Seismic data surface wave suppression method based on improved convolutional auto-encoding network

Abstract

The application provides a seismic data surface wave suppression method based on an improved convolutional auto-encoding network, and comprises the following steps: S1, stripping the trace head information of seismic data containing surface waves and seismic data not containing surface waves, separating the data according to shots, setting the data size, and obtaining N pairs of seismic data samples; S2, corresponding matching is performed on the obtained N pairs of seismic data samples to obtain the data samples required by the improved convolutional auto-encoding network model, wherein n1 pairs of data are randomly selected as a verification set, and the remaining n2 pairs of data are used as a training set for experiments; S3, an auto-encoding network is selected as a basic network framework, the seismic data containing surface waves are input into the auto-encoding network, and the seismic data without surface waves are output, so that the network model mapping from the seismic data containing surface waves to the seismic data without surface waves is obtained; and S4, after the basic network model is built, an optimization operation is performed on the network model. The adaptive intelligent denoising of the surface wave noise of seismic data is realized.

Description

Technical Field

[0001] This invention belongs to the field of seismic exploration technology and relates to a coherent noise cancellation method, particularly a seismic data surface wave suppression method based on an improved convolutional autoencoder network. Background Technology

[0002] Due to the complex environment in actual seismic exploration, the acquired seismic data contains a large amount of noise, which seriously affects the normal progress of subsequent seismic data processing and interpretation. Surface waves, as a typical coherent noise in seismic exploration, are low-frequency, low-velocity, high-amplitude regular interference waves that are widely present in seismic records and exhibit a distinct broom-like divergent distribution, severely impacting the accurate analysis of seismic data by interpreters. Therefore, suppressing surface wave noise in seismic noise has always been a key means to improve the signal-to-noise ratio and resolution of seismic data.

[0003] Based on the characteristics of surface waves—high energy, low propagation speed, and slow attenuation—common traditional surface wave suppression methods include frequency domain filtering, wavelet transform filtering, region filtering, anomalous amplitude attenuation, and three-dimensional FFK filtering. Frequency domain filtering and wavelet transform filtering consider the low frequency of surface waves, but their suppression effect is not ideal when there is frequency overlap between the effective signal and surface wave interference. Region filtering is a bandpass filter targeting a specific area, utilizing the low frequency characteristic of surface waves to perform high-pass filtering on the surface wave region to suppress them. The advantage of this method is its simplicity and practicality; it only processes signals within the surface wave region and has no impact on data outside the region. The disadvantage is that it only considers frequency, filtering the effective signal within the region while removing the surface wave, thus weakening the intensity of the effective information. Anomalous amplitude attenuation is essentially a median filter. Within a certain frequency band, channels whose amplitude deviates from a specified threshold value will be attenuated or interpolated based on adjacent channels. The advantages are that amplitude statistics can be performed across different frequency bands, attenuating shallow noise with amplitudes close to the effective value. The disadvantage is that surface waves remain in deeper layers, affecting the accuracy of subsequent energy compensation coefficient calculations. Additionally, three-dimensional FFK filtering is based on a cross-shaped arrangement, with two-dimensional FK filters spatially expanded from a fan shape to a cone shape. This method achieves good surface wave suppression from shallow to deep; however, it processes the entire gather data, including signals outside the surface wave region, leading to attenuation of the effective signal. Traditional surface wave suppression methods typically transform the signal to create a significant difference between the effective signal and the surface wave in a certain dimension, then separate the effective signal from the surface wave noise signal by setting a reasonable threshold. However, aliasing between the effective signal and the surface wave is unavoidable in some areas, causing traditional surface wave suppression methods to remove some effective waves along with the surface waves, resulting in a significant loss of effective information in the seismic data. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned problems in existing technologies by proposing a deep neural network method based on actual seismic data as the training set and an improved convolutional autoencoder network as the model framework. This method trains the seismic data to obtain a model mapping from seismic data containing surface waves to seismic data without surface waves, thereby achieving an adaptive and intelligent denoising method for seismic data surface wave suppression based on an improved convolutional autoencoder network.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A seismic data surface wave suppression method based on an improved convolutional autoencoder network includes the following steps:

[0007] S1. Separate the head information from the seismic data containing surface waves and the seismic data without surface waves, and separate the data by shot. Set the data size to obtain N pairs of seismic data samples.

[0008] S2. Match the obtained N pairs of earthquake data samples to obtain the data samples of the improved convolutional autoencoder network model. Randomly select n1 pairs of data as the validation set and the remaining n2 pairs of data as the training set for experiments.

[0009] S3. First, select an autoencoder network as the basic network framework, input seismic data containing surface waves into the autoencoder network, and output seismic data without surface waves to obtain the network model mapping from seismic data containing surface waves to seismic data without surface waves.

[0010] S4. After the basic network model is built, proceed with the optimization of the network model:

[0011] 1) Improve the basic network model by using convolutional layers:

[0012] A convolutional layer and an activation function form an encoding unit, and a deconvolutional layer and an activation function form a decoding unit. Several encoding units and corresponding decoding units form a complete improved convolutional autoencoder network.

[0013] 2) Improve the activation layers of the convolutional autoencoder network model using activation functions to ensure non-linear data transfer between different network layers:

[0014] The improved convolutional autoencoder network uses the Swish function, which is a smooth and non-monotonic function with no upper bound but a lower bound, and whose derivative is always greater than 0.

[0015] 3) After improving the network layers of the network model, the loss function of the network model is then improved:

[0016] Using the binary cross-entropy loss function in an improved convolutional autoencoder network:

[0017]

[0018]

[0019] In formula (1), x i y represents the predicted data obtained after network training. i The mean of the sum of squares of the errors between corresponding points in the predicted data and the original data is used to judge the network training effect; in formula (2), x represents the original data. i p(x) represents the label value obtained from network training. i ) represents x i The probability of the label value appearing, q(x) i () represents x under real data i The probability of a label value appearing; the binary cross-entropy loss function represents the distance between two probability distributions. The smaller the binary cross-entropy value, the closer the two probability distributions are, and the better the network training effect.

[0020] 4) The network model updates the network link weights using a loss function, therefore, the method for updating link weights is improved:

[0021] The improved convolutional autoencoder network uses an adaptive time estimation method function, which leverages momentum to reduce the likelihood of getting trapped in local minima. Simultaneously, it uses a separate learning rate for each learnable parameter, and this learning rate varies with each parameter during training to improve the network's training performance.

[0022]

[0023] In formula (3), m t v represents the average gradient at the first time step. t θ represents the gradient variance at the second time step. t θ represents the parameter at the previous time step. t+1 The parameter represents the parameter at the next time step, and η and ∈ are settable coefficient values;

[0024] 5) Import the prepared training set samples n2 into the improved convolutional autoencoder network model for network training, and store each trained network model.

[0025] 6) Test each of the stored network models one by one, input seismic data containing surface waves, obtain seismic data without surface waves, perform multi-angle result analysis on the obtained seismic data without surface waves, and comprehensively analyze to obtain the optimal one among the multiple network models;

[0026] 7) Using the selected optimal trained network model, perform transfer application verification on the validation set n1 for seismic data containing surface waves.

[0027] In the above-mentioned seismic data surface wave suppression method based on improved convolutional autoencoder network, in step S1, the number of seismic data samples is 78 pairs, the size of the seismic data samples is 512*512*3px, and the seismic data samples are converted into npy format.

[0028] In the above-mentioned seismic data surface wave suppression method based on improved convolutional autoencoder network, in step S2, 15 pairs of data are selected from 78 pairs of seismic data as a validation set, and the remaining 63 pairs of data are used as a training set for experiments.

[0029] In the above-mentioned seismic data surface wave suppression method based on improved convolutional autoencoder network, in step S3, the autoencoder network includes an input layer, a hidden layer and an output layer in sequence. The input x is input to the input layer and the high-dimensional data is mapped to low-dimensional data h through the encoding process. Then, the data is converted into y through the decoding process and output by the output layer.

[0030] In the above-mentioned seismic data surface wave suppression method based on improved convolutional autoencoder network, in step S4(1), x is input into the encoding unit one by one, and y is output by the decoding unit. The relationship mapping between x and y is obtained through network layer feature extraction.

[0031] In the above-mentioned seismic data surface wave suppression method based on improved convolutional autoencoder network, step S4(6) includes at least the following multi-angle result analysis: peak signal-to-noise ratio and structural similarity.

[0032] In the aforementioned seismic data surface wave suppression method based on an improved convolutional autoencoder network, the peak signal-to-noise ratio (PSNR) is used. snr The mathematical formula for evaluating the quality of network training results using structural similarity is:

[0033]

[0034] In formula (4), M MSE represents the root mean square error, and max(c) represents the maximum value of the pixel value in the noise-free data.

[0035]

[0036] In formula (5), x and y represent the data containing surface waves and the data after removing surface waves, respectively, where μ x , μ represents the mean and variance of x, respectively. y , These represent the mean and variance of y, respectively, and σ xyLet c1 and c2 represent the covariance between x and y, and c2 be constants.

[0037] In the above-mentioned seismic data surface wave suppression method based on improved convolutional autoencoder network, in step S4, 7), 15 prepared seismic data containing surface waves are specifically input into the already trained improved convolutional autoencoder network model, and the corresponding 15 seismic data with surface waves removed are output, thereby achieving effective suppression of surface wave noise in seismic data.

[0038] Compared with existing technologies, the seismic data surface wave suppression method based on improved convolutional autoencoder networks has the following advantages:

[0039] 1. A seismic data surface wave noise suppression model based on an improved convolutional autoencoder network. The model uses a convolutional autoencoder network as the basic model framework and replaces the loss function, gradient descent function and activation function in the original convolutional autoencoder network with the BCE function, Adam function and Swish function to improve the feature learning ability and convergence speed of the network model.

[0040] 2. Based on an improved convolutional autoencoder network seismic data surface wave suppression model, the seismic data containing surface waves is used as the input of the network model, and the seismic data without surface waves is used as the output of the network model. The network is trained using data samples to obtain a complex model mapping from data containing surface waves to data without surface waves, thereby realizing the effective suppression of surface wave noise in seismic data by the network model.

[0041] 3. Convolutional autoencoders reduce the number of required parameters and improve computational efficiency by utilizing the shared weights and local connections in convolutional and deconvolutional layers. At the same time, they enable the network to perceive the features of two-dimensional data. In the encoding part, features are extracted step by step through convolution to eliminate noise signals contained in the data. Then, in the decoding part, deconvolutional layers are used to gradually learn and restore image details, thus having the ability to separate clean data from noisy two-dimensional data, providing a foundation for noise suppression of seismic data.

[0042] 4. By leveraging the characteristics of an improved convolutional autoencoder network, effective seismic signal features are adaptively learned from seismic data containing surface waves, thereby removing prominent surface wave noise. Since the improved convolutional autoencoder network is an unsupervised learning method, it does not require surface wave noise-free annotation of the seismic data, allowing for flexible application. Attached Figure Description

[0043] Figure 1 This is a model framework diagram of autoencoder networks in existing technologies.

[0044] Figure 2 This is a model example diagram of the convolutional autoencoder network in this invention.

[0045] Figure 3It is an S-shaped excitation function diagram in the existing technology.

[0046] Figure 4 This is the Swish activation function diagram in this invention.

[0047] Figure 5 This is a model framework diagram of the improved convolutional autoencoder network in this invention.

[0048] Figure 6 This is a data training loss curve diagram of the present invention. Detailed Implementation

[0049] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and specific examples:

[0050] A seismic surface wave noise suppression method based on an improved convolutional autoencoder network (CARN) is proposed. This method uses an improved CARN for feature extraction, learns the distribution characteristics of effective information, and extracts effective seismic data from seismic data containing surface waves, thereby removing surface wave noise. The improved CARN is based on an autoencoder network framework, and the network model is optimized using convolutional layers, the BCE loss function, and the Swish activation function to enhance the model's overall surface wave feature extraction capability, while also improving the algorithm's robustness and computational efficiency.

[0051] This seismic data surface wave suppression method based on an improved convolutional autoencoder network includes the following steps:

[0052] S1. Separate the head information from the seismic data containing surface waves and the seismic data without surface waves, and separate the data by shot. Set the data size to obtain N pairs of seismic data samples.

[0053] The specific number of earthquake data samples is 78 shots, with 288 channels per shot and 2501 sampling points. The size of each earthquake data sample is 512*512*3px, resulting in a dataset of 156 labeled earthquake data points. To improve the feature extraction capability of the network, various convolutional kernel sizes such as 2*2, 8*8, and 7*7 were used to construct the encoding and decoding units. Each layer of the network has 256 channels with a stride of 2, ensuring that the data size does not change after training the network model. The earthquake data samples were converted to .npy format.

[0054] S2. Match the obtained N pairs of earthquake data samples to obtain the data samples of the improved convolutional autoencoder network model. Randomly select n1 pairs of data as the validation set and the remaining n2 pairs of data as the training set for experiments.

[0055] Specifically, 15 shots were selected from the 78-shot seismic data set as the validation set, and the remaining 63 shots were used as the training set for the experiment.

[0056] S3. First, select an autoencoder network as the basic network framework, input seismic data containing surface waves into the autoencoder network, and output seismic data without surface waves to obtain the network model mapping from seismic data containing surface waves to seismic data without surface waves.

[0057] like Figure 1 As shown, the autoencoder network consists of an input layer, a hidden layer, and an output layer. The input x is fed into the input layer, and the high-dimensional data is mapped to low-dimensional data h through the encoding process. Then, the data is converted into y through the decoding process and output by the output layer.

[0058] The goal is to make the output y as similar to the input x as possible, continuously driving the internal parameter tuning. This allows for the extraction of hidden layer features and parameter learning in an unsupervised manner, enabling the encoder to compress data. The operating principle of autoencoders is existing technology and will not be described in detail here.

[0059] S4. After the basic network model is built, proceed with the optimization of the network model:

[0060] 1) Improve the basic network model by using convolutional layers to optimize the network model's feature extraction capabilities:

[0061] Fully connected autoencoder networks take a one-dimensional array as input, which limits their application to two-dimensional seismic data and visualization images. In seismic data, local data are closely connected, while data that are far apart are not very correlated. However, the structure of a fully connected neural network allows it to connect to all upper-layer neurons, enabling the network to learn the global structure of the data while ignoring local similarities and resulting in redundant parameters.

[0062] A convolutional layer and an activation function form an encoding unit, and a deconvolutional layer and an activation function form a decoding unit. Several encoding units and corresponding decoding units form a complete improved convolutional autoencoder network.

[0063] For improving convolutional autoencoders, the encoder part of the autoencoder is replaced by convolutional layers and activation functions in the convolutional neural network, and the decoder part is replaced by deconvolutional layers and activation functions.

[0064] like Figure 2 As shown, x is input into the encoding unit one by one, and y is output by the decoding unit. The relationship mapping between x and y is obtained through network layer feature extraction.

[0065] 2) Improve the activation layer of the convolutional autoencoder network model by using activation functions to ensure non-linear data transfer between different network layers, thereby making full use of the feature extraction layer in the improved convolutional network model and optimizing the overall feature extraction capability of the network model:

[0066] S-shaped logistic functions were widely used in the early stages of neural network development, such as Figure 3 As shown, its shape is relatively consistent with the actual situation in nature. However, when the input value is large and continues to increase, the gradient becomes very small or even vanishes, making it impossible to update the weight values ​​using the gradient. Consequently, the network cannot continue to optimize, resulting in a network collapse problem.

[0067] like Figure 4 As shown, the improved convolutional autoencoder network uses the Swish function because the saturation property of the sigmoid function easily leads to gradient vanishing, while the Swish function is a smooth and non-monotonic function with no upper bound but a lower bound, and its derivative is always greater than 0. The number of iterations is set to 500, and the training time is approximately 5 days. The Swish function outperforms the sigmoid function in deep models.

[0068] 3) After improving the network layers of the network model, the loss function of the network model is then improved:

[0069] Improvements to the network layers enable the network model to extract more accurate information features from the input seismic data. The loss function is a standard for evaluating the training results during the network model training process. Essentially, it calculates the difference between the predicted and the true values; the smaller the loss function, the more successful the model. The commonly used loss function is Mean Squared Error (MSE). However, when the output probability is close to 0 or 1, the MSE value is very small, causing gradient vanishing at the beginning of model training, leading to network collapse and training failure.

[0070] In improving convolutional autoencoder networks, the binary cross entropy loss (BCE) function is used instead of MSE. This function penalizes both high-confidence erroneous outputs and low-confidence correct outputs, thereby improving the network's discriminative performance.

[0071]

[0072]

[0073] In formula (1), x i y represents the predicted data obtained after network training. i The mean of the sum of squares of the errors between corresponding points in the predicted data and the original data is used to judge the network training effect; in formula (2), x represents the original data. i p(x) represents the label value obtained from network training. i ) represents x i The probability of the label value appearing, q(x) i () represents x under real datai The probability of a label value appearing; the binary cross-entropy loss function represents the distance between two probability distributions. The smaller the binary cross-entropy value, the closer the two probability distributions are, and the better the network training effect.

[0074] 4) By improving the loss function of the network model, the training effect of the network model can be more accurately determined. The network model uses the loss function to update the connection weights, so the method for updating the connection weights needs to be improved:

[0075] The continuous optimization of a network during training hinges on the constant feedback and updates of its weights with each training iteration. Autoencoder networks typically employ stochastic gradient descent (SGD). However, this method has drawbacks, including the possibility of the function getting trapped in local minima and the use of a single learning rate for all learnable parameters, which can negatively impact the network's training performance.

[0076] The improved convolutional autoencoder network uses the Adaptive Moment Estimation (Adam) function with a learning rate of 0.001. This function leverages momentum to reduce the likelihood of getting trapped in local minima. Furthermore, it uses a separate learning rate for each learnable parameter, and this learning rate varies with each parameter during training to improve the network's training performance.

[0077]

[0078] In formula (3), m t v represents the average gradient at the first time step. t θ represents the gradient variance at the second time step. t θ represents the parameter at the previous time step. t+1 The parameter represents the parameter at the next time step, and η and ∈ are settable coefficient values;

[0079] 5) Import the prepared training set samples n2 into the improved convolutional autoencoder network model for network training, and store each trained network model.

[0080] 6) Test each of the stored network models individually, inputting seismic data containing surface waves to obtain seismic data with surface waves removed. Perform multi-angle result analysis on the obtained seismic data with surface waves removed, including at least: peak signal-to-noise ratio and structural similarity. Analyze comprehensively to determine the optimal network model from among the multiple models.

[0081] Using peak signal-to-noise ratio (P) snr The mathematical formula for evaluating the quality of network training results using structural similarity is:

[0082]

[0083] In formula (4), M MSE represents the root mean square error, and max(c) represents the maximum value of the pixel value in the noise-free data.

[0084]

[0085] In formula (5), x and y represent the data containing surface waves and the data after removing surface waves, respectively, where μ x , μ represents the mean and variance of x, respectively. y , These represent the mean and variance of y, respectively, and σ xy Let c1 and c2 represent the covariance between x and y, and c2 be constants.

[0086] Structural similarity is an indicator of the similarity between clean and denoised seismic profiles, which is more in line with the visual evaluation of image quality. By comparing seismic records without surface waves and records after denoising, the more similar the two are, the higher the SSIM value, indicating that the effective information is preserved better, thus proving that the model has a better surface wave noise suppression effect.

[0087] The training set of seismic data was visualized. The training set contained 126 samples, including seismic data with and without surface waves. The surface wave morphology varied across different data sets containing surface waves, ensuring the network model learned the surface wave characteristics as accurately and comprehensively as possible.

[0088] 7) Using the selected optimal trained network model, perform transfer application verification on the validation set n1 for seismic data containing surface waves.

[0089] like Figure 5 As shown, the 15 prepared seismic data containing surface waves are input into the already trained improved convolutional autoencoder network model, and the corresponding 15 seismic data with surface waves removed are output, thus achieving effective suppression of surface wave noise in the seismic data.

[0090] Analysis of the surface wave suppression effect on the validation set data shows that inputting seismic data containing surface waves into a pre-trained improved convolutional autoencoder network yields denoised seismic data. Surface waves exhibit a very pronounced broom-like divergence characteristic compared to effective information, severely weakening the effective information content and making its traces indistinct, thus affecting the imaging quality of the seismic data. The proposed method achieves virtually no residual surface wave noise on the denoised profile, more stable phase axis energy, and higher imaging quality.

[0091] Table 1. Average PSNR and SSIM of validation set data under different network models

[0092]

[0093] The final result obtained by the improved convolutional autoencoder network was compared with that of the original autoencoder network and the convolutional autoencoder network. Data visualization analysis shows that the improved convolutional autoencoder network outperforms the ordinary convolutional autoencoder network and the original autoencoder network in both surface wave noise suppression and effective information preservation. Simultaneously, the denoising performance was quantitatively compared by calculating the peak signal-to-noise ratio (PSNR) and structural similarity of different models. As shown in Table 1, the improved convolutional autoencoder network achieves the highest PSNR and structural similarity in surface wave suppression, indicating that it can effectively suppress surface wave noise while retaining more effective information.

[0094] Table 2. Average PSNR and SSIM of validation set data under different training iterations.

[0095]

[0096] To quantitatively demonstrate the robustness and superiority of this method, Table 2 shows the peak signal-to-noise ratio (PSNR) and structural similarity of the denoising effect after different training iterations. A higher PSNR indicates a higher effective information content, proving that the current model effectively suppresses surface noise; a higher structural similarity indicates better preservation of the effective qualitative information structure, proving that the current model effectively retains the effective information. Figure 6 It can be seen that before the number of iterations reaches 500, the network loss drops rapidly. After reaching 500 iterations, the loss function reaches near the minimum value and tends to level off, indicating that the denoising model trained at this time can stably perform surface wave denoising.

[0097] Compared with existing technologies, the seismic data surface wave suppression method based on improved convolutional autoencoder networks has the following advantages:

[0098] 1. A seismic data surface wave noise suppression model based on an improved convolutional autoencoder network. The model uses a convolutional autoencoder network as the basic model framework and replaces the loss function, gradient descent function and activation function in the original convolutional autoencoder network with the BCE function, Adam function and Swish function to improve the feature learning ability and convergence speed of the network model.

[0099] 2. Based on an improved convolutional autoencoder network seismic data surface wave suppression model, the seismic data containing surface waves is used as the input of the network model, and the seismic data without surface waves is used as the output of the network model. The network is trained using data samples to obtain a complex model mapping from data containing surface waves to data without surface waves, thereby realizing the effective suppression of surface wave noise in seismic data by the network model.

[0100] 3. Convolutional autoencoders reduce the number of required parameters and improve computational efficiency by utilizing the shared weights and local connections in convolutional and deconvolutional layers. At the same time, they enable the network to perceive the features of two-dimensional data. In the encoding part, features are extracted step by step through convolution to eliminate noise signals contained in the data. Then, in the decoding part, deconvolutional layers are used to gradually learn and restore image details, thus having the ability to separate clean data from noisy two-dimensional data, providing a foundation for noise suppression of seismic data.

[0101] 4. By leveraging the characteristics of an improved convolutional autoencoder network, effective seismic signal features are adaptively learned from seismic data containing surface waves, thereby removing prominent surface wave noise. Since the improved convolutional autoencoder network is an unsupervised learning method, it does not require surface wave noise-free annotation of the seismic data, allowing for flexible application.

[0102] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A seismic data surface wave suppression method based on an improved convolutional autoencoder network, characterized in that, Includes the following steps: S1. Separate the head information from the seismic data containing surface waves and the seismic data without surface waves, and separate the data by shot. Set the data size to obtain N pairs of seismic data samples. S2. Match the obtained N pairs of earthquake data samples to obtain the data samples of the improved convolutional autoencoder network model. Randomly select n1 pairs of data as the validation set and the remaining n2 pairs of data as the training set for experiments. S3. First, select an autoencoder network as the basic network framework, input seismic data containing surface waves into the autoencoder network, and output seismic data without surface waves to obtain the network model mapping from seismic data containing surface waves to seismic data without surface waves. S4. After the basic network model is built, proceed with the optimization of the network model: 1) Improve the basic network model by using convolutional layers: A convolutional layer and an activation function form an encoding unit, and a deconvolutional layer and an activation function form a decoding unit. Several encoding units and corresponding decoding units form a complete improved convolutional autoencoder network. 2) Improve the activation layers of the convolutional autoencoder network model using activation functions to ensure non-linear data transfer between different network layers: The improved convolutional autoencoder network uses the Swish function, which is a smooth and non-monotonic function with no upper bound but a lower bound, and its derivative is always greater than 0. 3) After improving the network layers of the network model, the loss function of the network model is then improved: Using the binary cross-entropy loss function in an improved convolutional autoencoder network: In formula (1), x i y represents the predicted data obtained after network training. i The mean of the sum of squares of the errors between corresponding points in the predicted data and the original data is used to judge the network training effect; in formula (2), x represents the original data. i p(x) represents the label value obtained from network training. i ) represents x i The probability of the label value appearing, q(x) i () represents x under real data i The probability of a label value appearing; the binary cross-entropy loss function represents the distance between two probability distributions. The smaller the binary cross-entropy value, the closer the two probability distributions are, and the better the network training effect. 4) The network model updates the network link weights using a loss function, therefore, the method for updating link weights is improved: The improved convolutional autoencoder network uses an adaptive time estimation method function, which leverages momentum to reduce the likelihood of getting trapped in local minima. Simultaneously, it uses a separate learning rate for each learnable parameter, and this learning rate varies with each parameter during training to improve the network's training performance. In formula (3), m t v represents the average gradient at the first time step. t θ represents the gradient variance at the second time step. t θ represents the parameter at the previous time step. t+1 The parameter represents the parameter at the next time step, and η and ∈ are settable coefficient values; 5) Import the prepared training set samples n2 into the improved convolutional autoencoder network model for network training, and store each trained network model. 6) Test each of the stored network models one by one, input seismic data containing surface waves, obtain seismic data without surface waves, perform multi-angle result analysis on the obtained seismic data without surface waves, and comprehensively analyze to obtain the optimal one among the multiple network models; 7) Using the selected optimal trained network model, perform transfer application verification on the validation set n1 for seismic data containing surface waves.

2. The seismic data surface wave suppression method based on an improved convolutional autoencoder network as described in claim 1, characterized in that, In step S1, the number of earthquake data samples is 78 pairs, the size of the earthquake data samples is 512*512*3px, and the earthquake data samples are converted into npy format.

3. The seismic data surface wave suppression method based on an improved convolutional autoencoder network as described in claim 2, characterized in that, In step S2, 15 pairs of data are selected from the 78 pairs of earthquake data as the validation set, and the remaining 63 pairs of data are used as the training set for the experiment.

4. The seismic data surface wave suppression method based on an improved convolutional autoencoder network as described in claim 1, characterized in that, In step S3, the autoencoder network sequentially includes an input layer, a hidden layer, and an output layer. The input x is fed into the input layer, and the high-dimensional data is mapped to low-dimensional data h through the encoding process. Then, the data is converted into y through the decoding process and output by the output layer.

5. The seismic data surface wave suppression method based on an improved convolutional autoencoder network as described in claim 4, characterized in that, In step S4(1), x is input into the encoding unit one by one, and y is output by the decoding unit. The relationship mapping between x and y is obtained through network layer feature extraction.

6. The seismic data surface wave suppression method based on an improved convolutional autoencoder network as described in claim 1, characterized in that, In step S4, 6), the multi-angle result analysis should include at least: peak signal-to-noise ratio and structural similarity.

7. The seismic data surface wave suppression method based on an improved convolutional autoencoder network as described in claim 6, characterized in that, Using peak signal-to-noise ratio P snr The mathematical formula for evaluating the quality of network training results using structural similarity is: In formula (4), M MSE represents the root mean square error, and max(c) represents the maximum value of the pixel value in the noise-free data. In formula (5), x and y represent the data containing surface waves and the data after removing surface waves, respectively, where μ x , μ represents the mean and variance of x, respectively. y , These represent the mean and variance of y, respectively, and σ xy Let c1 and c2 represent the covariance between x and y, and c2 be constants.

8. The seismic data surface wave suppression method based on an improved convolutional autoencoder network as described in claim 3, characterized in that, In step S4(7), the 15 prepared seismic data containing surface waves are specifically input into the trained improved convolutional autoencoder network model, and the corresponding 15 seismic data with surface waves removed are output, thereby effectively suppressing surface wave noise in the seismic data.

Images