Seismic wave impedance inversion method and system based on deep learning

Abstract

The application provides a seismic wave impedance inversion method and system based on deep learning, comprising the following steps: S1, constructing and preprocessing a data set; S2, constructing and training a DA-TransUNet model; extracting features of an input seismic image through a convolutional neural network of the DA-TransUNet model, extracting local features, and generating skip connection features; extracting global features through a Transformer network based on the local features; S3, performing upsampling processing through the convolutional neural network of the DA-TransUNet model, fusing the skip connection features with features of a current layer of a decoder, and obtaining a fusion result as an upsampling result; and S4, constructing a final seismic image prediction result. Through the DA-TransUNet network architecture, the application solves the problems of insufficient precision and low efficiency of traditional seismic wave impedance inversion technology in processing complex geological data.

Description

Technical Field

[0001] This invention belongs to the field of seismic wave impedance inversion technology, specifically, it relates to a seismic wave impedance inversion method and system based on deep learning. Background Technology

[0002] In the field of seismic impedance retrieval, existing techniques typically suffer from several significant limitations and drawbacks. First, many traditional seismic impedance retrieval techniques rely on linear or semi-linear processing flows, which often prove ineffective when dealing with complex geological structures. Second, traditional methods are generally computationally inefficient when processing high-dimensional data. Furthermore, some novel machine learning-based methods often overlook the complexity of seismic data, such as the connections between strata and geological variations at different scales. These models may perform well on specific datasets but lack sufficient generalization ability, making it difficult to adapt to diverse geological environments and different types of seismic data.

[0003] The August 2023 issue of the journal "Progress in Geophysics" published a method for inverting post-stack seismic impedance based on the TransUNet neural network. Although the model of this scheme combines the advantages of U-Net and Transformer, its performance still has room for improvement when dealing with finer-grained geological features. Moreover, it relies on a large amount of labeled data for effective training, which is often limited by data acquisition costs and data quality in practical applications.

[0004] The journal article information is as follows:

[0005] PENG Zhen,XU HuiQun.2024.Post-stack seismic impedance inversionmethod based on TransUNet neural network.Progress in Geophysics,39(2):704-715.doi:10.6038 / pg2024HH0025. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this invention is to provide a seismic wave impedance inversion method and system based on deep learning.

[0007] A seismic wave impedance inversion method based on deep learning provided by the present invention includes:

[0008] Step S1: Construct and preprocess the dataset;

[0009] Step S2: Based on the dataset, construct and train the DA-TransUNet model; extract features from the input seismic image using the convolutional neural network of the DA-TransUNet model, extract local features, and generate skip connection features; based on the local features, extract global features using a Transformer network;

[0010] Step S3: Based on global features, upsampling is performed through the convolutional neural network of the DA-TransUNet model, fusing the features of the skip connections with the features of the current layer of the decoder to obtain the fused result as the upsampling result;

[0011] Step S4: Based on the fusion results, construct the final seismic image prediction results.

[0012] Preferably, in step S1: the preprocessing includes: normalization, data augmentation, and format conversion;

[0013] The normalization process refers to the use of maximum and minimum normalization to process seismic wave impedance inversion, and the mathematical expression is:

[0014] y i ′ =(y i -min(y))÷(max(y)-min(y))

[0015] Among them, y i ′ The data is normalized; y i The original data is represented by y; max(y) is the largest value in the original data; min(y) is the smallest value in the original data; and during the normalization process, all data values ​​are within the range of 0 to 1.

[0016] The dataset includes: labeled data, unlabeled data, and validation data.

[0017] Preferably, in step S2:

[0018] Step S2.0: Based on the dataset, construct and train the DA-TransUNet model;

[0019] Step S2.1: Extract features from the seismic images in the dataset using the convolutional neural network of the DA-TransUNet model to obtain local features;

[0020] Step S2.2: Use DA-Block to refine the output of the convolutional neural network, i.e., the local features, to obtain skip connection features;

[0021] Step S2.3: Based on the skip connection features, extract global features through the Transformer network.

[0022] Preferably, the dataset includes a training set and a validation set;

[0023] The ratio of the training set to the validation set is three to one.

[0024] The validation set includes labeled data;

[0025] The training set is used to train the DA-TransUNet model;

[0026] The ratio of labeled to unlabeled data in the training set is adjusted according to the type of dataset.

[0027] Preferably, the DA-TransUNet model provides positional information for each element in the dataset sequence;

[0028] The DA-Block is composed of a spatial attention mechanism and a channel attention mechanism.

[0029] A seismic impedance inversion system based on deep learning according to the present invention includes:

[0030] Module M1: Constructs and preprocesses the dataset;

[0031] Module M2: Based on the dataset, construct and train the DA-TransUNet model; extract features from the input seismic image using the convolutional neural network of the DA-TransUNet model, extract local features, and generate skip connection features; based on the local features, extract global features using a Transformer network;

[0032] Module M3: Based on global features, it performs upsampling processing through the convolutional neural network of the DA-TransUNet model, fusing the features of skip connections with the features of the current layer of the decoder to obtain a fused result as the upsampling result;

[0033] Module M4: Based on the fusion results, construct the final seismic image prediction results.

[0034] Preferably, in module M1: the preprocessing includes: normalization, data augmentation, and format conversion;

[0035] The normalization process refers to the use of maximum and minimum normalization to process seismic wave impedance inversion, and the mathematical expression is:

[0036] y i ′ =(yi -min(y))÷(max(y)-min(y))

[0037] Among them, y i ′ The data is normalized; y i The original data is represented by y; max(y) is the largest value in the original data; min(y) is the smallest value in the original data; and during the normalization process, all data values ​​are within the range of 0 to 1.

[0038] The dataset includes: labeled data, unlabeled data, and validation data.

[0039] Preferably, in module M2:

[0040] Module M2.0: Constructs and trains the DA-TransUNet model based on the dataset;

[0041] Module M2.1: Extracts features from seismic images in the dataset using a convolutional neural network based on the DA-TransUNet model to obtain local features;

[0042] Module M2.2: Through DA-Block, the output of the convolutional neural network, i.e. the local features, is refined to obtain skip connection features;

[0043] Module M2.3: Based on skip connection features, global features are extracted through a Transformer network.

[0044] Preferably, the dataset includes a training set and a validation set;

[0045] The ratio of the training set to the validation set is three to one.

[0046] The validation set includes labeled data;

[0047] The training set is used to train the DA-TransUNet model;

[0048] The ratio of labeled to unlabeled data in the training set is adjusted according to the type of dataset.

[0049] Preferably, the DA-TransUNet model provides positional information for each element in the dataset sequence;

[0050] The DA-Block is composed of a spatial attention mechanism and a channel attention mechanism.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] 1. This invention solves the problems of insufficient accuracy and low efficiency of traditional seismic impedance inversion technology when processing complex geological data by adopting the DA-TransUNet network architecture.

[0053] 2. The DA-TransUNet model provided by this invention combines the efficient local feature extraction capability of U-Net with the global information processing advantage of Transformer, and significantly enhances the model's perception and processing capabilities at various scales by introducing a dual attention mechanism of spatial attention and channel attention.

[0054] 3. This invention effectively addresses the challenge of limited well logging annotation data through a semi-supervised learning module, significantly improving the model's prediction accuracy for seismic data in unexplored areas, while ensuring higher computational efficiency and lower resource consumption. Attached Figure Description

[0055] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0056] Figure 1 This is a schematic diagram of the process provided by the present invention. Detailed Implementation

[0057] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0058] This invention proposes a seismic impedance inversion method based on deep learning technology, which includes the use of a DA-TransUNet network architecture. This architecture effectively combines the local feature extraction capability of U-Net with the long-range information processing advantage of Transformer, and significantly improves the accuracy and efficiency of seismic impedance inversion through an optimized dual-attention mechanism.

[0059] A seismic wave impedance inversion method based on deep learning provided by the present invention includes:

[0060] Step 1: Construct and preprocess the dataset

[0061] Data preprocessing module: This module is responsible for collecting and preprocessing earthquake data. The preprocessed earthquake image dataset is then sent to the next feature extraction module, namely the convolutional neural network.

[0062] Step 2: Feature extraction using the DA-TransUNet network based on the dataset, i.e., the encoder stage.

[0063] Convolutional Neural Network Module: Using preprocessed seismic image data, local features are extracted through multi-layer convolution and activation functions, and skip connection features are generated. Both local features and skip connection features are refined by DA-Block.

[0064] The Transformer network module is used to receive local features generated from the convolutional neural network, process them through the Transformer's self-attention mechanism, and extract global features.

[0065] Step 3: Decoder and Feature Fusion Stage

[0066] The features processed by the encoder stage are progressively upsampled to gradually restore the feature map to the same resolution as the input image. During the upsampling process, the skip connection features generated by the convolutional neural network module are fused with the features of the current layer of the decoder, which helps to preserve important local feature information and integrate global context.

[0067] Step 4: Construct the final image prediction results

[0068] Based on skip connections and upsampling results, the final seismic image prediction results are constructed.

[0069] In other words, the seismic impedance inversion method based on deep learning provided by the present invention includes:

[0070] Step S1: Construct and preprocess the dataset;

[0071] Step S2: Based on the dataset, construct and train the DA-TransUNet model; extract features from the input seismic image using the convolutional neural network of the DA-TransUNet model, extract local features, and generate skip connection features; based on the local features, extract global features using a Transformer network;

[0072] Step S3: Based on global features, upsampling is performed through the convolutional neural network of the DA-TransUNet model, fusing the features of the skip connections with the features of the current layer of the decoder to obtain the fused result as the upsampling result;

[0073] Step S4: Based on the fusion results, construct the final seismic image prediction results.

[0074] The data preprocessing module is used to build a reliable dataset for model training;

[0075] Specifically, the data preprocessing module uses the Marmousi2 dataset and the SEAM dataset to perform wave impedance calculation, data normalization, and two-dimensional conversion of one-dimensional data to ensure the consistency and standardization of the input data.

[0076] The data preprocessing module applies data augmentation techniques such as random rotation, flipping, and scaling to increase data diversity and improve the robustness and adaptability of the model under different conditions.

[0077] The semi-supervised learning module is used to effectively improve model performance and enhance the accuracy and reliability of predictions for unknown seismic data when limited well logging labeled data is available. Semi-supervised learning is a machine learning paradigm that lies between supervised and unsupervised learning. In semi-supervised learning, the model is trained using only a small amount of labeled data and a large amount of unlabeled data, making it suitable for situations where obtaining labeled data is difficult.

[0078] The semi-supervised learning module is primarily used in the model training phase. It divides the dataset into a small amount of labeled data, a large amount of unlabeled data, and a validation dataset. Specifically, the data is divided into training and validation sets in a 3:1 ratio, with the validation set containing labeled data used to evaluate model performance and adjust parameters during iterations. The ratio of labeled to unlabeled data in the training set is dynamically adjusted based on the specific dataset. Supervised learning is performed using labeled data to ensure the model can initially identify geological structures. A semi-supervised learning strategy is applied, employing pseudo-labeling technology. Data with high confidence levels in the model's predictions of unlabeled data is used as training data. Through iterative optimization, the model is continuously adjusted to improve prediction accuracy on unlabeled data. After each iteration, the validation dataset is used to evaluate model performance, and model parameters are adjusted based on feedback to optimize algorithm efficiency.

[0079] When adapting the semi-supervised learning module to this system, it is necessary to ensure that the dataset covers a wide range of geological scenarios to avoid model overfitting. The model design should focus on its adaptability and stability, and a feedback mechanism should be established to adjust the training strategy and model structure based on validation results, ensuring continuous performance optimization.

[0080] This approach introduces a semi-supervised learning strategy, utilizing only a small amount of well logging annotation data during training. This effectively leverages unlabeled seismic data, reducing reliance on large amounts of labeled data and improving the model's generalization ability and robustness across a wide range of seismic scenarios. The DA-Block mechanism introduced in this study further enhances the model's ability to capture key geological information from seismic data by simultaneously utilizing spatial and channel attention mechanisms. This makes the model more accurate in identifying complex geological structures and variations, particularly demonstrating superior performance in recognizing stratigraphic connections and heterogeneous geological features.

[0081] The core of the system provided by this invention is the DA-TransUNet network architecture, which introduces a dual-attention mechanism on the basis of the TransUNet model. The TransUNet model combines the long-distance information processing capability of Transformer with the efficient local feature extraction capability of U-Net to form the DA-TransUNet model.

[0082] This structure is designed to effectively analyze complex geological structures in seismic data through its multi-scale and multi-region processing capabilities;

[0083] Building upon the TransUNet model, a dual-attention mechanism is introduced, which includes two attention mechanisms: spatial attention and channel attention, enhancing the model's perception capabilities at different scales.

[0084] Specifically, introducing a dual attention mechanism requires precise hyperparameter tuning and a comprehensive fusion strategy to ensure that the output information of spatial and channel attention can be fused in a balanced manner, thereby optimizing model efficiency and performance.

[0085] The dual attention mechanism performs more comprehensive feature capture on the basis of feature extraction from convolutional neural networks. The spatial attention mechanism generates a spatial weight map, which enables the model to focus on key parts of the image. The channel attention mechanism analyzes the global information of each channel and assigns different weights to different channels, thereby highlighting the feature channels that contribute the most to the final task.

[0086] Specifically, the specific algorithm flow of the DA-TransUNet model includes:

[0087] Step S1: Input data preprocessing:

[0088] Select suitable input data from the earthquake dataset and perform necessary preprocessing steps, such as format conversion and normalization, to ensure that the data is suitable for deep learning model processing.

[0089] To address the problem of seismic wave impedance inversion, a maximum-minimum normalization method is used, as shown in the following formula:

[0090] y i ′ =(y i -min(y))÷(max(y)-min(y))

[0091] Among them, y i ′ The data is normalized; y iThe original data is represented by ; max(y) represents the largest value in the original data; and min(y) represents the smallest value in the original data. This normalization process ensures that all data values ​​are within the range of 0 to 1, which is helpful for subsequent data processing and analysis.

[0092] To address the characteristics of the DA-TransUNet model, one-dimensional sequence data is converted into two-dimensional image data through a reshape operation. This meets the DA-TransUNet model's requirement for processing two-dimensional data, thereby effectively utilizing the model's spatial processing capabilities and improving the accuracy of wave impedance inversion.

[0093] Step S2: Encoder stage:

[0094] That is to say, in step S2:

[0095] Step S2.0: Based on the dataset, construct and train the DA-TransUNet model;

[0096] Step S2.1: Extract features from the seismic images in the dataset using the convolutional neural network of the DA-TransUNet model to obtain local features;

[0097] Step S2.2: Use DA-Block to refine the output of the convolutional neural network, i.e., the local features, to obtain skip connection features;

[0098] Step S2.3: Based on the skip connection features, extract global features through the Transformer network.

[0099] A convolutional neural network (CNN) is used to perform preliminary feature extraction on the input seismic image; the CNN includes: convolutional layers, pooling layers, and activation layers; the activation layers include nonlinear activation functions.

[0100] Specifically, convolutional neural networks automatically learn useful and complex features from data through multiple convolutional layers, pooling layers, and non-linear activation functions. Specifically, multiple convolutional layers generate feature maps representing feature responses at different locations in the input data by sliding convolutional kernels across the dataset; based on these feature maps, the ReLU activation function captures non-linear relationships; and based on these non-linear relationships, pooling layers are used to reduce the dimensionality of the feature maps, thereby improving the network's insensitivity to input variations.

[0101] The introduction of DA-Block further refines the features of the CNN output, enhancing the model's ability to capture key features of geological structures. DA-Block consists of a spatial attention mechanism and a channel attention mechanism, the functions of which are described below:

[0102] The spatial attention mechanism utilizes attention maps generated by lightweight convolutional networks. The DA-TransUNet model can focus on key regions in an image, thereby improving the accuracy and efficiency of resolving complex geological structures.

[0103] The channel attention mechanism uses fully connected layers to generate channel weights, emphasizing the most critical feature channels and optimizing feature representation. This helps the model more accurately distinguish the physical properties of the geological layers. A fully connected layer is one where each neuron is connected to all neurons in the previous layer. Based on global statistical information of each channel, it learns to determine which channels contribute the most to the output, thereby generating channel weights to emphasize or suppress specific feature channels.

[0104] DA-Block first processes the feature map through convolutional layers, then feeds the processed feature map into the spatial attention module and the channel attention module respectively, adds the outputs of the two modules, and then passes it through a convolutional layer to obtain the final output, thereby achieving further refinement of the features.

[0105] The features extracted by the CNN after further processing by DA-Block are fed into the Transformer network, which leverages its ability to handle long-range dependencies to extract more complex global features.

[0106] CNNs typically extract local features focusing on specific regions or aspects of data, such as edges, textures, and local shapes. In contrast, Transformer networks utilize their self-attention mechanism to process input features, allowing the DA-TransUNet model to learn the dependencies between each element and all other elements throughout the entire input sequence. This global feature extraction helps in understanding long-range patterns and relationships within data, such as the continuity and distribution characteristics of seismic records.

[0107] When extracting global features, the self-attention mechanism of the Transformer network and the DA-TransUNet model can simultaneously focus on all parts of the input data, helping to understand long-distance dependencies between elements; the positional encoding mechanism provides the DA-TransUNet model with positional information of each element in the sequence; layer normalization and residual connections work together to accelerate the training process and improve the stability of the DA-TransUNet model by directly forwarding the input and standardizing the output of each layer.

[0108] Step S3: Decoder and Feature Fusion:

[0109] The skip connection features, refined by DA-Block and retained from the encoder stage, are fused with features from the decoder. This step helps preserve important local feature information while integrating global context. Then, a CNN is used for progressive upsampling to gradually restore the feature map to the same resolution as the input image.

[0110] The skip connections, by directly passing features from lower layers to higher layers within the network, help preserve detailed information that might be lost in deep networks. The specific processing procedure is as follows:

[0111] In the encoder stage, the input data undergoes multiple convolutional and pooling layers to reduce resolution and increase feature depth. Each feature map is further refined using DA-Block, resulting in high spatial resolution features rich in local information. These features are directly passed to the corresponding layer in the decoder via skip connections, ensuring that high-level features captured during encoding are not lost during downsampling. In the decoder stage, the features from the skip connections are fused with the features from the current layer of the decoder. This fusion is achieved through feature concatenation, which helps combine the details captured by the encoder with the contextual information gradually recovered by the decoder.

[0112] Step S4: Final Output Construction:

[0113] After multiple layers of feature processing and fusion, the final image prediction result is constructed.

[0114] The present invention also provides a seismic impedance inversion system based on deep learning. The seismic impedance inversion system based on deep learning can be implemented by executing the process steps of the seismic impedance inversion method based on deep learning. That is, those skilled in the art can understand the seismic impedance inversion method based on deep learning as a preferred embodiment of the seismic impedance inversion system based on deep learning.

[0115] A seismic impedance inversion system based on deep learning according to the present invention includes:

[0116] Module M1: Constructs and preprocesses the dataset;

[0117] Module M2: Based on the dataset, construct and train the DA-TransUNet model; extract features from the input seismic image using the convolutional neural network of the DA-TransUNet model, extract local features, and generate skip connection features; based on the local features, extract global features using a Transformer network;

[0118] Module M3: Based on global features, it performs upsampling processing through the convolutional neural network of the DA-TransUNet model, fusing the features of skip connections with the features of the current layer of the decoder to obtain a fused result as the upsampling result;

[0119] Module M4: Based on the fusion results, construct the final seismic image prediction results.

[0120] The module M1 is the data preprocessing module;

[0121] In module M2, the DA-TransUNet model is trained, specifically through a semi-supervised learning module;

[0122] Preferably, in module M1: the preprocessing includes: normalization, data augmentation, and format conversion;

[0123] The normalization process refers to the use of maximum and minimum normalization to process seismic wave impedance inversion, and the mathematical expression is:

[0124] y i ′ =(y i -min(y))÷(max(y)-min(y))

[0125] Among them, y i ′ The data is normalized; y i The original data is represented by y; max(y) is the largest value in the original data; min(y) is the smallest value in the original data; and during the normalization process, all data values ​​are within the range of 0 to 1.

[0126] The dataset includes: labeled data, unlabeled data, and validation data.

[0127] Preferably, in module M2:

[0128] Module M2.0: Constructs and trains the DA-TransUNet model based on the dataset;

[0129] Module M2.1: Extracts features from seismic images in the dataset using a convolutional neural network based on the DA-TransUNet model to obtain local features;

[0130] Module M2.2: Through DA-Block, the output of the convolutional neural network, i.e. the local features, is refined to obtain skip connection features;

[0131] Module M2.3: Based on skip connection features, global features are extracted through a Transformer network.

[0132] The module M2.0 is the semi-supervised learning module;

[0133] Preferably, the dataset includes a training set and a validation set;

[0134] The ratio of the training set to the validation set is three to one.

[0135] The validation set includes labeled data;

[0136] The training set is used to train the DA-TransUNet model;

[0137] The ratio of labeled to unlabeled data in the training set is adjusted according to the type of dataset.

[0138] Preferably, the DA-TransUNet model provides positional information for each element in the dataset sequence;

[0139] The DA-Block is composed of a spatial attention mechanism and a channel attention mechanism.

[0140] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0141] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A seismic wave impedance inversion method based on deep learning, characterized in that, include: Step S1: Construct and preprocess the dataset; Step S2: Based on the dataset, construct and train the DA-TransUNet model; extract features from the input seismic image using the convolutional neural network of the DA-TransUNet model, extract local features, and generate skip connection features; based on the local features, extract global features using a Transformer network; Step S3: Based on global features, upsampling is performed through the convolutional neural network of the DA-TransUNet model, fusing the features of the skip connections with the features of the current layer of the decoder to obtain the fused result as the upsampling result; Step S4: Based on the fusion results, construct the final seismic image prediction results; In step S2: Step S2.0: Based on the dataset, construct and train the DA-TransUNet model; Step S2.1: Extract features from the seismic images in the dataset using the convolutional neural network of the DA-TransUNet model to obtain local features; Step S2.2: Use DA-Block to refine the output of the convolutional neural network, i.e., the local features, to obtain skip connection features; Step S2.3: Based on the skip connection features, extract global features through the Transformer network; The DA-TransUNet model provides positional information for each element in the dataset sequence; the DA-Block is composed of a spatial attention mechanism and a channel attention mechanism.

2. The seismic wave impedance inversion method based on deep learning according to claim 1, characterized in that, In step S1: the preprocessing includes: normalization, data augmentation, and format conversion; The normalization process refers to the use of maximum and minimum normalization to process seismic wave impedance inversion, and the mathematical expression is: in, Normalized data This is the original data; The largest value in the original data; This is the smallest value in the original data; and during the normalization process, all data values ​​are within the range of 0 to 1. The dataset includes: labeled data, unlabeled data, and validation data.

3. The seismic wave impedance inversion method based on deep learning according to claim 1, characterized in that, The dataset includes a training set and a validation set; The ratio of the training set to the validation set is three to one. The validation set includes labeled data; The training set is used to train the DA-TransUNet model; The ratio of labeled to unlabeled data in the training set is adjusted according to the type of dataset.

4. A seismic wave impedance inversion system based on deep learning, characterized in that, include: Module M1: Constructs and preprocesses the dataset; Module M2: Based on the dataset, construct and train the DA-TransUNet model; extract features from the input seismic image using the convolutional neural network of the DA-TransUNet model, extract local features, and generate skip connection features; based on the local features, extract global features using a Transformer network; Module M3: Based on global features, it performs upsampling processing through the convolutional neural network of the DA-TransUNet model, fusing the features of skip connections with the features of the current layer of the decoder to obtain a fused result as the upsampling result; Module M4: Based on the fusion results, construct the final seismic image prediction results; In module M2: Module M2.0: Constructs and trains the DA-TransUNet model based on the dataset; Module M2.1: Extracts features from seismic images in the dataset using a convolutional neural network based on the DA-TransUNet model to obtain local features; Module M2.2: Through DA-Block, the output of the convolutional neural network, i.e. the local features, is refined to obtain skip connection features; Module M2.3: Based on skip connection features, extracts global features through a Transformer network; The DA-TransUNet model provides positional information for each element in the dataset sequence; the DA-Block is composed of a spatial attention mechanism and a channel attention mechanism.

5. The deep learning-based seismic impedance inversion system according to claim 4, characterized in that, In module M1: the preprocessing includes: normalization, data augmentation, and format conversion; The normalization process refers to the use of maximum and minimum normalization to process seismic wave impedance inversion, and the mathematical expression is: in, Normalized data This is the original data; The largest value in the original data; This is the smallest value in the original data; and during the normalization process, all data values ​​are within the range of 0 to 1. The dataset includes: labeled data, unlabeled data, and validation data.

6. The seismic wave impedance inversion system based on deep learning according to claim 4, characterized in that, The dataset includes a training set and a validation set; The ratio of the training set to the validation set is three to one. The validation set includes labeled data; The training set is used to train the DA-TransUNet model; The ratio of labeled to unlabeled data in the training set is adjusted according to the type of dataset.

Images