A seismic random noise suppression method based on self-attention convolutional autoencoder

Abstract

The present invention discloses a method for suppressing seismic random noise based on a self-attention convolutional autoencoder, which is characterized in that the feature information contained in the original image is extracted through a self-attention mechanism to generate an attention weight matrix to distinguish the different features between seismic signals and noise, calibrate the key degrees of different regions and channels in the feature map for the network model, and apply different attentions to avoid distortion during signal reconstruction. The model uses a convolutional network as the core of the encoder network and a deconvolutional network as the core of the decoder network. In view of the complexity and particularity of seismic data, multi-scale convolutional modules are used in the encoding and decoding stages to extract seismic data features. The present invention pre-trains the network model with synthetic seismic data, adaptively adjusts the learning rate by the first- and second-order matrices of the gradient through an Adam optimizer to make the loss function converge quickly, and fine-tunes the network parameters with field seismic data, thereby strengthening the robustness of the model, making the data reconstructed by the model closer to the distribution and features of the target data, reducing the loss of effective seismic signals, realizing automatic and efficient suppression of seismic random noise, and improving the accuracy of seismic data processing.

Description

Technical Field

[0001] The present invention belongs to the fields of geophysics and artificial intelligence, and specifically relates to a seismic random noise suppression method based on a self-attention convolutional autoencoder. Background Art

[0002] Suppressing seismic random noise is a significant challenge in seismic exploration. Random noise, a strong interfering wave, appears in seismic exploration, significantly reducing the resolution and signal-to-noise ratio of seismic recordings, directly impacting the accuracy of seismic data processing. While some traditional methods can suppress noise in data, they often suffer from significant information loss and residual noise. Furthermore, most traditional seismic data noise suppression methods are based on fixed transformation basis functions and cannot adaptively process complex seismic data.

[0003] In recent years, deep learning has achieved remarkable results in the field of computer vision. It extracts deep features from data to achieve image classification, object detection, and image denoising. Deep learning methods based on labeled data have been successfully applied to seismic random noise suppression. These methods use the results of traditional noise suppression methods as labeled data to train a neural network, which is then used to suppress noise in actual seismic data. However, these methods produce inaccurate labeled data, which affects the noise suppression performance of the network model. Therefore, developing unsupervised learning methods that do not rely on labeled data is of great significance for seismic random noise suppression.

[0004] The seismic random noise suppression method based on self-attention convolutional autoencoder is an unsupervised learning method. It can learn the difference between effective signals and noise from a large number of data samples, adaptively establish a deep neural network to suppress noise, and directly recover noise-free seismic data from noisy seismic data, eliminating the large amount of work required to label data. It also reduces the loss of effective seismic signals and improves the signal-to-noise ratio of seismic signals while suppressing random noise. Summary of the Invention

[0005] In order to effectively suppress seismic random noise, overcome the interference it causes in seismic exploration, and improve the problem of low precision in seismic data processing, the present invention proposes a method for suppressing seismic random noise based on a self-attention convolutional autoencoder. A convolutional autoencoder network guided by self-attention is constructed. The self-attention mechanism is used to extract the feature information contained in the original image, generate an attention weight matrix to distinguish the different features of seismic signals and noise, calibrate the key degree of different regions and channels in the feature map for the network model, and apply different attentions to avoid distortion during signal reconstruction. Synthetic seismic data is used to pre-train the network. The Adam optimizer is used to adaptively adjust the learning rate of the first- and second-order matrices of the gradient, so that the loss function converges quickly. Field seismic data is used to fine-tune the network parameters, thereby enhancing the robustness of the model, making it closer to the distribution and characteristics of the target data, and improving the performance of suppressing random noise in actual seismic data.

[0006] To achieve the above object, the technical solution of the present invention mainly includes the following steps:

[0007] A. Construct a seismic data slice dataset:

[0008] The original seismic data is segmented into single-shot seismic records, and the header information is removed, leaving only the seismic trace data. The extracted data is sliced into n×n seismic data slices using a convolutional sliding window to obtain a clean seismic data slice set.

[0009] B. Calibrate the feature map using the attention weight matrix:

[0010] (1) Form a linear embedding sequence of noisy seismic data

[0011] Randomly select a batch of clean seismic data slices and add random noise. The noisy seismic data slices are split into small data blocks of m×m and arranged to form a 1×(m×m) linear embedding sequence, which is input into the self-attention module;

[0012] (2) Calculate the self-attention weight matrix

[0013] The self-attention mechanism is used to extract the feature information contained in the original image to distinguish the different features of seismic signals and noise, apply different attentions to avoid distortion during signal reconstruction, and allocate limited resources to important parts. The different seismic signals extracted are distinguished by weighted changes, and the information that is more critical for the noise suppression task is selected from among many representations. Calculate the query matrix Q = XW

[0012] , K , V , V , , Q , Q , K , ,

[0013] 、key matrix K = XW K and value matrix V = XW V , where X is the input data matrix, W Q 、W K 、W Vare three trainable parameter matrices;

[0014] (3) Calibrate the feature map

[0015] The parameter matrices change with the change of the input X matrix through the self-attention weight matrix, and the key degrees of different regions and channels in the calibrated feature map for the network model are utilized to perform targeted feature learning, avoiding wasting resources on redundant information.

[0016] C. Use synthetic seismic data to pre-train the self-attention convolutional autoencoder network:

[0017] (1) Design and construction of the convolutional layer

[0018] Feature extraction is performed on the input data through the convolutional layer. It contains multiple convolutional kernels inside, and each element constituting the convolutional kernel corresponds to a weight coefficient and a bias. Each neuron in the convolutional layer is connected to multiple neurons in a region with a close position in the previous layer. The convolutional kernel sweeps regularly over the input features, performs matrix element multiplication and summation on the input features within the receptive field and superimposes the bias;

[0019] (2) Construct the self-attention convolutional autoencoder network model

[0020] Combining the self-attention convolution and the autoencoder network model, using the method of enhancing the traditional convolution by the self-attention mechanism to dynamically generate filters, perform feature extraction according to the global dependence relationship of the input features, perform encoding and decoding, realize data compression and reconstruction, and enhance the noise suppression ability of the network. Use the convolutional network as the network core of the encoder and the transposed convolutional network as the network core of the decoder. Considering the complexity and particularity of seismic data, multi-scale convolutional modules are used to extract seismic data features in the encoding and decoding stages. The encoder (Encoder) of the self-attention convolutional autoencoder consists of k convolutional layers and g pooling layers, and the decoder (Decoder) consists of k + 2 transposed convolutional layers;

[0021] (3) Model training

[0022] Use the mean square error (MSE) as the distance metric function between the original data vector and the denoised data vector, and use this distance metric function as the loss function. Select Adam as the optimizer, and at the same time use the first-order and second-order matrices of the gradient to adaptively adjust the learning rate, and adjust the self-attention weight matrix and the parameters of the convolutional autoencoder network model until convergence. Input the noisy seismic data vector into the autoencoder to make the output as consistent as possible with the input, and realize the extraction of hidden layer features and parameter learning in an unsupervised form. Based on the backpropagation algorithm and the optimization method, use the input data X itself as the supervision to guide the neural network to try to learn a mapping relationship, so as to obtain a reconstructed output X R. In the seismic denoising task, the output X reconstructed by the autoencoder R If the difference between the original input exceeds the threshold λ, then iterative optimization continues.

[0023] D. Fine-tuning the parameters of the self-attention convolutional autoencoder network using field seismic data:

[0024] (1) Preprocessing of field seismic data

[0025] The field seismic data is segmented into single-shot seismic trace data, and the extracted data is sliced into n×n seismic data slices using a convolutional sliding window to obtain a field seismic data set;

[0026] (2) Parameter fine-tuning

[0027] Transfer the pre-trained model parameters based on synthetic seismic data to the field seismic data noise suppression task. Use the mean squared error (MSE) as the distance metric function between the original data vector and the denoised data vector, and use it as the loss function. Select Adam as the optimizer, and at the same time use the first and second order matrices of the gradient to adaptively adjust the learning rate. Retrain the model using actual field noise data as the target data to fine-tune the model parameters, thereby enhancing the robustness of the model, making it closer to the distribution and characteristics of the target data, and improving the noise suppression performance for actual seismic data.

[0028] The beneficial effects of the present invention are: building a seismic random noise suppression model of a self-attention convolutional autoencoder, adopting a model combining self-attention convolution and an autoencoder network and an adaptive learning rate adjustment based on the Adam optimizer, enhancing the traditional convolution method, dynamically generating filters, extracting features according to the global dependence relationship of input features, performing encoding and decoding, realizing data compression and reconstruction, and enhancing the noise suppression performance of the model. Description of the drawings

[0029] Figure 1 It is a flowchart of the seismic random noise suppression method of the self-attention convolutional autoencoder described in the present invention. Detailed implementation manners

[0030] The following combines Figure 1 to further describe the present invention in detail:

[0031] A. Constructing a seismic data slice data set:

[0032] The original seismic data is segmented into single-shot seismic records and the header information is removed, only retaining the seismic trace data. The extracted data is sliced into 64×64 seismic data slices using a convolutional sliding window to obtain a clean seismic data slice set.

[0033] B. Calibrate the feature map using the attention weight matrix:

[0034] (1) Form a linear embedding sequence of noisy seismic data

[0035] Randomly select batch = 64 slices of clean seismic data and add random noise. Split the noisy seismic data slices into small data blocks of 16×16 and arrange them to form a 1×256 linear embedding sequence, which is input into the self-attention module;

[0036] (2) Calculate the self-attention weight matrix

[0037] Use the self-attention mechanism to extract the feature information contained in the original image to distinguish the different features of seismic signals and noise and apply different attentions, avoid distortion during signal reconstruction, and allocate limited resources to important parts. Differentiate the extracted different seismic signals using weighted variation, select the information that is more critical for the noise suppression task from numerous representations, and calculate the query matrix Q = XW Q 、key matrix K = XW K and value matrix V = XW V , where X is the input data matrix, W Q 、W K 、W V are three trainable parameter matrices;

[0038] (3) Calibrate the feature map

[0039] Through the self-attention weight matrix, the parameter matrix changes with the change of the input X matrix, and uses it to calibrate the key degree of different regions and channels in the feature map for the network model, and conducts targeted feature learning to avoid wasting resources on redundant information.

[0040] C. Pre-train the self-attention convolutional autoencoder network using synthetic seismic data:

[0041] (1) Design and construction of the convolutional layer

[0042] Extract features from the input data through the convolutional layer. It contains multiple convolutional kernels, and each element constituting the convolutional kernel corresponds to a weight coefficient and a bias. Each neuron in the convolutional layer is connected to multiple neurons in the region close to its position in the previous layer. The convolutional kernel sweeps regularly over the input features, performs matrix element multiplication and summation on the input features within the receptive field and superimposes the bias, and the calculation method is as follows:

[0043]

[0044] The summation part in the formula is equivalent to solving a cross-correlation once, where b is the bias, Z l and Zl+1 Denote the convolution input and output of the (l + 1)-th layer, where L l+1 is the size of Z l+1 . Z(i, j) corresponds to the pixel of the feature map, K is the number of channels of the feature map, and f, s, p are the parameters of the convolutional layer, corresponding to the convolutional kernel size, convolutional stride, and padding layer number;

[0045] (2) Construct a self-attention convolutional autoencoder network model

[0046] Combine self-attention convolution and autoencoder network models, use the self-attention mechanism to enhance traditional convolution, dynamically generate filters, extract features according to the global dependence of input features, perform encoding and decoding, realize data compression and reconstruction, and enhance the noise suppression ability of the network. Use a convolutional network as the network core of the encoder and a transposed convolutional network as the network core of the decoder. Considering the complexity and particularity of seismic data, multi-scale convolutional modules are used to extract seismic data features in the encoding and decoding stages. The encoder (Encoder) of the self-attention convolutional autoencoder consists of 5 convolutional layers and 2 pooling layers, and the decoder (Decoder) consists of 7 transposed convolutional layers;

[0047] (3) Model training

[0048] Use the mean square error (MSE) as the distance metric function between the original data vector and the denoised data vector, and use this distance metric function as the loss function. Select Adam as the optimizer, and at the same time use the first-order and second-order matrices of the gradient to adaptively adjust the learning rate, and adjust the self-attention weight matrix and the parameters of the convolutional autoencoder network model until convergence. Input the noisy seismic data vector into the autoencoder to make the output as consistent as possible with the input, and realize the extraction of hidden layer features and parameter learning in an unsupervised form. Based on the backpropagation algorithm and optimization methods, use the input data X itself as the supervision to guide the neural network to try to learn a mapping relationship, so as to obtain a reconstructed output X R .

[0049] The data passes through two main parts: the Encoder and the Decoder. The encoder encodes the high-dimensional input X into a low-dimensional hidden variable h to force the neural network to learn the most informative features; the decoder restores the hidden variable h in the hidden layer to the initial dimension. In order for the output of the decoder to perfectly or approximately recover the original input, that is, X R ≈ X.

[0050] The encoding process of the original data from the input layer to the hidden layer is as follows:

[0051] h = g θ1 (x) = σ(W1x + b1)

[0052] Decoding process from the hidden layer to the output layer:

[0053]

[0054] The optimization objective function of the algorithm is:

[0055] MinimizeLoss = dist(X, X R )

[0056] where dist is the distance metric function between the two, calculated using the mean squared error (MSE).

[0057] The smaller the value of MSE, the closer the data reconstructed by the network is to the original clean data, that is, the denoised data has a higher peak signal-to-noise ratio compared to the noisy clean data. In the seismic noise suppression task, if the difference between the output X reconstructed by the autoencoder R and the original input exceeds the threshold λ = 0.01, then iterative optimization continues.

[0058] D. Fine-tuning the parameters of the self-attention convolutional autoencoder network using field seismic data:

[0059] (1) Preprocessing of field seismic data

[0060] The field seismic data is segmented into single-shot seismic trace data, and the extracted data is sliced into 64×64 seismic data slices using a convolutional sliding window to obtain a field seismic data set;

[0061] (2) Parameter fine-tuning

[0062] The pre-trained model parameters based on synthetic seismic data are migrated to the field seismic data noise suppression task. The mean squared error (MSE) is used as the distance metric function between the original data vector and the denoised data vector and as the loss function. Adam is selected as the optimizer, and the learning rate is adaptively adjusted using the first and second order matrices of the gradient. The model is retrained using a small amount of actual field noise data as the target data to fine-tune the model parameters, thereby enhancing the robustness of the model, making it closer to the distribution and characteristics of the target data, and improving the noise suppression performance for actual seismic data.

[0063] The above are only the preferred embodiments of the present invention. Any person skilled in the art may modify or change the above-described technical solutions into equivalent examples with equivalent changes. Any simple modification, change, or modification made to the above embodiments based on the technical solutions of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.

Claims

1. A seismic random noise suppression method based on a self-attention convolutional autoencoder, characterized in that Including the following steps: A. Calibrate the feature map using the attention weight matrix: Randomly select clean seismic data slices and add random noise. Split the noisy seismic data slices into small data blocks and arrange them to form a linear embedding sequence, which is input into the self-attention module to obtain the attention weight matrix to distinguish the different features of seismic signals and noise signals, so as to calibrate the key degrees of different regions and channels in the feature map for the network model; B. Pretrain the self-attention convolutional autoencoder network using synthetic seismic data: Take the convolutional network as the core of the encoder network and the deconvolutional network as the core of the decoder network to construct a multi-scale convolutional autoencoder module. Input the calibrated feature map into the convolutional autoencoder module, and use the Adam optimizer to adaptively adjust the learning rate of the first-order and second-order matrices of the gradient, so that the loss function gradually converges until the difference between the values of the loss function in two consecutive times is less than the threshold, and obtain the pre-trained model based on synthetic seismic data; C. Fine-tune the parameters of the self-attention convolutional autoencoder network using field seismic data: Use the actual field seismic data as the target data to fine-tune the network parameters, and use the test set to test the model. Calculate the peak signal-to-noise ratio of the test data to specifically evaluate the noise suppression performance of the model, and use the final model to process seismic data to achieve automatic and efficient seismic random noise suppression.

Images