of the present invention; Figure 4 FIG. 1 is a schematic diagram of an actual seismic data training set according to an embodiment of the present invention, wherein: Figure 4 (a) to (b) represent two actual seismic orbit diagrams randomly extracted from the actual data set. Figure 4 (c)~(d) represent the schematic diagrams of the sparse time-frequency spectrum predicted by the USTFDA network; Figure 5 A schematic diagram of frequency slices calculated by the USTFDA model according to an embodiment of the present invention, wherein Figure 5 (a) and (b) are schematic diagrams of frequency slices of 20 Hz and 45 Hz predicted by the USTFDA model; Figure 6 This is a schematic diagram of the attenuation results calculated at 20 Hz and 45 Hz using the USTFDA model according to an embodiment of the present invention. Figure 1-Figure 6 The present invention provides a seismic wave attenuation estimation method based on deep learning, and the seismic wave attenuation estimation method comprises the following steps:
[0039] S1. Use deep learning algorithm to build a sparse short-time Fourier transform model based on Unet network.
[0040] In the specific implementation, the seismic wave attenuation estimation method of the present invention is modified on the basis of the classic Unet model, and a sparse short-time Fourier transform model (Sparse Short Time Fourier Transform) is proposed in combination with the scenario of sparse time-frequency analysis. The specific network model structure is as follows: Figure 1 shown.
[0041] The sparse short-time Fourier transform model includes three parts, namely, a mapping structure, a feature extraction structure, and an upsampling structure. Among them, the mapping structure includes a 1x3 convolution operation unit, which can convert one-dimensional seismic data into a corresponding two-dimensional feature map, which is convenient for subsequent model training. The feature extraction structure consists of repeated 3x3 convolution and 2x2 maximum pooling (max-pooling) operation units. After each convolution, the number of channels of the network will double, so as to better perform feature learning. The upsampling structure is similar to the feature extraction structure, and is also composed of repeated 3x3 convolution operation units, but here 2x2 convolution is used to replace the 2x2 maximum pooling operation unit of the feature extraction structure, and the number of channels is reduced to half of the original after each convolution. The upsampling structure also includes a 1x1 convolution operation unit, which is used to convert the final feature map into the corresponding sparse time-frequency result.
[0042] S2. Calculate the sparse time-frequency spectrum of the synthetic seismic data as labels based on the sparse short-time Fourier transform model to form a synthetic data set.
[0043] Specifically, the seismic wave attenuation estimation method of the present invention generates 14010 seismic synthetic data using the reflection data model and Ricker wavelets of different frequencies, performs data normalization, and then calculates the corresponding sparse time-frequency spectrum as a training label. In a specific implementation, at least one of the maximum value normalization or the mean variance normalization can be used to normalize the data; in this embodiment, the mean variance normalization is used to normalize the data, so that all data are normalized to a distribution with a mean of 0 and a variance of 1.
[0044] Sparse time-frequency transform is proposed on the basis of short-time Fourier transform. For the seismic signal u(t), short-time Fourier transform can be defined as follows:
[0045]
[0046] In formula (1), p and q are discrete time samples and discrete frequency samples respectively. On this basis, the inverse problem of short-time Fourier transform can be expressed as:
[0047]
[0048] It can be expressed in matrix form as:
[0049] u=Gm, (3)
[0050] in, is a vector of u(t), is the dictionary generated by g(n). is generated by the time-frequency coefficients. When the dictionary G and the signal u(t) exist, m can be obtained according to formula (3). Therefore, the inverse problem can be described as:
[0051]
[0052] in, σ、 Represent the noise level and regularization term respectively.
[0053] In order to solve the inverse problem defined in formula (4), we adopt an iterative approach to divide the inverse problem into two optimization sub-problems:
[0054]
[0055] Here, k represents the number of iterations.
[0056] Through the above steps, we calculated the sparse time-frequency spectra of 14,010 synthetic data to form a synthetic dataset, and used 80% of the synthetic dataset as a training set and 20% as a validation set.
[0057] S3. Use the training set to train the sparse short-time Fourier transform model to obtain a trained sparse short-time Fourier transform model, and use the verification set to verify the training effect of the sparse short-time Fourier transform model.
[0058] Specifically, in this training, we took into account the low memory requirements and high computational efficiency of the Adam (Adaptive momentum) optimization algorithm and used it as the optimizer for the entire network. The Adam optimization algorithm is a stochastic optimization method for adaptive momentum and is often used as an optimizer algorithm in deep learning. Furthermore, the mean square error loss function is selected to measure the difference between the learning results and the corresponding labels, and back-propagation is performed to update the network parameters.
[0059] In order to verify the accuracy of the STFNet model, we used an additional 100 synthetic data to form a test set, such as Figure 2 shown. Figure 2 (a) to (b) are two synthetic data randomly selected from the test set. Figure 2 (c) to (d) are the sparse time-frequency results predicted by the STFNet network. It can be seen that our STFNet model has learned the nonlinear relationship between the seismic signal and the corresponding sparse time-frequency spectrum.
[0060] S4. Using a transfer learning method, an unsupervised sparse time-frequency domain adaptive transformation model is constructed based on the trained sparse short-time Fourier transform model.
[0061] Specifically, after obtaining the trained STFNet model, we introduce the knowledge of domain adaptation and construct an unsupervised sparse time-frequency domain adaptation (USTFDA) model, where the unsupervised sparse time-frequency domain adaptation model includes a source domain and a target domain. The main idea of domain adaptation is to minimize the difference in data distribution characteristics between the source domain and the target domain. Therefore, we use the Cosine distance to measure the similarity of the feature distribution of the two domains. Here, the source domain represents synthetic data, and the target domain represents actual seismic data.
[0062] Figure 3 The network structure and calculation process of the USTFDA model are shown. First, we obtain the parameters of the STFNet model trained with synthetic data and freeze the parameters of the mapping structure and feature extraction structure to achieve better migration effect. Secondly, for the remaining structure of the network, that is, the upsampling structure, we divide it into two processes. Figure 3 The downstream process shown is the source domain dataset process, where we upsample the input source domain data and calculate the mean square error loss with the corresponding labels, and the result constitutes part of the entire network loss function. Figure 3 The upstream process in is the target domain data set process. In this process, for the target domain data and the source domain data after each layer of upsampling, the corresponding Cosine distance is calculated, that is, the similarity between the two. The calculation result constitutes another part of the entire network loss function. The overall network loss function is as follows:
[0063]
[0064] In formula (6), loss MSE (R sour , L sour ) means calculating the mean square error loss between the source domain data training results and the labels. cos (d sour i , d tar i ) is the reciprocal of the Cosine distance between the source domain data and the target domain data after upsampling; because the larger the Cosine distance, the more similar the distances between the two data are, so we need to minimize the reciprocal of the Cosine distance in the loss function to maximize the Cosine distance. γ represents a weight parameter used to adjust the proportion of the Cosine distance in the overall loss function.
[0065] S5. Use unlabeled actual seismic data to train the unsupervised sparse time-frequency domain adaptive transformation model to obtain a trained unsupervised sparse time-frequency domain adaptive transformation model, and verify the training effect of the unsupervised sparse time-frequency domain adaptive transformation model.
[0066] Specifically, in order to train the USTFDA model, we randomly selected 1500 real data to form the actual data training set. In addition, we randomly selected 1500 synthetic data and corresponding labels as the synthetic data training set. In general, we use 1500 unlabeled real data and 1500 labeled synthetic data to train the USTFDA model.
[0067] During the training process, we set the learning rate dynamically so that it gradually decreases with the number of iterations, thus avoiding falling into the trap of local minima.
[0068] After training the USTFDA model, we selected 100 actual data from the actual dataset to form a test set to verify the effectiveness of USTFDA. Figure 4 (a) to (b) show two randomly selected test data and the corresponding prediction results. Figure 4 (c)~(d) are the sparse time-frequency spectra predicted by the USTFDA model.
[0069] S6. Applying the three-dimensional post-stack seismic data to the trained unsupervised sparse time-frequency domain adaptive transformation model to estimate the seismic attenuation amplitude.
[0070] Specifically, we used the trained USTFDA model to predict the overall 3D data and extracted the results with the main frequencies of 20 Hz and 45 Hz. Figure 5 (a) to (b) show the 20Hz and 45Hz frequency slices calculated using the USTFDA model. Figure 5 The low-frequency and high-frequency slices shown are used to calculate seismic attenuation. Figure 6 As shown in the figure, the black dots represent oil layers and the white dots represent dry layers. Figure 6 The displayed attenuation results are matched. The higher the attenuation value, the greater the difference between high and low frequencies, indicating that the high-frequency attenuation has more energy, which means that there is an oil reservoir. In the attenuation results predicted by the USTFDA model, all wells were correctly inferred, which shows that the USTFDA model has stability and accuracy, and has high practical value for seismic attenuation estimation.
[0071] The present invention also provides a seismic wave attenuation estimation system based on deep learning, and the seismic wave attenuation estimation system is implemented by the above-mentioned seismic wave attenuation estimation method.
[0072] The present invention discloses a seismic wave attenuation estimation system based on deep learning. The seismic wave attenuation estimation system introduces a deep learning neural network model on the basis of the calculation of sparse time-frequency transform to avoid multiple inefficient manual parameter selection experiments and a large amount of calculation time, overcomes the shortcomings of sparse time-frequency transform in the prior art, and further applies it to actual seismic data to perform seismic attenuation estimation, thereby enhancing the practicability and reliability of this method.
[0073] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.
[0074] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention without departing from the principles and intent of the present invention.
Claims
1. A seismic wave attenuation estimation method based on deep learning, characterized in that: The following steps are involved: Use deep learning algorithms to build a sparse short-time Fourier transform model based on the Unet network; Calculating the sparse time-frequency spectrum of synthetic seismic data as labels based on the sparse short-time Fourier transform model to form a synthetic data set; Using the synthetic data set to train the sparse short-time Fourier transform model to obtain a trained sparse short-time Fourier transform model, and verifying the training effect of the sparse short-time Fourier transform model; Using a transfer learning method, constructing an unsupervised sparse time-frequency domain adaptive transformation model based on the trained sparse short-time Fourier transform model; Using unlabeled actual seismic data to train the unsupervised sparse time-frequency domain adaptive transformation model to obtain a trained unsupervised sparse time-frequency domain adaptive transformation model, and verifying the training effect of the unsupervised sparse time-frequency domain adaptive transformation model; The three-dimensional post-stack seismic data are applied to the trained unsupervised sparse time-frequency domain adaptive transform model to estimate the seismic attenuation amplitude.
2. The method for estimating seismic wave attenuation based on deep learning according to claim 1, characterized in that: The sparse short-time Fourier transform model includes a mapping structure, a feature extraction structure and an upsampling structure. The mapping structure includes a 1x3 convolution operation for converting one-dimensional seismic data into a two-dimensional feature map. The feature extraction structure includes a repeated 3x3 convolution and a 2x2 maximum pooling operation unit. The upsampling structure includes a repeated 3x3 convolution operation unit, a 2x2 convolution operation unit and a 1x1 convolution operation unit.
3. The method for estimating seismic wave attenuation based on deep learning according to claim 1, characterized in that: Before the sparse time-frequency spectrum of the synthetic seismic data is used as a label, the following steps are also included: The synthetic seismic data is normalized.
4. The method for estimating seismic wave attenuation based on deep learning according to claim 3, characterized in that: The normalization processing performed on the synthetic seismic data includes at least one of maximum normalization and mean variance normalization.
5. The method for estimating seismic wave attenuation based on deep learning according to claim 1, characterized in that: The method of training the sparse short-time Fourier transform model by using the synthetic data set to obtain a trained sparse short-time Fourier transform model specifically comprises the following steps: The Adam optimization algorithm is used as the optimizer of the sparse short-time Fourier transform model, and the mean square error loss function is selected to measure the difference between the learning results of the sparse short-time Fourier transform model and its corresponding labels.
6. The method for estimating seismic wave attenuation based on deep learning according to claim 1, characterized in that: The method of using the transfer learning method to construct an unsupervised sparse time-frequency domain adaptive transformation model based on the trained sparse short-time Fourier transform model specifically includes the following steps: The unsupervised sparse time-frequency domain adaptive transformation model includes a source domain and a target domain, and the similarity between the source domain and the target domain is calculated by cosine distance.
7. The method for estimating seismic wave attenuation based on deep learning according to claim 1, characterized in that: The method of applying the three-dimensional post-stack seismic data to the trained unsupervised sparse time-frequency domain adaptive transformation model to estimate the seismic attenuation amplitude specifically includes the following steps: Using high-frequency three-dimensional post-stack seismic data and low-frequency three-dimensional post-stack seismic data to transform the trained unsupervised sparse time-frequency domain adaptive transformation model to obtain high-frequency seismic amplitude and low-frequency seismic amplitude respectively; The difference between the high-frequency earthquake amplitude and the low-frequency earthquake amplitude is calculated, and the intensity of earthquake attenuation is estimated according to the difference.
8. A seismic wave attenuation estimation system based on deep learning, characterized in that: The seismic wave attenuation estimation system is implemented by the seismic wave attenuation estimation method according to any one of claims 1-7.