A seismic data classification method based on a time attention mechanism
Through the seismic data classification method based on the time attention mechanism, the time period and global characteristics of seismic data are extracted using parallel branch structures and convolutional neural networks, the problem of improper resource allocation in the existing model is solved and the classification accuracy of seismic data is improved.
Patent Information
- Application Number
- CN202310272897.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-03-17
- Publication Date
- 2025-07-08
- Estimated Expiration
- 2043-03-17
AI Technical Summary
The existing deep learning models fail to effectively utilize the correlation between features when processing seismic data, resulting in improper allocation of computing resources and affecting the classification accuracy of seismic data.
The seismic data classification method based on the time attention mechanism is adopted, and the time period characteristics and global characteristics of the seismic data are extracted through parallel branch structures, and the feature processing capabilities are optimized using convolutional neural networks and residual networks, and the time attention mechanism is combined with the time attention mechanism to weight the time period characteristics to improve the feature attention.
The classification accuracy of earthquake data is improved, the correlation between data features is fully utilized, and the feature processing capability is optimized.
Smart Images
Figure CN116068651B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of seismic detection, and particularly to a seismic data classification method based on a temporal attention mechanism. Background Art
[0002] As an evaluation method applied to fracture exploration in low-permeability oilfields, hydraulic fracturing technology has played an extremely important role in oil and gas exploration and is widely used in the monitoring of oilfield production increase and the exploitation of new energy such as shale gas. By deploying seismographs in wells and on the ground, a large amount of seismic data generated by hydraulic fracturing can be recorded. Therefore, in the case of limited computing resources, when processing and analyzing data, it is necessary to extract as efficiently as possible the useful information contained in these data. Deep neural networks have the advantages of automatic feature extraction and self-learning and have achieved success in many classification problems. Most of the existing deep learning models are dedicated to solving related tasks of seismic data classification, but the same processing strategy is adopted for all features extracted from seismic data, so that the correlation between features is not well utilized.
[0003] The temporal attention mechanism is a resource allocation scheme that, in the case of limited computing resources, weakens the attention to useless information, allocates computing resources to more important tasks, and simultaneously solves the problem of information overload. Introducing the temporal attention mechanism when using a deep learning model to process seismic data enables the model to focus on information that is more critical to the current task, reduces the attention to useless information, and even filters out useless information, and makes full use of the correlation between features to solve the seismic data classification problem. Therefore, the present invention proposes a seismic data classification method based on a temporal attention mechanism to solve the above technical problems. Summary of the Invention
[0004] The purpose of the present invention is to propose a seismic data classification method based on a temporal attention mechanism for the above technical problems, optimize the ability to process seismic data features, and improve the classification accuracy of seismic data.
[0005] To achieve the above purpose, the present invention provides the following solution:
[0006] A seismic data classification method based on a temporal attention mechanism, comprising:
[0007] Collect seismic data, process the seismic data to generate a spectrogram, wherein the seismic data is time series data;
[0008] Extract features from the spectrogram to obtain time segment features and global features with weights;
[0009] Obtain the classification result of the seismic data according to the time segment features and global features with weights.
[0010] Further, processing the seismic data includes:
[0011] Converting the time series data into frequency domain data by using frequency domain transformation and saving the frequency domain data as a spectrogram, where the frequency domain transformation includes fast Fourier transform and short-time Fourier transform.
[0012] Further, obtaining the time period features and global features with weights includes:
[0013] Extracting the shallow features of the spectrogram and respectively extracting the time period features and global features of the spectrogram from the shallow features of the spectrogram by using a parallel branch structure.
[0014] Further, extracting the shallow features of the spectrogram includes:
[0015] Constructing a shallow feature extraction module based on a convolutional neural network, inputting the spectrogram into the shallow feature extraction module to extract the shallow features of the spectrogram, where the shallow feature extraction module includes a two-dimensional convolutional layer, a batch normalization layer, and an ELU activation layer, the two-dimensional convolutional layers are connected in the form of residuals, and the batch normalization layer and the ELU activation layer are connected after the two-dimensional convolutional layer.
[0016] Further, the parallel branch structure includes an upper branch and a lower branch;
[0017] The upper branch is used to extract the deep features of the spectrogram from the shallow features of the spectrogram and extract the time period features of the spectrogram based on the extracted deep features;
[0018] The lower branch is used to extract the global features of the spectrogram from the shallow features of the spectrogram.
[0019] Further, extracting the deep features of the spectrogram includes:
[0020] Constructing a deep feature extraction module based on a residual network, inputting the shallow features of the spectrogram into the deep feature extraction module to extract the deep features of the spectrogram, where the deep feature extraction module includes residual blocks, the residual blocks include a direct mapping unit and a residual unit, and the output results of the direct mapping unit and the residual unit are summed after being obtained, the direct mapping unit includes a two-dimensional convolutional layer, a batch normalization layer, an ELU activation layer, and a Dropout layer, and the residual unit includes an ELU activation layer and a max pooling layer.
[0021] Further, extracting the time period features of the spectrogram includes:
[0022] Construct a time-based attention mechanism to build a time period feature extraction module, and input the deep features of the spectrogram into the time period feature extraction module to extract the time period features of the spectrogram. Among them, the time period feature extraction module includes a weight calculation unit and a feature extraction unit, and perform feature splicing after obtaining the results output by the parallel calculation of the weight calculation unit and the feature extraction unit;
[0023] The weight calculation unit includes a global average pooling layer, a fully connected layer, a Mish activation layer, and a Sigmoid function. Based on the global average pooling layer, perform feature channel compression on the deep features, generate the weights of the feature channels through the fully connected layer, Mish activation layer, and Sigmoid function, and weight the weights to the deep features; the feature extraction unit includes a two-dimensional convolutional layer and an ELU activation layer, and perform further feature extraction on the deep features.
[0024] Further, extracting the global features of the spectrogram includes:
[0025] Construct a global feature extraction module based on a convolutional neural network, and input the shallow features of the spectrogram into the global feature extraction module to extract the global features of the spectrogram. Among them, the global feature extraction module includes a two-dimensional convolutional layer, a global average pooling layer, a fully connected layer, and an ELU activation layer. Perform feature dimensionality reduction and further deep feature extraction on the shallow features through the two-dimensional convolutional layer and the global average pooling layer, use the fully connected layer to map the further deep features extracted, obtain the global features, and connect the ELU activation layer after the fully connected layer.
[0026] Further, obtaining the classification result of the seismic data includes:
[0027] Fuse the time period features and global features of the extracted spectrogram, and input them into the Softmax function to obtain the classification result of the seismic data.
[0028] The beneficial effects of the present invention are:
[0029] The present invention extracts the shallow features of seismic data through a convolutional neural network, and then adopts a parallel branch structure. In the upper branch, a residual network is applied to extract deep features, and a time-based attention mechanism is used for the deep features to obtain time period features with weights; in the lower branch, a convolutional neural network is used to extract global features from the shallow features. Fusing and judging the extracted deep features can improve the attention of the deep learning model to the global attributes of seismic data, make full use of the features of the original data, not only optimize the processing ability of seismic data features, but also improve the classification accuracy of the two types of seismic data. Description of the Drawings
[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.
[0031] Figure 1 Flowchart of a seismic data classification method based on a time attention mechanism according to an embodiment of the present invention;
[0032] Figure 2 Block diagram of a deep feature extraction module according to an embodiment of the present invention;
[0033] Figure 3 Block diagram of a time segment feature extraction module according to an embodiment of the present invention;
[0034] Figure 4 Block diagram of a global feature extraction module according to an embodiment of the present invention. Detailed implementation manners
[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.
[0036] It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer executable instructions. And although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.
[0037] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the drawings and specific implementation manners.
[0038] This embodiment provides a seismic data classification method based on a time attention mechanism, as Figure 1 shown, and the specific process includes:
[0039] S1. Collect seismic data and perform frequency domain transformation on the seismic data
[0040] Detect signals through a seismograph to obtain time series data, and then apply frequency domain transformation methods such as fast Fourier transform and short-time Fourier transform to convert the time series data into frequency domain data and save it in the form of a spectrogram.
[0041] S2. Extract shallow features
[0042] First, a convolutional neural network is used to extract shallow features from the input spectrogram. The shallow feature extraction module consists of 15 two-dimensional convolutional layers, including 1 two-dimensional convolution with a convolution kernel of 7*7, 3 two-dimensional convolutions with a convolution kernel of 5*5, and 11 two-dimensional convolutions with a convolution kernel of 3*3. Different convolution kernels are connected in the form of residuals, and a batch normalization layer and an ELU activation layer are connected after each convolutional layer. The application of the batch normalization layer can achieve a certain regularization effect, improve the model convergence speed, and prevent the model from overfitting.
[0043] S3. Parallel branch structure is adopted to extract time segment features and global features
[0044] S3.1. Extract time segment features in the upper branch
[0045] The upper branch consists of a deep feature extraction module and a time segment feature extraction module, and the two modules are connected in series. In the upper branch, the shallow features are first used to extract deep features through a residual network, and then the time-based attention mechanism is used for the extracted deep features to obtain time segment features with importance weights.
[0046] As Figure 2 shown, in the deep feature extraction module, deep features are extracted through a residual network. The residual network consists of two types of residual blocks. Each type of residual block is divided into a direct mapping unit and a residual unit. After obtaining the output results of the direct mapping unit and the residual unit, a summation operation is performed.
[0047] Among them, the residual units of the two types of residual blocks are both composed of an ELU activation layer and a max pooling layer. The direct mapping unit of the first type of residual block is composed of two two-dimensional convolutional layers with a convolution kernel of 3*3, a batch normalization layer, an ELU activation layer, and a Dropout layer for the first deep feature extraction. The input shallow features pass through an ELU activation layer and a max pooling layer and then are feature concatenated with the deep features extracted for the first time. The first type of residual block is stacked twice; the direct mapping unit of the second type of residual block is composed of two two-dimensional convolutional layers with a convolution kernel of 3*3, two batch normalization layers, two ELU activation layers, and a Dropout layer for the second deep feature extraction. The input shallow features pass through two ELU activation layers and a max pooling layer and then are feature concatenated with the deep features extracted for the first time. The second type of residual block is stacked seven times.
[0048] As Figure 3As shown in the figure, in the time segment feature extraction module, a time-based attention mechanism is used to obtain time segment features with importance weights from deep features. The time segment feature extraction module is serially stacked six times, divided into a weight calculation branch and a feature extraction branch. The output results of the weight calculation branch and the feature extraction branch are concatenated to obtain time segment features with weights.
[0049] Among them, the weight calculation branch further extracts time segment features from the deep features extracted by the residual structure. It consists of a global average pooling layer, two fully connected layers, a Mish activation layer, and a Sigmoid function. The stacking order is a global average pooling layer, a fully connected layer, a Mish activation layer, a fully connected layer, and a Sigmoid function. This unit first compresses the above deep features along the spatial dimension through the global average pooling layer, compressing each two-dimensional feature channel into a real number, and the output dimension matches the number of input feature channels; then uses two fully connected layers, a Mish activation layer, and a sigmoid function to generate respective weights for each feature channel, that is, the correlation between feature channels; finally, all the calculated weights are weighted to the previous features channel by channel through multiplication to complete the recalibration of the original features in the channel dimension.
[0050] The feature extraction branch further extracts features from the deep features extracted by the residual network. The feature extraction unit consists of three two-dimensional convolutional layers and two ELU activation layers, stacked in the order of a two-dimensional convolutional layer with a kernel size of 3*3, an ELU activation layer, two two-dimensional convolutional layers with a kernel size of 3*3, and an ELU activation layer.
[0051] S3.2. Extract global features in the lower branch
[0052] In the lower branch, the shallow features are input for global feature extraction. As Figure 4 shown, the global feature extraction module first performs a feature dimensionality reduction on the shallow features using a two-dimensional convolutional layer with a kernel size of 3*3 and two two-dimensional global average pooling layers, stacked in the order of a global average pooling layer, a two-dimensional convolutional layer with a kernel size of 3*3, and a global average pooling layer; then uses four two-dimensional convolutional layers with a kernel size of 3*3 and two global average pooling layers to extract deeper features, stacked in the order of three two-dimensional convolutional layers with a kernel size of 3*3, a global average pooling layer, a two-dimensional convolutional layer with a kernel size of 3*3, and a global average pooling layer, serially stacked 3 times; finally, uses five fully connected layers to map the extracted deeper features to a new feature space to obtain global features, and an ELU activation layer is connected after each fully connected layer.
[0053] S4. Feature fusion is performed on the extracted segment features and global features, and then the result is input into the Softmax function to obtain the classification result of the seismic data.
[0054] In the present invention, shallow features of seismic data are extracted through a convolutional neural network, and then a parallel branch structure is adopted: in the upper branch, a residual network is applied to extract deep features, and a time-based attention mechanism is used for the deep features to obtain segment features with weights; in the lower branch, a convolutional neural network is used to extract global features from the shallow features. By performing fusion determination on the extracted deep features, the attention of the deep learning model to the global attributes of seismic data can be improved, and the features of the original data are fully utilized, which not only optimizes the processing ability of seismic data features, but also improves the classification accuracy of the two types of seismic data.
[0055] The embodiments described above are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.
Claims
1. A seismic data classification method based on a time attention mechanism, characterized in that, Including: Collecting seismic data, processing the seismic data to generate a spectrogram, wherein the seismic data is time series data; Extracting features from the spectrogram to obtain time segment features and global features with weights; Obtaining a classification result of the seismic data according to the time segment features and global features with weights; Wherein, obtaining the time segment features and global features with weights includes: Extracting shallow features of the spectrogram, and respectively extracting the time segment features and global features of the spectrogram from the shallow features of the spectrogram by using a parallel branch structure; The parallel branch structure includes an upper branch and a lower branch; The upper branch is used to extract deep features of the spectrogram from the shallow features of the spectrogram, and extract the time segment features of the spectrogram based on the extracted deep features; The lower branch is used to extract the global features of the spectrogram from the shallow features of the spectrogram; Extracting the time segment features of the spectrogram includes: Constructing a time segment feature extraction module based on an attention mechanism of time, inputting the deep features of the spectrogram into the time segment feature extraction module to extract the time segment features of the spectrogram, wherein the time segment feature extraction module includes a weight calculation unit and a feature extraction unit, and performing feature splicing after obtaining the results output by the parallel calculation of the weight calculation unit and the feature extraction unit; The weight calculation unit includes a global average pooling layer, a fully connected layer, a Mish activation layer and a Sigmoid function, compressing the feature channels of the deep features based on the global average pooling layer, generating weights of the feature channels through the fully connected layer, the Mish activation layer and the Sigmoid function, and weighting the weights to the deep features; the feature extraction unit includes a two-dimensional convolutional layer and an ELU activation layer, and further extracting features from the deep features.
2. The seismic data classification method based on the temporal attention mechanism according to claim 1, wherein Processing the seismic data includes: Converting the time series data into frequency domain data by using a frequency domain transformation, and saving the frequency domain data as a spectrogram, wherein the frequency domain transformation includes a fast Fourier transform and a short-time Fourier transform.
3. The seismic data classification method based on a time attention mechanism according to claim 1, characterized in that Extracting the shallow features of the spectrogram includes: Constructing a shallow feature extraction module based on a convolutional neural network, inputting the spectrogram into the shallow feature extraction module to extract the shallow features of the spectrogram, wherein the shallow feature extraction module includes a two-dimensional convolutional layer, a batch normalization layer and an ELU activation layer, the two-dimensional convolutional layers are connected in a residual form, and the batch normalization layer and the ELU activation layer are connected after the two-dimensional convolutional layer.
4. The seismic data classification method based on a time attention mechanism according to claim 1, characterized in that Extracting the deep features of the spectrogram includes: Construct a deep feature extraction module based on a residual network, and input the shallow features of the spectrogram into the deep feature extraction module to extract the deep features of the spectrogram. Among them, the deep feature extraction module includes residual blocks, and each residual block includes a direct mapping unit and a residual unit. After obtaining the output results of the direct mapping unit and the residual unit, sum them. The direct mapping unit includes a two-dimensional convolutional layer, a batch normalization layer, an ELU activation layer, and a Dropout layer. The residual unit includes an ELU activation layer and a max pooling layer.
5. The seismic data classification method based on a time attention mechanism according to claim 1, wherein Extracting the global features of the spectrogram includes: Construct a global feature extraction module based on a convolutional neural network, and input the shallow features of the spectrogram into the global feature extraction module to extract the global features of the spectrogram. Among them, the global feature extraction module includes a two-dimensional convolutional layer, a global average pooling layer, a fully connected layer, and an ELU activation layer. The two-dimensional convolutional layer and the global average pooling layer are used to perform feature dimensionality reduction and further deep feature extraction on the shallow features. The fully connected layer is used to map the further deep features extracted, and the global features are obtained. And the ELU activation layer is connected after the fully connected layer.
6. The seismic data classification method based on a temporal attention mechanism according to claim 1, wherein Obtaining the classification result of the seismic data includes: Fuse the extracted temporal features and global features of the spectrogram, and input them into the Softmax function to obtain the classification result of the seismic data.