A short-time passive source seismic data direct reconstruction method based on MDUNETR
Patent Information
- Application Number
- CN202311128908.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-04
- Publication Date
- 2026-10-09
- Estimated Expiration
- 2043-09-04
AI Technical Summary
现有深度学习技术可从短时段的重构结果预测得到长时段的重构结果，但仍需要依赖互相关运算
[0035]This invention, based on deep learning technology, requires only a short period of passive source seismic records to obtain reconstruction results. This demonstrates the significant role of this model in capturing complex spatiotemporal patterns of seismic data and generating accurate predictions. Furthermore, this invention reduces the impact of prior assumptions on the reconstruction results, improves data utilization efficiency, and enhances time efficiency, enabling rapid real-time data processing. Specifically, the following features are described:
Smart Images
Figure CN117169956B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of seismic exploration technology, and in particular relates to a direct reconstruction method for short-time passive source seismic data based on MDUNETR. Background Technology
[0002] Seismic interferometry can convert passive noise source signals into deterministic, effective signals, and use these reconstructed signals to image and invert subsurface media. However, traditional seismic interferometry is usually based on some theoretical assumptions: such as the passive source surface needing a sufficiently wide spatial distribution to form a closed structure with the free surface and to enclose the entire detector array to ensure wavefield dispersion, i.e., the medium cannot be illuminated from only one side; and the medium within the closed structure should be lossless. Although subsequent developments based on the convolution reciprocity theorem, such as channel-wise deconvolution and multidimensional deconvolution, have mitigated the influence of anisotropic source distribution and dissipative media, an important integral term remains in passive source seismic interferometry theory: recording time. Every passive source seismic interferometry technique requires a sufficiently long observation record for the retrieved Green's function to converge. Although from a mathematical perspective, using long-term seismic records in calculations can enhance the statistical analysis of physical events, from a practical computational perspective, retrieving physical events from long-term observation records means a significant increase in computational costs and a loss of immediacy in monitoring changes in the subsurface media.
[0003] Deep learning methods overcome the limitations of traditional methods in seismic interferometry, possessing advantages such as automatic feature learning, robustness, nonlinear modeling capabilities, and efficient processing. They offer new possibilities and potential for seismic data processing and subsurface structure research. Due to their powerful nonlinear modeling capabilities and numerous adjustable parameters, deep learning models can automatically learn and optimize feature representations at different levels during training. This automatic learning characteristic enables deep learning models to extract and represent complex features of subsurface structures from large amounts of seismic data, thereby improving the accuracy and reliability of seismic interferometry results. While existing deep learning techniques can predict long-term reconstruction results from short-term reconstruction results, they still rely on cross-correlation calculations.
[0004] Therefore, in view of the above situation, there is an urgent need to develop a direct reconstruction method for short-time passive source seismic data based on MDUNETR to overcome the shortcomings in current practical applications. Summary of the Invention
[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a direct reconstruction method for short-time passive source seismic data based on MDUNETR, thereby solving the problems mentioned in the background.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] A direct reconstruction method for short-time passive source seismic data based on MDUNETR includes the following steps:
[0008] Step 1: Passive source seismic data simulation: Establish a velocity model for complex media, define the simulation range and its boundary conditions, set relevant parameters of the seismic source, and use the numerical method of finite difference to simulate the propagation of seismic waves in the subsurface medium to obtain numerical simulation results;
[0009] Step 2: Create a training dataset: Process the numerical simulation results from Step 1 using seismic interferometry to obtain the reconstruction results and use them as labels;
[0010] Step 3: Construct the MDUNETR network model: Use skip connections to input the features extracted from the Transformer encoding module into the CNN-based decoder module, and add a multi-scale feature dense fusion method in the decoding part to realize the reuse of multi-scale features and obtain the constructed MDUNETR network model.
[0011] Step 4, Model Training: Use the training dataset from Step 2 to train the model built in Step 3. Randomly select 80% as the training set and the remaining 20% as the validation set.
[0012] As a further technical solution of the present invention, in step one, the velocity model of the complex medium includes the velocity, density and attenuation properties of the formation.
[0013] As a further technical solution of the present invention, in step two, the input data includes short-term passive source noise data and virtual source location data, and the label data includes the relevant reconstruction results of long-term passive source seismic records at the corresponding virtual source locations, and the label data is matched with the input data.
[0014] As a further technical solution of the present invention, in step two, the specific method for processing the numerical simulation results in step one is as follows:
[0015] Short-term passive source noise data is extracted and used as part of the network input.
[0016] Seismic interferometry was used to reconstruct long-term passive source seismic records through cross-correlation, and the results were used as network labels.
[0017] Prepare 2880 sets of data, use the Min-Max scaling normalization method to perform a linear transformation on the original data so that the result is mapped to [0,1], and input it into the network for training;
[0018] The following formula is the time-domain calculation formula for cross-correlation type seismic interferometry of noise sources:
[0019]
[0020] C(x B ,x A ,t)=u(x B ,t)*u(x A ,-t).
[0021] Wherein G(x) B ,x A ,t) is x in a homogeneous, lossless acoustic medium A and x B The Green's function between positions, ±t represents the causal and non-causal parts of the Green's function, N(t) represents the virtual source wavelet, and u(x) B ,t) and u(x A (,-t) represents detector x A With x B The received signal.
[0022] As a further technical solution of the present invention, the specific steps for constructing the MDUNETR network model in step three are as follows:
[0023] Step a: Represent the virtual seismic source location as a vector and concatenate it with the noise data to form extended input data of size H×W×2;
[0024] Step b: Using the Vit algorithm-based encoding module, the image is first transformed. The image slicing part divides the input image into a series of equal-sized image blocks, each of size P×P×2, transforming the H×W×2 image into an N×(P) image. 2 A sequence of (×2), containing N = H × W / P 2 The Transformer structure takes a number of image patches and converts the input image into a one-dimensional sequence. Each image patch corresponds to a vector. The input is a two-dimensional matrix with the shape (N, D), where N is the length of the sequence and D is the dimension of each vector in the sequence.
[0025] Step c: Perform a linear transformation on each segmented image patch using the input embedding part, mapping it to a continuous vector representation. Add positional encoding to the input embedding vector to help the model distinguish image patches at different locations.
[0026] Step d: Use an 8-layer Transformer encoder to process the image sequence and set P=16. Each encoder layer consists of a multi-head self-attention layer and a multi-layer perceptron layer. The multi-head self-attention layer consists of 8 parallel self-attention heads, which are used to establish long-distance dependencies between different positions of the model and capture global information. The multi-layer perceptron layer independently performs nonlinear transformations on the representation vector at each position to enhance the model's expressive and modeling capabilities. While preserving global information, it introduces more complex nonlinear relationships and performs deeper feature extraction on the input vector. Four feature maps (L2, L4, L6, and L8) are extracted from the stacked Transformer encoder output and reshaped and convolutional operations are performed to change their resolution and number of channels and match them with the decoding module.
[0027] Step e: Establish connections between the four layers of features from step d and the preceding layers, including between layers of different branches, and perform connection operations on all layers as the output of the module.
[0028] Step f: Reshape the output of the Transformer into tensors of H / 2×W / 2 (L2), H / 4×W / 4 (L4), H / 6×W / 6 (L6), and H / 8×W / 8 (L8), respectively. Upsample L8 using a deconvolutional layer, then connect L6 to it, and input it into a multi-scale feature dense fusion method to reuse the output and capture more comprehensive features.
[0029] As a further technical solution of the present invention, in step f, the multi-scale feature dense fusion method uses two 3×3 and 7×7 convolutional layers for parallel operation to comprehensively learn information from different receptive fields.
[0030] As a further technical solution of the present invention, in step four, the model training employs a loss function and an optimizer. The loss function is selected as root mean square error, and the calculation formula is as follows:
[0031]
[0032] Among them, P i Y represents the predicted value of the i-th sample. i The table represents the label value of the i-th sample, and n is the number of samples in the training set.
[0033] As a further technical solution of the present invention, the optimizer is selected as Adam optimizer, the learning rate is set to 0.0001, the batch size is set to 64, the number of training rounds is set to 1000, and the training data is shuffled in each training round.
[0034] Compared with the prior art, the beneficial effects of the present invention are:
[0035] This invention, based on deep learning technology, requires only a short period of passive source seismic records to obtain reconstruction results. This demonstrates the significant role of this model in capturing complex spatiotemporal patterns of seismic data and generating accurate predictions. Furthermore, this invention reduces the impact of prior assumptions on the reconstruction results, improves data utilization efficiency, and enhances time efficiency, enabling rapid real-time data processing. Specifically, the following features are described:
[0036] 1. Reduced upfront assumptions: Traditional methods typically require a series of upfront assumptions, such as stationarity or uniformity of background noise. Deep learning models can learn directly from the data, thus reducing the need for these presuppositions;
[0037] 2. Data utilization efficiency: The model can make full use of noisy data to extract information about underground structures. This allows us to obtain more data on underground geological features in a relatively short time, thereby improving data utilization efficiency;
[0038] 3. Fast real-time processing: After the model is trained, its prediction speed is very fast, and it can achieve near real-time seismic interferometric reconstruction.
[0039] To more clearly illustrate the structural features and effects of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Attached Figure Description
[0040] Figure 1 The flowchart illustrates the direct reconstruction method for short-time passive source seismic data based on MDUNETR, as provided in this embodiment of the invention.
[0041] Figure 2 This is a velocity model diagram for step one of the methods provided in this embodiment of the invention.
[0042] Figure 3 This is a network architecture diagram for step three of the method provided in this embodiment of the invention.
[0043] Figure 4 This is a diagram showing the original passive source record in this invention.
[0044] Figure 5 This is a display diagram showing the network prediction results for the virtual source location x = 1000m in this invention.
[0045] Figure 6 This is a display diagram showing the result of cross-correlation reconstruction of the virtual source position x = 1000m in this invention.
[0046] Figure 7 This is a comparison chart showing the difference between the prediction results and the label data in this invention.
[0047] Figure 8This is a display diagram of the single-channel record normalization comparison results in this invention.
[0048] Figure 9 This is a display diagram showing the imaging results of the predicted data in this invention. Detailed Implementation
[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0050] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.
[0051] like Figures 1 to 3 As shown, a direct reconstruction method for short-time passive source seismic data based on MDUNETR, provided as an embodiment of the present invention, includes the following steps:
[0052] Step 1: Passive source seismic data simulation: Establish a velocity model for complex media, define the simulation range and its boundary conditions, set relevant parameters of the seismic source, and use the numerical method of finite difference to simulate the propagation of seismic waves in the subsurface medium to obtain numerical simulation results;
[0053] Step 2: Create a training dataset: Process the numerical simulation results from Step 1 using seismic interferometry to obtain the reconstruction results and use them as labels;
[0054] Step 3: Construct the MDUNETR network model: Use skip connections to input the features extracted from the Transformer encoding module into the CNN-based decoder module, and add a multi-scale feature dense fusion method in the decoding part to realize the reuse of multi-scale features and obtain the constructed MDUNETR network model.
[0055] Step 4, Model Training: Use the training dataset from Step 2 to train the model built in Step 3. Randomly select 80% as the training set and the remaining 20% as the validation set.
[0056] In this embodiment, a network structure called Multi-scale Dense Block (MDB) based on the Transformer and multi-scale feature-dense fusion method is used, with the passive source reconstruction result as the label and the original noise record as the input. The network model obtained by this method can achieve accurate reconstruction of short-term passive source noise records.
[0057] like Figure 2As shown, in a preferred embodiment of the present invention, in step one, the velocity model of the complex medium includes the velocity, density and attenuation properties of the strata. In the numerical simulation stage, the finite difference forward modeling method is used. In the Marmousi model, 5000 randomly excited seismic sources are set uniformly distributed underground, with a grid spacing of 10m and a sampling interval of 1ms. 192 uniformly distributed detectors are deployed on the surface to continuously receive noise records.
[0058] In a preferred embodiment of the present invention, in step two, the input data includes short-term passive source noise data and virtual source location data (all detectors are regarded as virtual sources), and the label data includes the relevant reconstruction results of long-term passive source seismic records at the corresponding virtual source locations, and the label data is matched with the input data.
[0059] In a preferred embodiment of the present invention, the specific method for processing the numerical simulation results in step one in step two is as follows:
[0060] Short-term passive source noise data is extracted and used as part of the network input.
[0061] Seismic interferometry was used to reconstruct long-term passive source seismic records through cross-correlation, and the results were used as network labels.
[0062] Prepare 2880 sets of data, use the Min-Max scaling normalization method to perform a linear transformation on the original data so that the result is mapped to [0,1], and input it into the network for training;
[0063] The following formula is the time-domain calculation formula for cross-correlation type seismic interferometry of noise sources:
[0064]
[0065] C(x B ,x A ,t)=u(x B ,t)*u(x A ,-t).
[0066] Wherein G(x) B ,x A ,t) is x in a homogeneous, lossless acoustic medium A and x B The Green's function between positions, ±t represents the causal and non-causal parts of the Green's function, N(t) represents the virtual source wavelet, and u(x) B ,t) and u(x A (,-t) represents detector x A With x B The received signal.
[0067] In a preferred embodiment of the present invention, in step three, although the Transformer has the ability to learn global information, its ability to capture local information is not strong. The features extracted by the four Transformer encoders are skipped and connected to the CNN-based decoder. Furthermore, a multi-scale feature fusion method is used in the decoding part to achieve comprehensive learning of local and global information.
[0068] The specific steps for constructing the MDUNETR network model are as follows:
[0069] Step a: Represent the virtual seismic source location as a vector and concatenate it with the noise data to form extended input data of size H×W×2;
[0070] Step b: Using the Vit algorithm-based encoding module, the image is first transformed. The image slicing part divides the input image into a series of equal-sized image blocks, each of size P×P×2, transforming the H×W×2 image into an N×(P) image. 2 A sequence of (×2), containing N = H × W / P 2 The Transformer structure takes a number of image patches and converts the input image into a one-dimensional sequence. Each image patch corresponds to a vector. The input is a two-dimensional matrix with the shape (N, D), where N is the length of the sequence and D is the dimension of each vector in the sequence.
[0071] Step c: At this point, the dimension of each image patch is P. 2 However, the actual required vector dimension is D, so image patches need to be processed. The input embedding part is used to perform a linear transformation on each segmented image patch, mapping it to a continuous vector representation. Position encoding is added to the input embedding vector so that the model can distinguish image patches at different positions.
[0072] Step d: Use an 8-layer Transformer encoder to process the image sequence and set P=16. Each encoder layer consists of a multi-head self-attention layer and a multi-layer perceptron layer. The multi-head self-attention layer consists of 8 parallel self-attention heads, which are used to establish long-distance dependencies between different positions of the model and capture global information. The multi-layer perceptron layer independently performs nonlinear transformations on the representation vector at each position to enhance the model's expressive and modeling capabilities. While preserving global information, it introduces more complex nonlinear relationships and performs deeper feature extraction on the input vector. Four feature maps (L2, L4, L6, and L8) are extracted from the stacked Transformer encoder output and reshaped and convolutional operations are performed to change their resolution and number of channels and match them with the decoding module.
[0073] Step e: Establish connections between the four layers of features from step d and the preceding layers, including between layers of different branches, and perform connection operations on all layers as the output of the module.
[0074] Step f: Reshape the output of the Transformer into tensors of H / 2×W / 2 (L2), H / 4×W / 4 (L4), H / 6×W / 6 (L6), and H / 8×W / 8 (L8), respectively. Upsample L8 using a deconvolutional layer, then connect L6 to it, and input it into a multi-scale feature dense fusion method to reuse the output and capture more comprehensive features.
[0075] In a preferred embodiment of the present invention, in step f, the multi-scale feature dense fusion method uses two 3×3 and 7×7 convolutional layers for parallel computation to comprehensively learn information from different receptive fields.
[0076] In a preferred embodiment of the present invention, in step four, the model training employs a loss function and an optimizer. The loss function is selected as root mean square error (MSE), and the calculation formula is as follows:
[0077]
[0078] Among them, P i Y represents the predicted value of the i-th sample. i The table represents the label value of the i-th sample, and n is the number of samples in the training set.
[0079] In a preferred embodiment of the present invention, the optimizer is Adam optimizer, the learning rate is set to 0.0001, the batch size is set to 64, the number of training epochs is set to 1000, and the training data is shuffled in each training epoch.
[0080] like Figures 4 to 9 As shown in the figure, as a preferred embodiment of the present invention, the test results demonstrate the network's ability to learn and capture potential patterns in seismic data. To illustrate this, the results at a virtual source location x = 1000m were selected for demonstration. It was found that effective signals could not be captured in the original passive source seismic records, while the network's predictions showed a high degree of consistency with the labeled data. Single-channel records were selected from the test results for normalization comparison, revealing excellent fitting between the prediction results and the expected data, with a high degree of matching in their relative amplitudes. The imaging effect of the prediction results demonstrates its ability to accurately recover underground structures. This proves that the present invention can recover real underground reflection information and accurately predict and reconstruct the required seismic features.
[0081] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A direct reconstruction method for short-time passive source seismic data based on MDUNETR, characterized in that, Includes the following steps: Step 1: Passive source seismic data simulation: Establish a velocity model for complex media, define the simulation range and its boundary conditions, set relevant parameters of the seismic source, and use the numerical method of finite difference to simulate the propagation of seismic waves in the subsurface medium to obtain numerical simulation results; Step 2: Create a training dataset: Process the numerical simulation results from Step 1 using seismic interferometry to obtain the reconstruction results and use them as labels; Step 3: Construct the MDUNETR network model: Use skip connections to input the features extracted from the Transformer encoding module into the CNN-based decoder module, and add a multi-scale feature dense fusion method in the decoding part to realize the reuse of multi-scale features and obtain the constructed MDUNETR network model. Step 4, Model Training: Use the training dataset from Step 2 to train the model built in Step 3. Randomly select 80% as the training set and the remaining 20% as the validation set.
2. The method for direct reconstruction of short-time passive source seismic data based on MDUNETR according to claim 1, characterized in that, In step one, the velocity model of the complex medium includes the velocity, density, and attenuation properties of the formation.
3. The method for direct reconstruction of short-time passive source seismic data based on MDUNETR according to claim 1, characterized in that, In step two, the input data includes short-term passive source noise data and virtual source location data, and the label data includes the relevant reconstruction results of long-term passive source seismic records at the corresponding virtual source locations.
4. The method for direct reconstruction of short-time passive source seismic data based on MDUNETR according to claim 1, characterized in that, In step two, the specific method for processing the numerical simulation results from step one is as follows: Short-term passive source noise data is extracted and used as part of the network input. Seismic interferometry was used to reconstruct long-term passive source seismic records through cross-correlation, and the results were used as network labels. Prepare 2880 sets of data, use the Min-Max scaling normalization method to perform a linear transformation on the original data so that the result is mapped to [0,1], and input it into the network for training; The following formula is the time-domain calculation formula for cross-correlation type seismic interferometry of noise sources: ； ； in, In a homogeneous and lossless acoustic medium and Green's function between positions, Represents the causal and non-causal parts of the Green's function. Represents the virtual source wavelet. and Representative detector and The received signal.
5. The method for direct reconstruction of short-time passive source seismic data based on MDUNETR according to claim 1, characterized in that, In step three, the specific steps for constructing the MDUNETR network model are as follows: Step a: Represent the virtual seismic source location as a vector and concatenate it with the noise data to form expanded input data of size [value missing]. ; Step b: Using the Vit algorithm-based encoding module, the image is first converted. The image slicing part divides the input image into a series of equal-sized image blocks, each small image block being [size missing].
2. The image becomes a The sequence contains The Transformer structure converts the input image into a one-dimensional sequence of image patches, where each patch corresponds to a vector. The input is a two-dimensional matrix, the shape of which can be represented as... , is the length of the sequence, and D is the dimension of each vector in the sequence; Step c: Perform a linear transformation on each segmented image patch using the input embedding part, mapping it to a continuous vector representation. Add positional encoding to the input embedding vector to help the model distinguish image patches at different locations. Step d: Use an 8-layer Transformer encoder to process the image sequence, and make... Each encoder layer consists of a multi-head self-attention layer and a multi-layer perceptron layer. The multi-head self-attention layer consists of eight parallel self-attention heads, which are used to establish long-distance dependencies between different positions of the model and capture global information. The multi-layer perceptron layer independently performs nonlinear transformations on the representation vector at each position to enhance the model's expressive and modeling capabilities. While preserving global information, it introduces more complex nonlinear relationships and performs deeper feature extraction on the input vector. Four layers of feature maps are extracted from the output of the stacked Transformer encoder, including L2, L4, L6, and L8 feature maps, and then reshaped and convolutional operations are performed to change their resolution and number of channels and match them with the decoding module. Step e: Establish connections between the four layers of features from step d and the preceding layers, including between layers of different branches, and perform connection operations on all layers as the output of the module. Step f: Reshape the Transformer output into L2: L4: L6: And L8: The tensor of L8 is upsampled using a deconvolutional layer, then connected to L6 and fed into a multi-scale feature-dense fusion method to reuse the output and capture more comprehensive features.
6. The method for direct reconstruction of short-time passive source seismic data based on MDUNETR according to claim 5, characterized in that, In step f, the multi-scale feature dense fusion method uses two 3×3 and 7×7 convolutional layers for parallel computation to comprehensively learn information from different receptive fields.
7. The method for direct reconstruction of short-time passive source seismic data based on MDUNETR according to claim 1, characterized in that, In step four, model training employs a loss function and an optimizer. The root mean square error (RMSE) is chosen as the loss function, and its calculation formula is as follows: ； in, This represents the predicted value of the i-th sample. This represents the label value of the i-th sample. This represents the number of samples in the training set.
8. The method for direct reconstruction of short-time passive source seismic data based on MDUNETR according to claim 7, characterized in that, The optimizer is Adam, the learning rate is set to 0.0001, the batch size is set to 64, the number of training epochs is set to 1000, and the training data is shuffled in each training epoch.