Seismic data sub-component extraction method based on rock physics and deep learning
By combining rock physics and deep learning methods, a seismic data subcomponent extraction model is constructed, which solves the problem of extracting small changes in seismic data and improves the accuracy of reservoir oil and gas detection.
Patent Information
- Application Number
- CN202510829562.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-20
- Publication Date
- 2025-08-05
AI Technical Summary
The prior art is difficult to effectively extract subcomponent information that reacts to slight changes in the properties of geological media from seismic data, resulting in insufficient oil and gas detection accuracy in deep buried reservoirs.
Combining rock physics and deep learning, a viscoelastic biphasic medium wave equation model is constructed. Through high-fidelity and weak-proof signal processing and semi-supervised learning, a variety of network models are used to extract seismic data subcomponents, including bidirectional time domain convolution networks, bidirectional selective state space models and attention mechanism algorithms, and feature extraction is carried out in combination with seismic wave propagation theory and modal decomposition technology.
The accuracy of reservoir oil and gas detection can be improved, and subcomponent information that reacts to slight changes in the properties of geological media can be extracted from seismic data, enhancing the accuracy of oil and gas detection.
Smart Images
Figure CN120428336A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of reservoir prediction and oil and gas detection, and in particular to a method for extracting secondary components from seismic data based on rock physics and deep learning. Background Art
[0002] Directly predicting the distribution of oil and gas reservoirs using seismic data is a highly sought-after goal in oil and gas exploration. The difficulty in predicting oil and gas in deeply buried reservoirs lies in identifying or extracting information about the seismic response signals of the reservoir pore fluids. If such responses are observable, they are likely to manifest only as micro-events in the seismic record, reflected in the secondary components of the seismic signature, making them difficult to visually identify. Current wave equations describing seismic waves are approximate equations derived under certain assumptions and are unlikely to explicitly reflect the response of the rock pore fluids.
[0003] Therefore, extracting secondary component information that can reflect subtle changes in geological medium properties from complex seismic wave fields, such as changes or differences in rock pore fluid properties and their saturation, is of great significance and value for improving the accuracy of deep-buried reservoir fluid prediction and oil and gas detection. However, how to effectively extract this information is a difficult problem that needs to be solved urgently in this technical field. Summary of the Invention
[0004] To solve the above technical problems, the present invention provides a method for extracting secondary components of seismic data based on rock physics and deep learning, so as to solve the problem of effective extraction of secondary components of seismic data and thereby improve the accuracy of reservoir oil and gas detection.
[0005] In order to achieve the above purpose, the technical solution adopted by the present invention is:
[0006] A method for extracting secondary components from seismic data based on rock physics and deep learning, comprising the following steps: S1. Based on well logging data and rock physics test analysis, combined with prior seismic geological information, a deep-buried viscoelastic two-phase macroscopic rock physics model is constructed. The viscoelastic two-phase medium wave equation is used to simulate and calculate the model seismic response with a series of fine-tuned model parameters. S2. Perform high-fidelity weak signal processing on the actual seismic data, then calculate the difference between it and the model seismic response. Combined with the well logging data, analyze the impact of model parameter changes on the seismic response, obtain a series of seismic data subcomponents under model disturbances, and construct a sample data set. S3: Based on the bidirectional time-domain convolutional network, bidirectional selective state-space model, bidirectional gated recurrent unit network and attention mechanism algorithm, combined with seismic wave propagation theory and semi-supervised learning algorithm to build a physics-oriented deep network model; S4. Use modal decomposition fused with sample entropy clustering to decompose the seismic data into data of different frequency bands, and input the original data and data of different frequency bands into different branch networks of the deep network model for model training to obtain a seismic data sub-component extraction model; S5. Input the actual seismic data after the fidelity and weakness preservation processing into the trained model to obtain the secondary component data of the seismic data, and calibrate and analyze the extracted results in combination with the actual exploration data; Preferably, the deep burial mentioned in step S1 requires consideration of the influence of various factors such as temperature, static pressure and pore pressure during rock physics modeling.
[0007] Preferably, the process of performing high-fidelity weak signal protection processing on actual seismic data described in step S1 may include fine static correction processing combining tomographic static correction and reflection wave residual static correction, broadband high-fidelity weak signal noise suppression processing, pre-stack high-fidelity amplitude compensation and high-resolution processing, and wide-azimuth data OVT domain high-precision migration imaging processing, and the purpose of weak signal (secondary component) protection is taken into consideration in each step.
[0008] Preferably, in the calculation process of the seismic data secondary components in step S2, the model seismic response is regarded as the main component, and the difference between the actual data and the model response is regarded as the secondary component of the actual medium "disturbance" relative to the model.
[0009] Preferably, the deep network model described in step S3 is mainly composed of a bidirectional time-domain convolutional network module, a bidirectional selective state-space model module, a feature fusion module, a bidirectional gated recurrent network module and a fully connected feedforward network module. By integrating the advantages of various classic network models and algorithms, it has stronger representation learning ability and higher stability, and has the flexibility to process different types of data. Moreover, only the last feedforward neural network module needs to be adjusted to be able to model and learn the data from different angles.
[0010] Specifically, the bidirectional time-domain convolutional network module and the bidirectional selective state-space model module are first used in parallel to identify and capture local and global features in the data, and the feature information is fused in the channel dimension through the feature fusion module; then the bidirectional gated recurrent unit network combined with the attention mechanism is used to finely model the fused feature information; finally, the fully connected feedforward network module is used to map the output extraction results.
[0011] The bidirectional time-domain convolutional network module in the deep network model is used to capture local features and long-term dependencies in sequence data. The bidirectional selective state-space model module is used to capture long-range related information rather than local features in sequence data. The feature fusion module is used to fuse the feature information extracted by the bidirectional time-domain convolutional network module and the bidirectional selective state-space model. The bidirectional gated recurrent unit network module combined with the attention mechanism is used to capture the temporal features and global contextual association information in sequence data.
[0012] The bidirectional time-domain convolutional network consists of forward and backward time-domain convolutional network layers. It can extract feature information from sequence data in both forward and reverse directions, and can take into account past and future information at the same time. The time-domain convolutional network is an improved version of the convolutional neural network, which can better understand the dynamic changes in sequence data. However, it can only learn feature information along a single direction of the sequence data, and it is difficult to effectively capture future sequence feature information in the data.
[0013] The bidirectional selective state-space model, also known as the bidirectional Mamba model, is a network model that rivals or even surpasses the Transformer. It integrates time-varying parameters into a structured state-space sequence model and leverages hardware-aware algorithms to accelerate training and inference. This overcomes the computational efficiency bottlenecks of traditional Transformers and offers powerful contextual modeling capabilities and linear complexity. The bidirectional Mamba network is an innovative design that combines the Mamba network with bidirectional processing capabilities. This overcomes the limitations of the Mamba model's unidirectional processing and can simultaneously capture richer global feature information in sequence data.
[0014] The bidirectional gated recurrent unit network is constructed using two gated recurrent unit network layers in opposite directions to realize the processing of bidirectional temporal information. The forward gated recurrent unit network layer calculates the hidden layer information along the forward order of the time data, and the backward gated recurrent unit network layer calculates the hidden layer information along the reverse order of the time data. The two hidden layer information are then fused to obtain the output of the bidirectional gated recurrent unit network.
[0015] Preferably, the semi-supervised learning used in step S3 is a learning method between supervised learning and unsupervised learning, which can use both labeled and unlabeled seismic data for training, capturing explicit physical knowledge from labeled data while mining implicit knowledge from a large amount of unlabeled data. The physics-oriented deep network model is a closed-loop semi-supervised learning framework formed by combining a deep network model with a neural network forward operator for assisting inversion model training. It can fully utilize well seismic data and train both networks simultaneously. Preferably, the cascade modal decomposition of seismic data by fusion sample entropy clustering described in step S4 can introduce frequency domain information into network model training. It first performs modal decomposition on the seismic data to obtain a series of intrinsic mode function components, then uses the sample entropy algorithm to calculate the sample entropy of each component to quantify the complexity of each component and applies the clustering algorithm to perform cluster reconstruction, and performs secondary modal decomposition on the components with high complexity to further extract features and filter noise.
[0016] Preferably, in step S4, a sequence-to-sequence approach, an adaptive learning rate adjustment strategy, an early stopping mechanism, and the like may be adopted during the training process of the deep network model.
[0017] Compared with the prior art, the present invention has at least the following beneficial effects: This paper proposes a method for extracting secondary components from seismic data that combines rock physics with deep learning. This method can extract secondary component information from seismic data that reflects subtle changes in geological properties. By integrating the advantages of multiple classical network models and algorithms, this method overcomes the problem of insufficient data feature information extracted by a single network model and extracts richer seismic geological feature information from seismic data. By constructing a physics-driven semi-supervised learning architecture, this method can capture explicit physical knowledge from labeled data while mining implicit knowledge from large amounts of unlabeled data, thereby improving the model's applicability to a certain extent. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION
[0019] In order to enable people skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of the specification, not all of the embodiments.
[0020] like Figure 1 As shown, the present invention provides a method for extracting secondary components from seismic data based on rock physics and deep learning, which mainly includes the following steps: Step S1: Based on well logging data and rock physics test analysis, combined with prior seismic geological information, a deep-buried viscoelastic two-phase macroscopic rock physics model is constructed, and the viscoelastic two-phase medium wave equation is used to simulate and calculate the model seismic response with a series of fine-tuned model parameters; Specifically, the well logging data of the target area are first statistically analyzed, and rock physical testing and analysis are carried out based on this to construct a deep-buried viscoelastic two-phase macroscopic rock physical model. The constructed rock physical model is then improved and optimized in combination with seismic geological conditions and prior information. The viscoelastic two-phase medium wave equation is then used to simulate and calculate the seismic response of the model. On this basis, a series of seismic responses with fine-tuning of the model parameters are obtained by adjusting the changes in the model parameters.
[0021] Step S2: Perform high-fidelity weak signal processing on the actual seismic data, then calculate the difference between the actual seismic data and the model seismic response, and analyze the impact of model parameter changes on the seismic response in combination with well logging data to obtain a series of seismic data subcomponents under model disturbances and construct a sample data set; Specifically, the acquired original seismic data are first subjected to fine static correction processing combining tomographic static correction and reflected wave residual static correction, noise suppression processing of broadband high-fidelity weak signal, pre-stack high-fidelity amplitude compensation and high-resolution processing, and wide-azimuth data OVT domain high-precision migration imaging processing, so as to protect the low-frequency, high-frequency weak signals and weak energy weak signals in the original seismic data and obtain high-quality fidelity and weak seismic data; on this basis, the difference between the seismic data after fidelity and weak signal processing and the model seismic response is calculated, and the influence of model parameter changes on the seismic response is analyzed in combination with seismic geological prior information to obtain the difference between the actual data and the seismic responses of different models, and then obtain a series of seismic data subcomponents under model disturbances, which are then combined with the actual seismic data and the model seismic response to construct a sample data set.
[0022] Step S3: Based on the bidirectional time-domain convolutional network, the bidirectional selective state-space model, the bidirectional gated recurrent unit network, and the attention mechanism algorithm, a physics-oriented deep network model is constructed in combination with the earthquake wave propagation theory and the semi-supervised learning algorithm; Specifically, the deep network module mainly consists of a bidirectional time-domain convolutional network module, a bidirectional selective state-space model module, a feature fusion module, a bidirectional gated recurrent network module, and a fully connected feedforward network module. First, the bidirectional time-domain convolutional neural network and the bidirectional selective state-space model are used in parallel to identify and capture local and global features in the data. The former is used to capture local features and dependencies in the sequence data, while the latter captures long-range related information rather than local features in the sequence data. The feature fusion module fuses the feature information in the channel dimension. Subsequently, a bidirectional gated recurrent unit network combined with an attention mechanism is used to finely model the fused feature information, capturing the temporal features and previous and next correlations in the sequence data. Finally, the fully connected feedforward network module maps the output extraction results. On this basis, the constructed deep network model is combined with a neural network forward operator used to assist in inversion model training to form a closed-loop semi-supervised learning framework. A physics-driven semi-supervised learning deep network model is constructed, and both network models are trained simultaneously while fully utilizing well-seismic data.
[0023] Step S4: using modal decomposition of fused sample entropy clustering to decompose the seismic data into data of different frequency bands, and inputting the original data and the data of different frequency bands into different branch networks of the deep network model for model training, thereby obtaining a seismic data sub-component extraction model; Specifically, the seismic data is first subjected to adaptive noise complete set empirical mode decomposition (EMD) to obtain a series of intrinsic mode function components. The sample entropy of all modal components is then calculated using a sample entropy algorithm, and clustering is then applied to reconstruct the components. A variational mode decomposition algorithm is then used to perform secondary modal decomposition on the more complex modal function components to further extract features and filter out noise, decomposing the seismic data into data of different frequency bands. This data is then combined with the original data, forming labeled and unlabeled sample data, which are then fed into different branches of a deep network model for model training. This training is performed using a sequence-to-sequence approach combined with an adaptive learning rate adjustment strategy and an early stopping mechanism. Model training terminates when the model's error loss is minimized or a pre-set termination condition is met, resulting in a satisfactory seismic data subcomponent extraction model.
[0024] Step S5: Input the actual seismic data after the fidelity and weakness preservation processing into the trained model to obtain the secondary component data of the seismic data, and calibrate and analyze the extracted results in combination with the actual exploration data.
[0025] Specifically, the saved trained model is loaded, and the actual seismic data that has been processed with high fidelity and weak preservation is input into the trained model. The hidden layer receives the data and uses the trained model for calculation, and the calculation results are passed to the output layer to obtain the sub-component results of the seismic data and perform visual analysis; on this basis, the extracted results are corrected and analyzed according to the logging interpretation and production data of the wells drilled in the target area.
[0026] Those skilled in the art will appreciate that the above embodiments are merely illustrative of the beneficial effects of the present invention and are not exhaustive. Any modifications, equivalent substitutions, improvements, etc. made without departing from the principles and spirit of this specification shall not be excluded from the scope of protection of the present invention.
Claims
1. A method for extracting secondary components from seismic data based on rock physics and deep learning, characterized in that: Specifically include: (1) Based on well logging data and rock physics test analysis, combined with prior seismic geological information, a deep-buried viscoelastic two-phase macroscopic rock physics model is constructed. The viscoelastic two-phase medium wave equation is used to simulate and calculate the seismic response of a series of model parameters fine-tuned; (2) Perform high-fidelity weak signal processing on the actual seismic data, then calculate the difference between it and the model seismic response, combine it with the well logging data to analyze the impact of model parameter changes on the seismic response, obtain a series of seismic data subcomponents under model disturbances, and construct a sample data set; (3) Based on the bidirectional time-domain convolutional network, bidirectional selective state-space model, bidirectional gated recurrent unit network and attention mechanism algorithm, combined with seismic wave propagation theory and semi-supervised learning algorithm to build a physics-oriented deep network model; (4) Using modal decomposition fused with sample entropy clustering, the seismic data is decomposed into data of different frequency bands, and the original data and data of different frequency bands are respectively input into different branch networks of the deep network model for model training to obtain a seismic data sub-component extraction model; (5) The actual seismic data after the fidelity and weakness treatment is input into the trained model to obtain the secondary component data of the seismic data, and the extracted results are corrected and analyzed in combination with the actual exploration data.
2. The method for extracting secondary components from seismic data based on rock physics and deep learning according to claim 1, characterized in that: The deep burial mentioned in step (1) requires consideration of various factors such as temperature, static pressure, and pore pressure during rock physics modeling.
3. The method for extracting secondary components from seismic data based on rock physics and deep learning according to claim 1, characterized in that: In the calculation process of the secondary components of the seismic data described in step (2), the model seismic response is regarded as the main component, and the difference between the actual data and the model response is regarded as the secondary component of the "disturbance" of the actual medium relative to the model.
4. The method for extracting secondary components from seismic data based on rock physics and deep learning according to claim 1, characterized in that: The deep network model described in step (3) first identifies and captures local and global features in the data using a bidirectional time-domain convolutional network and a bidirectional selective state-space model in parallel, and fuses the feature information in the channel dimension; then uses a bidirectional gated recurrent unit network combined with an attention mechanism to finely model the fused feature information; finally, the extraction result is output through a fully connected feedforward network mapping; by integrating the advantages of various classic network models and algorithms, the deep network model has stronger representation learning ability and higher stability, and has the flexibility to process different types of data. Moreover, only the final feedforward neural network module needs to be adjusted to model and learn the data from different angles.
5. The method for extracting secondary components from seismic data based on rock physics and deep learning according to claim 1, characterized in that: The semi-supervised learning used in step (3) is a learning method between supervised learning and unsupervised learning. It can use both labeled and unlabeled seismic data for training, capturing explicit physical knowledge from labeled data while mining implicit knowledge from a large amount of unlabeled data. The physics-oriented deep network model is a closed-loop semi-supervised learning framework formed by combining a deep network model with a neural network forward operator used to assist in inversion model training. It can fully utilize well seismic data and train both networks simultaneously.
6. The method for extracting secondary components from seismic data based on rock physics and deep learning according to claim 1, characterized in that: The cascade modal decomposition of seismic data by fusion sample entropy clustering described in step (4) can introduce frequency domain information into network model training. It first performs modal decomposition on the seismic data to obtain a series of intrinsic mode function components, then uses the sample entropy algorithm to calculate the sample entropy of each component to quantify the complexity of each component and applies the clustering algorithm to perform cluster reconstruction, and performs secondary modal decomposition on the components with high complexity to further extract features and filter noise.