A desert seismic denoising method based on multiple attention mechanism

Abstract

The application discloses a desert seismic denoising method based on a multiple attention mechanism, and comprises the following steps: step one, constructing an initial denoising network based on the multiple attention mechanism, which comprises a mean shift layer, a convolution layer, a main activation function layer, a first enhanced attention module, a supervised attention module, a second enhanced attention module, a convolution layer and a mean shift layer connected in sequence; step two, training the initial denoising network based on the multiple attention mechanism to obtain an ideal denoising network based on the multiple attention mechanism; and step three, inputting an original noisy seismic signal into the ideal denoising network based on the multiple attention mechanism to obtain a denoised seismic signal. The application has the characteristics that random noise can be suppressed and a signal can be restored even in the case of a low signal-to-noise ratio.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geophysics, and more particularly, the present application relates to a desert seismic denoising method based on a multiple attention mechanism. BACKGROUND

[0002] For seismic exploration, excellent seismic records are of great significance. However, due to long-term large-scale exploitation, existing natural resources are mainly distributed in complex areas with weak signals and low signal-to-noise ratio (SNR), in addition, in the time domain and frequency domain, non-coherent noise and signals overlap, these characteristics all put forward higher requirements for denoising methods.

[0003] In order to extract signals from noise records, in the past few years, a large number of traditional denoising algorithms have rapidly developed in this field. For example, initial random denoising (L.L. Canales et al. 1984), wavelet transform (Z.Yu et al. 2017), curvelet transform (J.Li et al. 2017), contourlet transform (R.Moore et al. 2014), empirical mode decomposition (EMD) (M.Bekara et al. 2009), variational mode decomposition (VMD) (F.Li et al. 2018), f-x predictive filtering (K.Chen et al. 2017), t-x (R.Abma et al. 1995), shearlet transform (C.Zhang et al. 2018), model-based optimization methods and time-frequency peak filtering (TFPF) (B.Boashash. 2004) and the like, these methods have important guiding significance for the research of seismic denoising, although the above-mentioned classical methods can improve the denoising performance to a certain extent, but they rely on parameters to a great extent, these parameters are determined by the experience of field workers and specific geological conditions, considering the large amount of data that needs to be processed, if these methods are used, a lot of time is needed to adjust the parameters, in addition, most methods rely on the assumption of Gaussian white noise, not only ignoring the phenomenon of spectral aliasing, but also ignoring the similarity between signals and noise, also unable to maintain the texture characteristics of complex seismic data, even more unable to introduce ambiguity (C.Yang et al. 2021), that is, these algorithms cannot meet the more complex conditions urgently needed in today's research.

[0004] In recent years, convolutional neural network (CNN) has become a hotspot of researchers due to its wide application in detection, segmentation, image processing, estimation, retrieval, etc. Compared with traditional methods, the CNN-based algorithm no longer needs prior information about the signal or noise, can avoid manual adjustment of parameters (Chao Zhang et al. 2019), and can better restore the weak signal covered by noise. On the basis of CNN, many improved methods have been proposed. Zhang (2017) proposed a convolutional neural denoising network (DnCNN), which adopts a combination of convolutional layers, batch normalization layers and ReLU functions. Moreover, the algorithm uses a residual structure to improve the stability of training and denoising ability. When using quantitative indicators and visual quality as evaluation indicators, DnCNN achieves remarkable results in dealing with known and unknown noise levels of Gaussian noise. The CNN-based denoising method focuses on the nonlinear mapping mechanism generated by a large amount of labeled seismic data to extract features from noisy data (D. Liu et al. 2019). However, the neural network-based method also has some disadvantages. First, there are a large number of parameters in the network that need to be adjusted according to experience. Second, the neural network is prone to overfitting, and some signals that do not originally exist may appear in the processing results, which poses great challenges to related methods. SUMMARY

[0005] The purpose of the present application is to design and develop a desert seismic denoising method based on a multiple attention mechanism. By extracting important features repeatedly from noisy records based on detailed information extracted by two attention modules, the denoising ability of the network is enhanced, and the effective signal in the seismic data is recovered under low signal-to-noise ratio.

[0006] The technical scheme provided by the present application is as follows:

[0007] A desert seismic denoising method based on a multiple attention mechanism, comprising the following steps:

[0008] Step 1: Construct an initial denoising network based on a multiple attention mechanism;

[0009] The denoising network based on a multiple attention mechanism comprises a first mean shift layer, a convolutional layer, a main activation function layer, a first enhanced attention module, a supervised attention module, a second enhanced attention module, a convolutional layer and a second mean shift layer connected in sequence.

[0010] The first and second enhanced attention modules each comprise an 18-layer structure, the first layer is an expanded convolution module, the second to sixth layers are composed of convolution layers and activation function layers in turn and at intervals, the seventh layer is a convolution layer, the eighth to thirteenth layers are composed of convolution layers and activation function layers in turn and at intervals, the fourteenth layer is an average pooling layer, the fifteenth to seventeenth layers are composed of convolution layers and activation function layers in turn and at intervals, and the eighteenth layer is an activation function layer; the input of the first layer and the output of the third layer form a residual structure as the input of the fourth layer, the input of the fourth layer and the output of the sixth layer form a residual structure as the input of the seventh layer, the output of the seventh layer and the output of the twelfth layer form a residual structure as the input of the thirteenth layer, and the output of the thirteenth layer and the output of the eighteenth layer form a residual structure as the output of the first enhanced attention module;

[0011] The expanded convolution module comprises a first expanded convolution layer, a second expanded convolution layer, a third expanded convolution layer and a fourth expanded convolution layer, the inputs of the first and third expanded convolution layers are the output of the main activation function layer, the output of the first expanded convolution layer is the input of the second expanded convolution layer, the output of the third expanded convolution layer is the input of the fourth expanded convolution layer, and the outputs of the second and fourth expanded convolution layers are the inputs of the second layer; the dilation factor of the first expanded convolution layer is 1, the dilation factor of the second expanded convolution layer is 2, the dilation factor of the third expanded convolution layer is 3, and the dilation factor of the fourth expanded convolution layer is 4; the size of the convolution filter of the twelfth layer of the first enhanced attention module is set to 1x1, and the size of the remaining convolution filters in the denoising network based on the multiple attention mechanism is set to 3x3.

[0012] Step two, training the initial denoising network based on the multiple attention mechanism to obtain an ideal denoising network based on the multiple attention mechanism;

[0013] Step three, inputting the original noisy seismic signal into the first mean shift layer of the ideal denoising network based on the multiple attention mechanism;

[0014] Step four, the data processed by the main activation function layer enter the first enhanced attention module, the supervised attention module and the second enhanced attention module in turn;

[0015] Step five, outputting the data received after being processed by the second mean shift layer to obtain a denoised seismic signal.

[0016] Preferably, the expanded convolution module further comprises:

[0017] A plurality of activation function layers are arranged at the output ends of the first, second, third and fourth expanded convolution layers one by one.

[0018] Preferably, the supervision attention module is composed of a first convolutional layer, a second convolutional layer, a third convolutional layer, and a secondary activation function layer.

[0019] The output of the first enhanced attention module is used as the input of the first convolutional layer and the second convolutional layer, the output of the second convolutional layer is used as the input of the third convolutional layer in a residual structure with the original noisy seismic signal, the output of the third convolutional layer is used as the input of the secondary activation function layer, and the output of the secondary activation function layer is used as the output of the supervision attention module after being combined with the output of the first convolutional layer to form a first-level residual structure and combined with the output of the first enhanced attention module to form a second-level residual structure.

[0020] Preferably, the secondary activation function layer is a Sigmoid function.

[0021] Preferably, the 18th layer of the first enhanced attention module and the second enhanced attention module is a Sigmoid function, and the other activation function layers in the first enhanced attention module and the second enhanced attention module are ReLU functions.

[0022] Preferably, the main activation function layer is a ReLU function.

[0023] Preferably, the calculation formula of the average pooling layer is:

[0024] y1=f sig (f up (f ReLU (f down (f avg (x1)))))×x1;

[0025] In the formula, x1 is the input of the first enhanced attention module or the second enhanced attention module, y1 is the corresponding output of the first enhanced attention module or the second enhanced attention module, f avg is an average pooling layer, f down is a down-sampling layer, f ReLU is a ReLU activation function, f up is a corresponding up-sampling layer, f sig is a Sigmoid activation function.

[0026] Preferably, the performance of the ideal denoising network based on the multiple attention mechanism satisfies:

[0027]

[0028] In the formula, is the denoising result obtained after the denoising network is processed, and M is the noisy data. Noise calculated by the denoising network, S is the overall denoising network structure, and mu is the parameter in the denoising network.

[0029] Preferably, the parameter in the denoising network is obtained by adjusting the L1 loss function.

[0030] Preferably, the training set for training the initial denoising network based on the multiple attention mechanism includes pure seismic signals and noise.

[0031] The pure seismic signal is synthesized by a Ricker wavelet, and contains 128 recording channels, each channel has 1024 sampling points, and the sampling frequency is 512 Hz.

[0032] The noise is cut from the real noise record collected in the desert environment, contains 2332 channels, each channel has 59998 sampling points, and the noise is cut to the same size as the pure seismic signal.

[0033] When the training set is used, the multiple of the adjusted noise is adjusted so that the signal-to-noise ratio of the noisy desert signal is between-2dB and-6dB.

[0034] The beneficial effects of the application are:

[0035] The desert seismic denoising method based on the multiple attention mechanism designed and developed by the application can more effectively extract the features in the original data through the hamburger type attention structure of "inter-channel attention mechanism-supervised attention mechanism-inter-channel attention mechanism", and further improve the overall denoising ability of the network. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 The structure diagram of the desert seismic denoising method based on the multiple attention mechanism.

[0037] Figure 2 The network diagram of the EAM.

[0038] Figure 3 The network diagram of the SAM.

[0039] Figure 4 The schematic diagram of the pure record of the simulation record.

[0040] Figure 5 The schematic diagram of the noisy record of the simulation record.

[0041] Figure 6 The schematic diagram of the band-pass filter denoising result of the simulation record.

[0042] Figure 7A schematic diagram of the WNNM denoising result of the simulated recording according to the present application.

[0043] Figure 8 A schematic diagram of the DnCNN denoising result of the simulated recording according to the present application.

[0044] Figure 9 A schematic diagram of the denoising network of the simulated recording according to the present application.

[0045] Figure 10 A schematic diagram of the added noise of the simulated recording according to the present application.

[0046] Figure 11 A schematic diagram of the band-pass filter denoising difference of the simulated recording according to the present application.

[0047] Figure 12 A schematic diagram of the WNNM denoising difference of the simulated recording according to the present application.

[0048] Figure 13 A schematic diagram of the DnCNN denoising difference of the simulated recording according to the present application.

[0049] Figure 14 A schematic diagram of the denoising network denoising difference of the simulated recording according to the present application.

[0050] Figure 15 A schematic diagram of the F-K spectrum of the pure recording of the simulated recording according to the present application.

[0051] Figure 16 A schematic diagram of the F-K spectrum of the noisy recording of the simulated recording according to the present application.

[0052] Figure 17 A schematic diagram of the F-K spectrum of the band-pass filter denoising result of the simulated recording according to the present application.

[0053] Figure 18 A schematic diagram of the F-K spectrum of the WNNM denoising result of the simulated recording according to the present application.

[0054] Figure 19 A schematic diagram of the F-K spectrum of the DnCNN denoising result of the simulated recording according to the present application.

[0055] Figure 20 A schematic diagram of the F-K spectrum of the denoising network denoising result of the simulated recording according to the present application.

[0056] Figure 21 A schematic diagram of the single-channel comparison of the band-pass filter denoising result and the pure seismic signal of the simulated recording according to the present application.

[0057] Figure 22A single-channel comparison graph of the denoising result of the simulated recording of the WNNM of the application and a pure seismic signal.

[0058] Figure 23 A single-channel comparison graph of the denoising result of the simulated recording of the DnCNN of the application and a pure seismic signal.

[0059] Figure 24 A single-channel comparison graph of the denoising result of the simulated recording of the denoising network of the application and a pure seismic signal.

[0060] Figure 25 A comparison curve graph of the denoising effects of the six different network structures of the application.

[0061] Figure 26 A single-channel comparison curve graph of the SAM of the application and a pure signal.

[0062] Figure 27 A single-channel comparison curve graph of the EAM of the application and a pure signal.

[0063] Figure 28 A single-channel comparison curve graph of the 2EAM of the application and a pure signal.

[0064] Figure 29 A single-channel comparison curve graph of the 3EAM of the application and a pure signal.

[0065] Figure 30 A single-channel comparison curve graph of the 4EAM of the application and a pure signal.

[0066] Figure 31 A single-channel comparison curve graph of the denoising network of the application and a pure signal.

[0067] Figure 32 A schematic diagram of the original noisy seismic signal of the actual seismic data of the application.

[0068] Figure 33 A schematic diagram of the band-pass filter denoising result of the actual seismic data of the application.

[0069] Figure 34 A schematic diagram of the WNNM denoising result of the actual seismic data of the application.

[0070] Figure 35 A schematic diagram of the DnCNN denoising result of the actual seismic data of the application.

[0071] Figure 36 A schematic diagram of the denoising network denoising result of the actual seismic data of the application. DETAILED DESCRIPTION

[0072] The application will be further described in detail below, so that those skilled in the art can implement the application according to the description.

[0073] The desert seismic denoising method based on the multiple attention mechanism provided by the application is composed of three parts and includes two different feature attention modules, namely the inter-channel attention mechanism in the enhanced attention module and the supervised attention module, forming a hamburger-style attention mechanism structure for extracting important features from seismic data.

[0074] 1. Constructing an initial denoising network based on the multiple attention mechanism;

[0075] As shown in Figure 1 The denoising network based on the multiple attention mechanism includes a first mean shift layer (M), a convolution layer (Conv), a main activation function layer, a first enhanced attention module (EAM), a supervised attention module (SAM), a second enhanced attention module, a convolution layer, and a second mean shift layer, which are connected in sequence.

[0076] That is, the denoising network based on the multiple attention mechanism is composed of the first EAM, the SAM, and the second EAM, the first EAM and the second EAM have the same structure, and the SAM is used to extract important features from data.

[0077] In front of the first EAM and behind the second EAM, the first mean shift mechanism and the second mean shift mechanism are used to eliminate the influence of noise extreme values, and the algorithm is based on kernel density estimation (KDE), which assumes that all data points are collected from a probability distribution, KDE mainly places a kernel on each data point, then adds each kernel to form a probability surface, and the multivariate kernel density estimation formula is derived from the radially symmetric kernel, and the clustering result will be profoundly affected by the bandwidth.

[0078] In this embodiment, through a large number of experiments, the RGB mean values of the first mean shift mechanism and the second mean shift mechanism are set to 0.4488, and the variance is 1.0.

[0079] In this embodiment, the main activation function layer is a ReLU function.

[0080] As shown in Figure 2As shown, the first EAM and the second EAM are both composed of 18 layers, the first layer is an expanded convolution module, which is used to expand the receptive field of the network, and can also maintain the internal data structure while ensuring the number of parameters unchanged, the second to sixth layers are composed of convolution layers and activation function layers in turn, the seventh layer is a convolution layer, the eighth to thirteenth layers are composed of convolution layers and activation function layers in turn, the fourteenth layer is an average pooling layer, the fifteenth to seventeenth layers are composed of convolution layers and activation function layers in turn, the eighteenth layer is an activation function layer, the input of the first layer and the output of the third layer constitute a residual structure as the input of the fourth layer, the input of the fourth layer and the output of the sixth layer constitute a residual structure as the input of the seventh layer, the output of the seventh layer and the output of the twelfth layer constitute a residual structure as the input of the thirteenth layer, and the output of the thirteenth layer and the output of the eighteenth layer constitute a residual structure as the output of the first enhanced attention module.

[0081] Wherein, the residual structure is all multiplication.

[0082] The expanded convolution module is composed of a first expanded convolution layer, a second expanded convolution layer, a third expanded convolution layer, a fourth expanded convolution layer and four activation function layers, the inputs of the first expanded convolution layer and the third expanded convolution layer are both the output of the main activation function layer, the output of the first expanded convolution layer is taken as the input of the second expanded convolution layer after passing through an activation function layer, the output of the third expanded convolution layer is taken as the input of the fourth expanded convolution layer after passing through an activation function layer, the outputs of the second expanded convolution layer and the fourth expanded convolution layer are respectively taken as the input of the second layer after passing through an activation function layer, and the convolution layer of the second layer is used to connect the features of the expanded convolution module.

[0083] If multiple expanded convolutions with the same expansion factor are used, grid effect will be caused, in the image field, the grid effect refers to a large number of unprocessed pixels, and this problem may destroy the continuity and integrity of the seismic record in the field of earthquakes, therefore, four expanded convolution layers with different expansion factors are used to contain all features in the application, the expansion factor of the first expanded convolution layer is 1, the expansion factor of the second expanded convolution layer is 2, the expansion factor of the third expanded convolution layer is 3, and the expansion factor of the fourth expanded convolution layer is 4, the size of the convolution filter of the twelfth layer of the first enhanced attention module is set to 1*1, and the size of the remaining convolution filters in the denoising network based on the multiple attention mechanism is all set to 3*3.

[0084] Wherein, the convolution layers with residual structures of the fourth to sixth layers are used to learn features from data, the subsequent seventh to eleventh layers are used to compress the features by using three convolution layers to improve the calculation speed, the 1*1 convolution of the twelfth layer makes the features flat, and the last fourteenth to eighteenth layers are attention mechanisms, i.e. inter-channel attention mechanisms.

[0085] In the traditional denoising method, different channel features are usually not distinguished, which is not suitable for restoring some detail information, and more channels mean that each layer can learn more detail information, so EAM is added in the denoising network based on multiple attention mechanisms, the inter-channel attention mechanism in the internal focuses on the channel features, at the same time, by adjusting the number of passes to enhance the computing power of the network, for the GPU which is good at processing parallel computing, it is also more convenient to process multiple channels at the same time, therefore, in order to focus on the consistency of channel features, for the inter-channel attention mechanism, first of all, the global average pooling layer is selected to process the overall data.

[0086] The formula of the global average pooling layer is as follows:

[0087] y1=f sig (f up (f ReLU (f down (f avg (x1)))))×x1;

[0088] In the formula, x1 is the input of the first enhanced attention module or the second enhanced attention module, y1 is the output corresponding to the first enhanced attention module or the second enhanced attention module, f avg is an average pooling layer, f down is a down-sampling layer, f ReLU is a ReLU activation function, f up is a corresponding up-sampling layer, f sig is a Sigmoid activation function.

[0089] The above module must consider the overall nonlinear synergistic effect and mutual exclusion relationship of the channel, therefore, ReLU activation function and sigmoid activation function are selected to achieve this purpose, in addition, the output of the Sigmoid activation function will be multiplied by x1 to adjust the dimension of the data.

[0090] In the first EAM and the second EAM, a large number of residual structures are used, and it is known that the nonlinear activation function (such as ReLU) increases the fitting characteristics while adding nonlinearity to the module, but these functions reduce the dimension and cause some data to collapse, and this error cannot be completely compensated by subsequent processing, in fact, the nonlinear activation layer can remove potential sparsity, so for some data with less sparsity, it is very appropriate to select a linear activation function, therefore, the 18th layer of the first EAM and the second EAM uses the Sigmoid function as the activation function, and the other activation function layers use the ReLU function, the combination of the two functions can maintain integrity and nonlinearity, and reduce the possibility of gradient disappearance in the back propagation process, and because the convergence speed of the parameters is faster, the repeated use of features can improve the learning speed of the module.

[0091] As shown in Figure 3 , the SAM provides noisy input data that is conducive to progressive denoising, and in local denoising, an attention map is formed to eliminate less important features and pass more important features to the next stage. The supervised attention module is composed of a first convolutional layer (Conv1), a second convolutional layer (Conv2), a third convolutional layer (Conv3), and a secondary activation function layer;

[0092] where N H x N W represents the dimension of the data, N C represents the number of channels, and the output of the first EAM block is first passed through Conv2 to form a residual record A H x N W x 3, then A is multiplied by the original noisy seismic signal to obtain B H x N W x 3, then B H x N W x 3 is sequentially passed through Conv3 and the Sigmoid activation function to form an attention mask C, and the output of the first EAM block is then passed through Conv1 to form local features D, and D is multiplied by C to obtain features E, which are used to calibrate the features, and finally, the output E H x N W x N C and the output of the first EAM block is output as the output of the SAM, which is the input of the second EAM block, waiting for further processing.

[0093] In this embodiment, the secondary activation function layer uses the Sigmoid function.

[0094] The denoising network based on the multiple attention mechanism mainly emphasizes the distinction between different channels by using the inter-channel attention mechanism in the enhanced attention module, so that more comprehensive features can be extracted from the seismic record; for the supervised attention module, the input noisy record is focused on again, and the existing features are further calibrated; through the hamburger-style attention structure of "inter-channel attention mechanism-supervised attention mechanism-inter-channel attention mechanism", the features in the original data can be more effectively extracted, and the overall denoising ability of the network is improved.

[0095] 2. Training the denoising network based on the multiple attention mechanism to obtain an ideal denoising network based on the multiple attention mechanism;

[0096] In order to enhance the robustness and processing capacity of the model, the range of signal distribution is expanded as much as possible when constructing the training set of the neural network, so as to improve the processing capacity of the network.

[0097] The training set includes pure seismic signals and noise, the pure seismic signals are constructed by using the Ricker wavelet, and contain 128 recording channels, each channel has 1024 sampling points, the sampling frequency is 512Hz, and the frequency of the reflected wave is between 25Hz and 40Hz;

[0098] The noise contains 2332 channels, each channel has 59998 sampling points, in the desert noise record, a plurality of different positions are selected, and the noise is cut to the same size as the simulated pure desert signal, that is, 128 channels and 1024 sampling points;

[0099] When training using the training set, the noise multiple is adjusted to ensure that the signal-to-noise ratio of the noisy desert signal is between-2dB and-6dB.

[0100] All these values are determined according to the analysis of existing seismic records, the noise added to the pure seismic signal is the real noise collected by the sensor without any mining activity, therefore, it includes some environmental noise and background noise, and has the same statistical and time-frequency characteristics as the actual desert noise.

[0101] 3. Input the original noisy seismic signal into the first mean shift layer of the denoising network based on the multiple attention mechanism;

[0102] 4. The data processed by the main activation function layer enters the first enhanced attention module, the supervised attention module and the second enhanced attention module in turn;

[0103] 5. The second mean shift layer processes the received data and outputs to obtain a denoised seismic signal;

[0104] In order to better understand the denoising principle of seismic data, the following formula is used to explain:

[0105] M = I + X;

[0106] The present application attempts to recover the pure data from the noisy data M as much as possible, and the observed actual noise is usually simulated by zero-mean Gaussian white noise X, and I represents the pure record.

[0107] The performance of the network can be described by the following formula:

[0108]

[0109] In the formula, is the denoising result obtained after the denoising network processing, M is the noisy data, is the noise calculated by the denoising network, S is the overall denoising network structure, and μ is the parameter in the denoising network.

[0110] In order to better evaluate the denoising performance of the network, the present application uses the L1 loss function to adjust the network parameters.

[0111] In order to prove the denoising ability of the denoising network described in the present application, three denoising methods of band-pass filter, WNNM and DnCNN are selected for comparison with the denoising network described in the present application.

[0112] Firstly, the training set is constructed, and Ricker wavelet is used to construct the pure simulated seismic record. The synthetic record contains 128 record channels and 1024 sampling points, and the sampling frequency is 512Hz. The frequency of the reflected wave is randomly distributed between 25Hz and 40Hz. The noise in the pure simulated seismic record is cut and clipped from the real noise record collected from the desert environment to the same size as the pure simulated seismic record signal, and the noise multiple is adjusted to ensure that the signal-to-noise ratio of the noisy desert signal is between-2dB and-6dB.

[0113] The signal-to-noise ratios of the synthetic seismic record, the band-pass filter result, the WNNM result, the DnCNN result and the denoising network are-6.0955dB, 1.4317dB, -4.2590dB, -3.9084dB and 15.6409dB respectively.

[0114] As shown in Figures 4-9 , the band-pass filter can retain the effective signal, but cannot suppress the noise to the maximum extent; when WNNM and DnCNN lose a large amount of signal, the denoising network can both retain the signal and suppress the noise.

[0115] As shown in Figures 10-14As shown, the residual graphs corresponding to the four methods can be seen, and the existing three methods have a certain degree of signal residual in the residual graph, and all the noise is not completely suppressed, by comparing the residual of the denoising network and the noise, the two records are very similar, which verifies the denoising performance of the denoising network again.

[0116] As Figures 15-20 shown, the F-K spectrum of the pure seismic signal, the noisy signal and the four denoising methods, the F-K spectrum of the denoising result of the denoising network of the application is most similar to the F-K spectrum of the pure seismic signal, which means that the denoising network of the application can retain almost all the signal of the frequency band, while the WNNM and DnCNN will lose a large amount of signal, and the denoising network of the application can suppress noise while retaining signal.

[0117] In addition, in order to further prove the denoising ability of the denoising network of the application, more synthetic seismic records are processed using the four methods, and the processing results are shown in Tables 1 and 2, and these records are different from the above seismic records, which shows the robustness of the denoising network of the application, and the signal-to-noise ratio and mean square error (MSE) are selected to judge the performance of different methods, as shown in the table, even in some extreme cases where the signal-to-noise ratio is lower than-5dB, the denoising network of the application can suppress noise and save most of the signal.

[0118] Table 1 Comparison of signal-to-noise ratio of seven noisy records processed by four denoising methods (unit: dB)

[0119]

[0120]

[0121] Table 2 Comparison of MSE of seven noisy records processed by four denoising methods

[0122] Original data SNR (dB) Bandpass WNNM DnCNN Ours -9.7952 0.0238 0.0967 0.0228 0.0035 -7.8570 0.0162 0.0579 0.0177 0.0023 -5.3582 0.0104 0.0339 0.0152 0.0014 -3.7746 0.0081 0.0252 0.0143 0.0012 -1.8364 0.0062 0.0181 0.0133 0.0010 1.2617 0.0045 0.0114 0.0120 9.1292e-04 3.7604 0.0037 0.0091 0.0113 8.5203e-04

[0123] As Figures 21-24 shown, the single-channel comparison graph of the denoising results of the four methods is drawn, which is used to analyze the amplitude preservation ability of the four methods in detail, wherein the 57th record channel of the simulation denoising result is randomly selected to clearly see the peak value of each record channel, and the denoising results of the four methods are compared with the pure signal record to analyze the processing ability of different methods, as Figure 23 and Figure 24The peak amplitude of DnCNN in the figure is higher than that of the pure record, and the peak amplitude of the denoising network is slightly lower than that of the noise record. Because most denoising methods will cause a certain loss to the pure seismic signal while suppressing noise, considering this phenomenon, by comprehensively comparing the denoising results of the four methods, the loss of the denoising network to the signal is the smallest, and the other three methods will produce a large amount of false signals, the peak value of the denoising result is greatly different from that of the original pure seismic signal, which seriously affects the subsequent interpretation of the formation texture.

[0124] In order to prove the effectiveness of the network structure of the denoising network, ablation experiments are used to compare the denoising effects of different network structures.

[0125] Table 3 Comparison of denoising performance of different network structures (SNR unit: dB)

[0126]

[0127]

[0128] From Table 3, it can be seen that under the condition of four different SNR of noisy signals, the method can exhibit good denoising performance, among which, using only one EAM block has good effect, and with the increase of the number of EAM blocks, the denoising effect decreases. Figure 25 The data in Table 3 is a line graph, from which it can be more clearly seen that the denoising network has better denoising performance than other network structures in the ablation experiment.

[0129] In addition, in order to further prove the denoising performance of the application, the single channel of the denoising results of seven network structures is analyzed, and the single channel of the denoising results of the seven network structures is compared with the single channel of the pure signal, and the results are shown in Figures 26-31 respectively, wherein the single channel comparison selects the denoising result corresponding to the noisy signal of-5.1330dB, and selects the 53rd record channel in the data.

[0130] From Figure 26 it can be seen that one SAM structure has poor signal peak recovery effect; from Figure 27 it can be seen that one EAM structure is prone to over-recovery phenomenon, and the peak value of the denoising result is greater than that of the pure signal, and is seriously affected by noise; from Figure 28 it can be seen that two EAM structures are seriously affected by noise, and the parts that should be 0 values are all treated as positive values; from Figure 29 it can be seen that three EAM structures are seriously affected by noise, and the parts that should be 0 values are all treated as negative values; from Figure 30As can be seen, the four EAM structures are prone to over-recovery, with the denoising results showing higher trough values ​​than the clean signal, and are significantly affected by noise; in contrast, Figure 31 The method described in this paper achieves good recovery results, further demonstrating the denoising performance of the denoising network described in this invention.

[0131] To ensure the practical application of this invention, a record of a public firing point obtained in a desert region of China is used, such as... Figure 32 As shown, the characteristics of this field reflection record are as follows: 121 channels, 801 samples, 2000 μs sample interval, many tiny reflection events are covered by severe random noise, and there are a large number of weak signals that are difficult to extract in this record.

[0132] like Figures 33-36 As shown, comparing the denoising results of the four methods, the denoising network described in this invention can extract more weak signals, while the other three methods lose weak signals to some extent. Furthermore, in deeper layers where the signal is weak, covered by strong noise, and difficult to identify, the denoising network described in this invention can extract more weak signals compared to the other three methods. Bandpass filters cannot retain the lower and higher frequency components of the recording, which will seriously affect the accuracy of subsequent inversion work. WNNM severely reduces the resolution of the recording, retaining too few signals to identify the overall structure of the ground from the recording. In the areas highlighted by the blocks in the recording, the denoising network described in this invention can recover more weak signals compared to other methods. Although DnCNN performs well in denoising analog recordings, it cannot extract weak signals in actual recordings. In the areas highlighted by the black arrows in the recording, the signal continuity recovered by the denoising network described in this invention is better than other methods.

[0133] This invention presents a desert earthquake denoising method based on a multi-attention mechanism. Two attention modules are added to recover weak signals from noisy data. The enhanced attention module incorporates a soft-threshold mechanism within its channel attention mechanism to emphasize more important parts of the attention feature map. The supervised attention module relearns the structure in the noisy data, allowing only the most important features to proceed to the next stage. This hamburger-like structure—"inter-channel attention mechanism – supervised attention mechanism – inter-channel attention mechanism"—formed by the two attention mechanisms allows the network to extract effective features from the original data from different aspects, thereby enabling the network to extract some weak signals masked by noise. Denoising results using simulated and real data verify the denoising performance of the denoising network described in this invention. Compared with three other methods, this method exhibits superior denoising performance, shorter training time, and better signal preservation.

[0134] While embodiments of the application have been disclosed in connection with the above specification, it will be apparent to those skilled in the art that numerous modifications can be made thereto without departing from the overall concept of the application. Accordingly, it is intended that all such modifications be included within the scope of the claims and their equivalents.

Claims

1. A desert earthquake denoising method based on a multiple attention mechanism, characterized in that, Includes the following steps: Step 1: Construct an initial denoising network based on a multiple attention mechanism; The denoising network based on the multiple attention mechanism includes a first mean shift layer, a convolutional layer, a main activation function layer, a first enhanced attention module, a supervised attention module, a second enhanced attention module, a convolutional layer, and a second mean shift layer connected in sequence. Both the first and second enhanced attention modules consist of an 18-layer structure. The first layer is a dilated convolutional module, layers 2-6 are composed of convolutional layers and activation function layers alternately, the seventh layer is a convolutional layer, layers 8-13 are composed of convolutional layers and activation function layers alternately, the 14th layer is an average pooling layer, layers 15-17 are composed of convolutional layers and activation function layers alternately, and the 18th layer is an activation function layer. The input of the first layer and the output of the third layer form a residual structure as the input of the fourth layer, the input of the fourth layer and the output of the sixth layer form a residual structure as the input of the seventh layer, the output of the seventh layer and the output of the 12th layer form a residual structure as the input of the 13th layer, and the output of the 13th layer and the output of the 18th layer form a residual structure as the output of the first enhanced attention module. The dilated convolution module consists of a first dilated convolution layer, a second dilated convolution layer, a third dilated convolution layer, and a fourth dilated convolution layer. The inputs of the first and third dilated convolution layers are both the outputs of the main activation function layer. The output of the first dilated convolution layer serves as the input of the second dilated convolution layer, the output of the third dilated convolution layer serves as the input of the fourth dilated convolution layer, and the outputs of the second and fourth dilated convolution layers serve as the input of the second layer. The dilation factor of the first dilated convolution layer is 1, the dilation factor of the second dilated convolution layer is 2, the dilation factor of the third dilated convolution layer is 3, and the dilation factor of the fourth dilated convolution layer is 4. The size of the convolutional filter in the 12th layer of the first enhanced attention module is set to 1×1, and the size of the remaining convolutional filters in the denoising network based on the multiple attention mechanism is set to 3×3. Step 2: Train the initial denoising network based on the multiple attention mechanism to obtain the ideal denoising network based on the multiple attention mechanism; Step 3: Input the original noisy seismic signal into the first mean-shift layer of the ideal denoising network based on the multiple attention mechanism; Step 4: The data processed by the main activation function layer sequentially enters the first enhanced attention module, the supervised attention module, and the second enhanced attention module; Step 5: The second mean shift layer processes the received data and outputs the denoised seismic signal.

2. The desert earthquake denoising method based on a multiple attention mechanism as described in claim 1, characterized in that, The dilated convolution module also includes: Multiple activation function layers are set one-to-one at the output of the first dilated convolutional layer, the second dilated convolutional layer, the third dilated convolutional layer, and the fourth dilated convolutional layer.

3. The desert earthquake denoising method based on a multiple attention mechanism as described in claim 2, characterized in that, The supervised attention module consists of a first convolutional layer, a second convolutional layer, a third convolutional layer, and a sub-activation function layer; Specifically, the output of the first enhanced attention module serves as the input to the first and second convolutional layers. The output of the second convolutional layer, together with the original noisy seismic signal, forms a residual structure that serves as the input to the third convolutional layer. The output of the third convolutional layer serves as the input to the sub-activation function layer. The output of the sub-activation function layer, together with the output of the first convolutional layer, forms a first-level residual structure, which, together with the output of the first enhanced attention module, forms a second-level residual structure that serves as the output of the supervised attention module.

4. The desert earthquake denoising method based on a multiple attention mechanism as described in claim 3, characterized in that, The secondary activation function layer is the Sigmoid function.

5. The desert earthquake denoising method based on a multiple attention mechanism as described in claim 4, characterized in that, The 18th layer of the first and second enhanced attention modules is a Sigmoid function, and the other activation function layers in the first and second enhanced attention modules are ReLU functions.

6. The desert earthquake denoising method based on a multiple attention mechanism as described in claim 5, characterized in that, The main activation function layer is the ReLU function.

7. The desert earthquake denoising method based on a multiple attention mechanism as described in claim 6, characterized in that, The formula for calculating the average pooling layer is: y1=f sig (in up (in ReLU (in down (in avg (x1)))))×x1; In the formula, x1 is the input of the first or second enhanced attention module, y1 is the corresponding output of the first or second enhanced attention module, and f avg For average pooling layers, f down For downsampling layer, f ReLU f is the ReLU activation function. up For the corresponding upsampling layer, f sig This is the Sigmoid activation function.

8. The desert earthquake denoising method based on a multiple attention mechanism as described in claim 7, characterized in that, The performance of the ideal denoising network based on the multiple attention mechanism satisfies: In the formula, The denoising result is obtained after processing by the denoising network, where M is the noisy data. The noise is calculated by the denoising network, S is the overall denoising network structure, and μ is the parameter in the denoising network.

9. The desert earthquake denoising method based on a multiple attention mechanism as described in claim 8, characterized in that, The parameters in the denoising network are obtained by adjusting the L1 loss function.

10. The desert earthquake denoising method based on a multiple attention mechanism as described in claim 9, characterized in that, The training set used to train the initial denoising network based on the multiple attention mechanism includes clean seismic signals and noise; The pure seismic signal is synthesized from Ricker wavelets and contains 128 recording channels, each with 1024 sampling points and a sampling frequency of 512Hz. The noise was extracted from real noise records collected in the desert environment, containing 2,332 channels, each with 59,998 sampling points, and the noise was trimmed to the same size as the pure seismic signal; When using the training set, adjust the noise factor so that the signal-to-noise ratio of the noisy desert signal is between -2dB and -6dB.

Images