CN116186498A - Earthquake signal noise suppression method and system based on self-supervision learning - Google Patents

Abstract

The invention discloses a self-supervision learning-based seismic signal noise suppression method and a self-supervision learning-based seismic signal noise suppression system, which are used for processing a single noisy seismic signal; after original seismic signals are read, data are normalized to a [0,1] interval; then carrying out Bernoulli sampling experiments for a plurality of times on the normalized seismic signals, and constructing data pairs of auxiliary tasks; constructing the whole structure of a denoising network, and determining the optimization target of the network; training the constructed Bernoulli sampling data pair by using a denoising network, and learning a mapping relation from a noise-containing seismic signal to a clean seismic signal; and after the network iteration converges, the parameters of the network model are saved, and the input single noisy seismic signal is recovered. The method effectively solves the problem that the clean seismic signal label is difficult to acquire, reduces the damage to useful signals as much as possible on the basis of effectively suppressing noise, and has good fidelity and practicability.

Description

A method and system for suppressing seismic signal noise based on self-supervised learning

Technical Field

The present invention belongs to the technical field of seismic signal processing, and in particular relates to a seismic signal noise suppression method and system based on self-supervised learning.

Background Art

In field seismic exploration, due to the influence of exploration technology, equipment and environment, the collected seismic signals are inevitably mixed with a large amount of random noise. Noise will seriously affect the resolution of seismic signals, the accuracy of reservoir prediction and the subsequent interpretation of seismic data. Therefore, the research on seismic signal noise suppression methods has always been a hot topic in the field of seismic signal processing.

Traditional methods for suppressing noise in seismic signals have achieved very good results. They often require clear physical information when establishing mathematical models, and then separate useful signals and random noise through mathematical optimization. Traditional methods for suppressing noise in seismic signals can be roughly divided into the following three categories: The first category is based on filtering. The useful signal and noise are transformed into the F-K domain through two-dimensional Fourier transform, and a suitable one-pass and one-resistance region is constructed in the transform domain to achieve the separation of the two signals; the second category is a denoising method based on matrix decomposition. According to the physical characteristics of the useful signal and noise, the noisy signal is decomposed into different components through local SVD, and the effective components are selected to reconstruct the useful signal; the third category is based on sparse representation of the signal. This type of method usually looks for a suitable dictionary in the transform domain to sparsely represent the useful signal, but cannot sparsely represent random noise.

In recent years, with the successful application of deep learning technology represented by convolutional neural networks in the fields of image processing and computer vision, deep learning has also brought new inspiration to the noise suppression of seismic signals. Supervised deep learning methods represented by DnCNN and 3D-DnCNN construct sample-labeled data sets and perform data augmentation, and then train end-to-end convolutional neural networks to enable them to separate noise and successfully restore clean useful signals.

The ability of supervised denoising networks to separate random noise is highly dependent on training data sets. However, in actual seismic exploration, clean seismic signal labels are often difficult to obtain, requiring a lot of manpower, material and financial resources. It is usually necessary to use traditional denoising methods to process noisy seismic signals to construct a data set, so the performance of the denoising network is limited by the traditional denoising methods of preprocessing. When preprocessing massive seismic data, the processing speed will also be very slow.

Summary of the invention

The technical problem to be solved by the present invention is to provide a seismic signal noise suppression method and system based on self-supervised learning in view of the deficiencies in the above-mentioned prior art, so as to solve the technical problem that it is difficult to obtain clean seismic signals in actual seismic exploration.

The present invention adopts the following technical solutions:

A method for suppressing noise of seismic signals based on self-supervised learning comprises the following steps:

Normalize a single seismic signal, use Bernoulli sampling to process the normalized seismic signal, and construct the Bernoulli sampling data pair of the auxiliary task through repeated experiments.

Construct a denoising network based on the encoder-decoder structure;

Determine the optimization objective function of the denoising network;

Train the denoising network to convergence based on the optimization objective function;

The constructed Bernoulli sampling data is input into the denoising network for learning, a single noisy seismic signal is restored, and a clean useful signal is reconstructed.

如下：Specifically, the data of the auxiliary task is

as follows:

Among them, _mi is the mask used for the i-th Bernoulli sampling, ⊙ is the Hadamard product, N is the number of Bernoulli sampling data pairs, and y _norm is the noisy seismic signal after normalization.

Furthermore, the Bernoulli experiment was repeated 100 times.

经过解码器的二维地震信号恢复为原来尺寸；残差噪声分离模块用于计算输入含噪地震信号和网络预测的干净有用信号之间的差值，得到网络分离的噪声，并将噪声先验且均值为0的条件作为正则化约束。Specifically, the denoising network based on the codec structure includes a data processing module, an encoder, a decoder and a residual noise separation module; the data processing module includes a normalization operation and a Bernoulli sampling; the two-dimensional seismic signal encoded by the encoder becomes a two-dimensional seismic signal of the size of the input actual seismic signal.

The two-dimensional seismic signal after the decoder is restored to its original size; the residual noise separation module is used to calculate the difference between the input noisy seismic signal and the clean useful signal predicted by the network, obtain the noise separated by the network, and use the condition that the noise prior and the mean value are 0 as the regularization constraint.

Furthermore, the encoder has 5 encoding modules, each of which includes partial convolution, void convolution, residual learning unit and maximum pooling; the decoder has 5 decoding modules, each of which includes upsampling with a scaling factor of 2, skip connection and a standard convolution layer with Dropout.

Specifically, the optimization objective function L _total of the denoising network is:

L _total =L _target +αL _zm +βL _tv

Among them, L _target is the target loss function, α is the weight coefficient of the noise zero mean loss function, L _zm is the noise zero mean loss function, β is the weight coefficient of the total variation loss function, and L _tv is the total variation loss function.

Specifically, the training of the denoising network is as follows:

The Adam gradient descent algorithm is used to optimize the entire optimization objective function L _total of the denoising network. The initial learning rate is 0.0001, the Epoch is set to 15000, and Dropout is turned on during the training process. After the optimization objective function converges, the parameters of the denoising network are saved.

Specifically, a single noisy seismic signal is input for prediction. Dropout is also turned on during prediction and the experiment is repeated N times. Finally, the average value of each experimental result is selected as the final result.

Furthermore, the average value x′ of each experimental result is calculated as follows:

为第i次伯努利实验中去噪网络对含噪地震信号的恢复结果。in,

is the recovery result of the noisy seismic signal by the denoising network in the i-th Bernoulli experiment.

In a second aspect, an embodiment of the present invention provides a seismic signal noise suppression system based on self-supervised learning, comprising:

The data module normalizes a single seismic signal, processes the normalized seismic signal using Bernoulli sampling, and constructs the Bernoulli sampling data pair for the auxiliary task through repeated experiments;

Building modules, constructing denoising networks based on codec structures;

Function module, which determines the optimization objective function of the denoising network obtained by the construction module;

A training module, which trains the denoising network obtained by the construction module based on the optimization objective function obtained by the function module until convergence;

The suppression module inputs the Bernoulli sampling data obtained by the data module into the denoising network obtained by the training module for learning, recovers the single noisy seismic signal, and reconstructs a clean and useful signal.

Compared with the prior art, the present invention has at least the following beneficial effects:

A noise suppression method for seismic signals based on self-supervised learning. Since the data range of actual seismic signals is large, in order to maintain the consistency of data dimensions, all collected seismic signals are mapped to the interval [0, 1] through normalization operations. At the same time, in order to solve the problem that it is difficult to obtain clean seismic signals, data pairs of auxiliary tasks are constructed through Bernou sampling, and the original problem that cannot be directly solved is converted into a dual problem. Since the original problem and the dual problem have the same solution, the optimization objective function can be constructed and the denoising network can be built for training under the premise that the clean label is unknown, so as to realize blind denoising of a single seismic signal. When constructing the structure of the denoising network, since the encoder-decoder network can extract high-level semantic features of the data through the encoder, the decoder uses upsampling and jump connections to convert the extracted features into target task data, and constructs a denoising network based on the encoder-decoder structure to optimize and solve the objective function. After a certain number of iterations, the denoising network finally separates the useful signal and random noise at the output.

有了伯努利采样数据对，就可以将原始问题转换为对偶问题，得到自监督网络的优化目标。同时伯努利采样也解决了单个地震数据训练样本不足的问题。Furthermore, we construct Bernoulli sampling data pairs

With the Bernoulli sampling data pair, the original problem can be converted into a dual problem and the optimization target of the self-supervised network can be obtained. At the same time, Bernoulli sampling also solves the problem of insufficient training samples of single earthquake data.

一方面解决了单个地震数据训练样本不足的问题；另一方面把去噪网络对每个伯努利采样数据对处理后的结果求平均值，可以提高网络预测结果的准确度。Furthermore, Bernoulli sampling is repeated N = 100 times to construct the data pair

On the one hand, it solves the problem of insufficient training samples for a single earthquake data; on the other hand, averaging the results of the denoising network for each Bernoulli sampling data pair can improve the accuracy of the network's prediction results.

编码器通过一步一步的卷积和池化操作，可以有效地提取输入地震数据地高级语义特征；解码器通过拼接和上采样操作，充分融合深层和浅层网络提取的特征，实现将高级语义特征转为到目标任务的数据，残差噪声分离模块通过计算输入实际地震信号和网络预计有用信号之间的残差得到分离的噪声，并用噪声独立且均值为零先验信息进行约束。残差噪声模块的引入也能在一定程度上避免网络过拟合。Furthermore, the entire denoising network adopts a codec-based structure, including a data processing module, an encoder, a decoder, and a residual noise separation module, wherein the data processing module constructs a Bernoulli sampling data pair through normalization and Bernoulli sampling operations.

The encoder can effectively extract high-level semantic features of input seismic data through step-by-step convolution and pooling operations; the decoder can fully integrate the features extracted by deep and shallow networks through splicing and upsampling operations to realize the conversion of high-level semantic features into data for target tasks. The residual noise separation module obtains the separated noise by calculating the residual between the input actual seismic signal and the network's expected useful signal, and constrains it with the prior information that the noise is independent and has a zero mean. The introduction of the residual noise module can also avoid network overfitting to a certain extent.

Furthermore, the encoder aims to extract high-level semantic features of actual seismic data. Each encoding module includes partial convolution, dilated convolution, residual learning unit and maximum pooling operation. Partial convolution can selectively obtain contextual information through mask to repair the imaging results; dilated convolution introduces holes on the basis of ordinary convolution, which can increase the receptive field of the network; at the same time, residual learning unit is introduced to prevent network degradation; pooling operation can compress imaging and extract semantic features. The decoder includes upsampling with a scaling factor of 2, and the jump connection can fuse the information of shallow network and deep network, and gradually restore the details and size of the target seismic imaging.

尽可能地接近真实值x_i；另一方面最小化了噪声n_i的能量。L_zm保证了去噪网络分离的噪声满足均值为零，防止网络分离的噪声中掺杂有用信号，提高了恢复有用信号的保真性。L_tv通过约束水平和垂直方向上的梯度变化，能在一定程度上压制噪声。这三个损失函数一起组成整个网络的优化目标，提高整个网络的去噪性能。Furthermore, introducing a regularization term based on the network optimization target L _target can prevent overfitting and improve the fidelity of the recovered clean seismic signal. Minimizing L _target can make the prediction value of the denoising network

As close to the true value x _i as possible; on the other hand, it minimizes the energy of the noise n _i . L _zm ensures that the noise separated by the denoising network satisfies the mean of zero, prevents the noise separated by the network from being mixed with useful signals, and improves the fidelity of the restored useful signals. L _tv can suppress noise to a certain extent by constraining the gradient changes in the horizontal and vertical directions. These three loss functions together constitute the optimization goal of the entire network and improve the denoising performance of the entire network.

Furthermore, the Adam gradient descent method is used to optimize the loss function of the entire network, with an initial learning rate of 0.0001. The purpose of the gradient descent method is to find the minimum value of the network loss function and determine the optimal parameters of the entire denoising network. In addition, Adam can automatically adjust the learning rate for each input variable of the loss function and update the variables by using a moving average that exponentially reduces the gradient.

Furthermore, a single noisy input seismic signal is predicted, and Bernoulli sampling is repeated N times to construct data pairs and send them to the denoising network for learning. The average of the predicted results is taken as the final result. Multiple experiments not only expand the data, but also improve the accuracy of the network.

Furthermore, multiple Bernoulli sampling experiments are equivalent to training multiple denoising networks, and the prediction results of each denoising network for a single noisy seismic signal are averaged as the final result. This can improve the accuracy of the network in predicting clean and useful signals.

It can be understood that the beneficial effects of the second aspect mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here.

In summary, the present invention transforms the problem that cannot be directly solved by Bernoulli sampling based on the mathematical characteristics of noise, and obtains the approximate optimal solution of the original problem by solving the dual problem. The present invention saves a lot of time cost required to construct clean seismic signals with artificial labels, and also solves the problem that clean seismic signals are difficult to obtain in actual seismic exploration.

The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments.

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 is a flow chart of the invention;

FIG2 is a diagram showing the structure of the entire denoising neural network of the present invention;

FIG3 is a diagram showing the processing results of actual seismic data, wherein (a) is the seismic signal actually acquired, (b) is the useful seismic signal recovered by the network of the present invention, and (c) is the noise separated by the network of the present invention;

FIG4 is a diagram of the F-K spectrum of FIG3 , wherein (a) is the F-K spectrum corresponding to the actual seismic signal, (b) is the F-K spectrum of the useful seismic signal, and (c) is the spectrum of the noise;

FIG5 is a diagram showing the processing results of actual seismic data, wherein (a) is the seismic signal actually acquired, (b) is the useful seismic signal recovered by the network of the present invention, and (c) is the noise separated by the network of the present invention.

DETAILED DESCRIPTION

The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

In the description of the present invention, it should be understood that the terms “include” and “comprises” indicate the presence of described features, wholes, steps, operations, elements and/or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and/or collections thereof.

It should also be understood that the terms used in the present specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, the singular forms "a", "an" and "the" are intended to include plural forms unless the context clearly indicates otherwise.

It should be further understood that the term "and/or" used in the present specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes these combinations. For example, A and/or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character "/" in this article generally indicates that the associated objects are in an "or" relationship.

It should be understood that, although the terms first, second, third, etc. may be used to describe preset ranges, etc. in the embodiments of the present invention, these preset ranges should not be limited to these terms. These terms are only used to distinguish preset ranges from each other. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be referred to as the second preset range, and similarly, the second preset range may also be referred to as the first preset range.

The word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)", depending on the context.

Various structural schematic diagrams of the embodiments disclosed in the present invention are shown in the accompanying drawings. These figures are not drawn to scale, and some details are magnified and some details may be omitted for the purpose of clear expression. The shapes of various regions and layers shown in the figures and the relative sizes and positional relationships therebetween are only exemplary, and may deviate in practice due to manufacturing tolerances or technical limitations, and those skilled in the art may further design regions/layers with different shapes, sizes, and relative positions according to actual needs.

The present invention provides a method for suppressing noise of seismic signals based on self-supervised learning. Before processing with an end-to-end convolutional neural network, it is not necessary to construct a clean label. Instead, the mathematical characteristics of the data itself are mined as supervisory information (not the original task label, but the constructed auxiliary task label) through auxiliary tasks, so that the convolutional neural network has the ability to separate useful signals and noise. Due to the uneven distribution of seismic signal data, the input seismic signal is first normalized to the [0, 1] interval using a normalization operation. Then, the normalized seismic signal is repeated multiple times using Bernoulli sampling experiments to construct the label of the auxiliary task. An end-to-end deep learning network based on a codec structure is built, and the data pairs constructed by Bernoulli sampling are trained. When the network converges after a certain number of iterations, the parameters of the network model are saved, and the mapping from a single noisy seismic signal to a clean useful signal is realized.

Referring to FIG1 , a method for suppressing noise of seismic signals based on self-supervised learning of the present invention comprises the following steps:

S1, normalize the input single seismic signal to the interval [0, 1];

First, read the actual seismic data y and perform data normalization, as shown in the following formula (1):

Where y _min and y _max represent the minimum and maximum values of the two-dimensional seismic signal, respectively. The normalization operation converts the input noisy seismic signal to the interval [0, 1].

S2, use Bernoulli sampling to process seismic signals and repeat the experiment multiple times to construct data pairs for auxiliary tasks;

如下所示：Perform multiple Bernoulli sampling on the normalized seismic signal to construct a data pair

As shown below:

Among them, _mi is the mask of the i-th Bernoulli sampling, ⊙ is the Hadamard product, and y _norm is the noisy seismic signal after normalization.

Bernoulli sampling is equivalent to a binary mask consisting of only 0 and 1 acting on a two-dimensional seismic signal. The probability of retaining each position on the mask is 0.7. The Bernoulli sampling data pair is constructed in this one-pass and one-block manner. At the same time, in order to solve the problem of insufficient data volume of a single noisy seismic signal, the Bernoulli experiment was repeated 100 times.

S3, build a denoising network based on the codec structure;

解码器同样有5个解码模块，每个模块包括比例因子为2的上采样、跳跃连接和带有Dropout的标准卷积层。经过解码器，二维地震信号最终恢复原来尺寸。最后是残差噪声分离模块，计算输入含噪地震信号和网络预测的干净信号之间的差值得到网络分离的噪声，并将噪声先验且均值为0的条件作为正则化约束来防止网络过拟合，这样就可以利用去噪网络对目标函数进行优化求解。Please refer to Figure 2. The denoising network based on the codec structure is similar to the traditional codec network UNet, which mainly includes a data processing module, an encoder, a decoder, and a residual noise separation module. Among them, the data processing module includes normalization operations and Bernoulli sampling. The encoder has a total of 5 encoding modules, each of which includes partial convolution, void convolution, residual learning unit, and maximum pooling. After the encoder, the encoded two-dimensional seismic signal is finally converted into an input size of

The decoder also has 5 decoding modules, each of which includes upsampling with a scale factor of 2, skip connections, and standard convolutional layers with Dropout. After the decoder, the two-dimensional seismic signal is finally restored to its original size. Finally, there is the residual noise separation module, which calculates the difference between the input noisy seismic signal and the clean signal predicted by the network to obtain the noise separated by the network, and uses the noise prior and mean 0 as a regularization constraint to prevent the network from overfitting, so that the denoising network can be used to optimize and solve the objective function.

S4, determining the optimization objective function of the denoising network obtained in step S3;

Considering that the clean seismic signal _xi is unknown, it is impossible to directly use network optimization to solve equation (4):

According to the data pairs constructed by Bernoulli sampling in step 02 and the characteristics that the noise is independent and the mean is 0, equation (4) is transformed into equations (5a) and (5b):

In order to avoid any useful signal being mixed in the separated noise, the prior information that the noise is independent and has a mean of 0 is used as part of the optimization objective in the form of a regularization penalty term. At the same time, the total variation loss function is introduced as a regularization term to further suppress the noise by constraining the gradient changes in the horizontal and vertical directions. Finally, the optimization objective function of the entire denoising network is determined as shown in the following equations (6a)-(6d):

L _total ＝L _target +aL _zm +βL _tv (6a)

和，

为伯努利采样构建的数据对，m_i为第i次伯努利采样使用的掩膜，θ为去噪网络的参数，n′_h，w为去噪网络分离的噪声n′在(h，w)位置的值，x′_h，w为去噪网络恢复的干净地震信号x′在(h，w)位置的值。Where L _target is the target loss function, L _zm is the noise zero mean loss function, α is the weight coefficient of L _zm , L _tv is the total variation loss function, β is the weight coefficient of L _tv , H is the height of the two-dimensional seismic data (number of sampling points), W is the width of the two-dimensional seismic data (number of traces), F _θ (·) is the denoising network,

and,

is the data pair constructed for Bernoulli sampling, _mi is the mask used for the i-th Bernoulli sampling, θ is the parameter of the denoising network, n′h _,w is the value of the noise n′ separated by the denoising network at the position (h,w), _{and x′h,w} is the value of the clean seismic signal x′ restored by the denoising network at the position (h,w).

S5, training the denoising network obtained in step S3 based on the optimization objective function obtained in step S4 until convergence, and saving the parameters of the denoising network;

The Adam gradient descent algorithm is used to optimize the objective function (6a). The initial learning rate is set to 0.0001, the Epoch size is 15000, and Dropout is turned on during the training process. After the objective function converges, the parameters of the network model are saved.

S6. Input the noisy seismic signal into the denoising network obtained in step S5 for learning, restore the single noisy seismic signal, and reconstruct a clean useful signal.

For the prediction of a single noisy seismic signal, Dropout is also turned on and the test is repeated N times. Finally, the average value of each test result is selected as the final result, as shown below:

为第i次伯努利实验中去噪网络对含噪地震信号的恢复结果。in,

is the recovery result of the noisy seismic signal by the denoising network in the i-th Bernoulli experiment.

The denoising network of the present invention can process the input noisy seismic signal end-to-end, explore the characteristics of random noise, and finally map the noisy seismic signal into a clean useful signal.

In another embodiment of the present invention, a seismic signal noise suppression system based on self-supervised learning is provided. The system can be used to implement the above-mentioned seismic signal noise suppression method based on self-supervised learning. Specifically, the seismic signal noise suppression system based on self-supervised learning includes a data module, a construction module, a function module, a training module and a suppression module.

Among them, the data module normalizes a single seismic signal, uses Bernoulli sampling to process the seismic signal obtained by normalization, and constructs the Bernoulli sampling data pair of the auxiliary task through repeated experiments;

Building modules, constructing denoising networks based on codec structures;

Function module, which determines the optimization objective function of the denoising network obtained by the construction module;

A training module, which trains the denoising network obtained by the construction module based on the optimization objective function obtained by the function module until convergence;

The suppression module inputs the Bernoulli sampling data obtained by the data module into the denoising network obtained by the training module for learning, recovers the single noisy seismic signal, and reconstructs a clean and useful signal.

In another embodiment of the present invention, a terminal device is provided, which includes a processor and a memory, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is suitable for implementing one or more instructions, and is specifically suitable for loading and executing one or more instructions to implement the corresponding method flow or corresponding function; the processor described in the embodiment of the present invention can be used for the operation of the seismic signal noise suppression method based on self-supervised learning, including:

A single seismic signal is normalized and processed using Bernoulli sampling. After repeated experiments, a Bernoulli sampling data pair for auxiliary tasks is constructed. A denoising network based on a codec structure is constructed. The optimization objective function of the denoising network is determined. The denoising network is trained until convergence based on the optimization objective function. The constructed Bernoulli sampling data pair is input into the denoising network for learning, a single noisy seismic signal is restored, and a clean and useful signal is reconstructed.

In another embodiment of the present invention, the present invention also provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a terminal device for storing programs and data. It can be understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and the extended storage medium supported by the terminal device. The computer-readable storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory, or a non-volatile memory (Non-Volatile Memory), such as at least one disk memory.

The processor may load and execute one or more instructions stored in a computer-readable storage medium to implement the corresponding steps of the method for suppressing seismic signal noise based on self-supervised learning in the above embodiment; the processor may load and execute the following steps:

A single seismic signal is normalized and processed using Bernoulli sampling. After repeated experiments, a Bernoulli sampling data pair for auxiliary tasks is constructed. A denoising network based on a codec structure is constructed. The optimization objective function of the denoising network is determined. The denoising network is trained until convergence based on the optimization objective function. The constructed Bernoulli sampling data pair is input into the denoising network for learning, a single noisy seismic signal is restored, and a clean and useful signal is reconstructed.

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention described and shown in the drawings here can usually be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

Please refer to Figure 3. Figure 3 (a) shows the actual collected two-dimensional seismic signal, with a total of 500 channels, each channel has 800 sampling points, and the sampling time interval is 2ms. It can be seen that the horizontal event axis is strongly interfered by random noise, which seriously affects the continuity and smoothness of the useful signal event axis and also reduces the signal-to-noise ratio. The input signal is processed using a seismic signal noise suppression method based on self-supervised learning.

Figure 3(b) and Figure 3(c) are respectively the useful signal reconstructed by the denoising network of the present invention and the separated random noise. It can be seen that the denoising network of the present invention has a significant suppression effect on random noise, and the continuity and smoothness of the event axis of the seismic signal have been restored. And by observing the noise separated by the network, we can't see any event axis structure. That is, our network has good fidelity and basically does not cause damage to the useful signal. In addition, observing Figure 3(c), the present invention also suppresses the coherent noise in the tilt direction to a certain extent. In order to further analyze the effectiveness of the denoising network of the present invention, Figure 3(a), Figure 3(b) and Figure 3(c) are transformed into the F-K domain, as shown in Figure 4(a), Figure 4(b) and Figure 4(c). It can be seen that the energy of the useful signal is mainly concentrated in the middle part, while the random noise is mainly distributed around it. The useful signal and random noise separated by the denoising network are also well separated in the transform domain, which also indirectly indicates the effectiveness of the network of the present invention.

Figures 5(a), 5(b) and 5(c) are another set of actual seismic data, which are noisy seismic signals, useful signals recovered by the present invention and separated noise, respectively. There are 1000 actual seismic signals in total, each with 3001 sampling points, and the sampling time interval is 2ms. In Figure 5(b), it can be seen that the denoising network of the present invention can suppress noise well in both areas containing weak noise data and areas containing strong noise data, ensuring the continuity of the seismic signal event axis. And observing Figure 5(c), no horizontal event axis structure can be seen, indicating that the recovered useful signal has basically no energy residue and has good fidelity.

In summary, the present invention is a method and system for suppressing noise of seismic signals based on self-supervised learning. According to the mathematical characteristics of noise, the problem that cannot be solved directly is transformed through Bernoulli sampling, and the approximate optimal solution of the original problem is obtained by solving the dual problem. This solves the problem that clean seismic signals are difficult to obtain in actual seismic exploration. In addition, when constructing a denoising network, in order to avoid overfitting of a single input signal during training, a residual learning unit is introduced to improve the performance of the network. And in order to better mine the characteristics of seismic data, a hole convolution is used to increase the receptive field of the entire network. Finally, a residual noise separation module is added, which separates random noise by calculating the difference between the noisy seismic signal and the clean useful signal predicted by the network, and uses the characteristics of noise independence and zero mean as a regularization constraint to avoid the separation of noise mixed with relevant information of useful signals, so that the entire network has fidelity. The experimental results show that the present invention can not only effectively remove random noise, but also has little damage to the recovered useful signal, and has good fidelity and practicality.

The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

In the embodiments provided by the present invention, it should be understood that the disclosed devices/terminals and methods can be implemented in other ways. For example, the device/terminal embodiments described above are only schematic. For example, the division of the modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

If the integrated module/unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

The present application is described with reference to the flowchart and/or block diagram of the method, device (system) and computer program product according to the embodiment of the present application. It should be understood that each process and/or box in the flowchart and/or block diagram, and the combination of the process and/or box in the flowchart and/or block diagram can be realized by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for realizing the function specified in one process or multiple processes in the flowchart and/or one box or multiple boxes in the block diagram.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device that implements the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.

These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.

The above contents are only for explaining the technical idea of the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims (10)

1. A seismic signal noise suppression method based on self-supervised learning, is characterized in that, comprises the following steps: Normalize a single seismic signal, use Bernoulli sampling to process the normalized seismic signal, and construct Bernoulli sampling data pairs for auxiliary tasks through repeated experiments; Build a denoising network based on codec structure; Determining the optimization objective function of the denoising network; Train the denoising network to convergence based on the optimized objective function; The constructed Bernoulli sampling data is input into the denoising network for learning, and a single noisy seismic signal is recovered and reconstructed to obtain a clean and useful signal. 2. the seismic signal noise suppression method based on self-supervised learning according to claim 1, is characterized in that, the data pair of auxiliary task

as follows:

Among them, m _i is the mask used for the i-th Bernoulli sampling, ⊙ is the Hadamard product, N is the number of Bernoulli sampling data pairs, and y _norm is the noisy seismic signal after normalization. 3. The seismic signal noise suppression method based on self-supervised learning according to claim 2, wherein the Bernoulli experiment is repeated 100 times. 4. the seismic signal noise suppression method based on self-supervised learning according to claim 1, is characterized in that, the denoising network based on codec structure comprises data processing module, coder, decoder and residual noise separation module; The processing module includes normalization operation and Bernoulli sampling; the two-dimensional seismic signal encoded by the encoder becomes the input actual seismic signal size

The two-dimensional seismic signal after the decoder is restored to its original size; the residual noise separation module is used to calculate the difference between the input noisy seismic signal and the clean and useful signal predicted by the network to obtain the noise separated by the network, and the noise priori And the condition with a mean value of 0 is used as a regularization constraint. 5. the seismic signal noise suppression method based on self-supervised learning according to claim 4, is characterized in that, encoder has 5 encoding modules, and each encoding module comprises partial convolution, hole convolution, residual learning unit and Max pooling; the decoder has 5 decoding modules, each including upsampling with a scale factor of 2, skip connections and standard convolutional layers with dropout. 6. the seismic signal noise suppression method based on self-supervised learning according to claim 1, is characterized in that, the optimized objective function L _total of denoising network is specifically: L _total ＝ L _target + αL _zm + βL _tv Among them, L _target is the target loss function, α is the weight coefficient of the noise zero-mean loss function, L _zm is the noise zero-mean loss function, β is the weight coefficient of the total variation loss function, and L _tv is the total variation loss function. 7. the seismic signal noise suppression method based on self-supervised learning according to claim 1, is characterized in that, training denoising network is specially: Use the Adam gradient descent algorithm to optimize the entire optimization objective function L _total of the denoising network. The initial learning rate is 0.0001, and the Epoch is set to 15000. Dropout is turned on during the training process, and the parameters of the denoising network are saved after the optimization objective function converges. 8. The seismic signal noise suppression method based on self-supervised learning according to claim 1, characterized in that, the input single noisy seismic signal is predicted, and Dropout is also opened during prediction and N times of trials are repeated, finally selecting The average of the experimental results was taken as the final result. 9. the method for suppressing seismic signal noise based on self-supervised learning according to claim 8, is characterized in that, calculates the average value x ' of each experimental result specifically as:

in,

is the restoration result of the noise-containing seismic signal by the denoising network in the i-th Bernoulli experiment. 10. A seismic signal noise suppression system based on self-supervised learning, characterized in that, comprising: The data module normalizes a single seismic signal, uses Bernoulli sampling to process the normalized seismic signal, and constructs Bernoulli sampling data pairs for auxiliary tasks through repeated experiments; Building blocks to build a denoising network based on the codec structure; A function module, which determines the optimization objective function of the denoising network obtained by the building module; The training module is based on the optimized objective function obtained by the function module to train the denoising network obtained by the building module to convergence; In the compression module, the Bernoulli sampling data obtained by the data module is input into the denoising network obtained by the training module for learning, recovering a single noisy seismic signal, and reconstructing to obtain a clean and useful signal.

Images