Microseismic profile denoising method based on time-frequency domain combination

Abstract

The invention relates to a microseismic profile denoising method based on time-frequency domain combination, which comprises the following steps of: constructing a microseismic signal denoising model comprising a WNet sub-network and a TNet sub-network, inputting an acquired time sequence model into the WNet sub-network and the TNet sub-network, extracting features through a space self-attention module STAB, fusing the features through an adaptive attention fusion module AAFB, and carrying out time-frequency domain combination. The method comprises the steps of obtaining a trained micro-seismic signal denoising model through time-frequency information fusion and jump connection, and finally inputting a real micro-seismic signal profile map to be denoised into the trained micro-seismic signal denoising model to obtain a denoised micro-seismic profile map.

Description

Technical Field

[0001] The present invention relates to the technical field of microseismic monitoring, and in particular to a microseismic profile denoising method based on time-frequency domain combination. Background Art

[0002] Microseismic monitoring technology is a new method developed in recent years and is mainly used for oil exploration. Random noise suppression is an important stage in microseismic signal processing. Improving the signal-to-noise ratio (SNR) of microseismic signals can improve subsequent data analysis and processing. Due to the complexity of random noise, existing microseismic signal denoising methods are difficult to achieve a good effect between removing random noise and restoring effective signals. Summary of the invention

[0003] The technical problems of the present invention are:

[0004] Microseismic profile signal denoising is difficult to recover weak effective signals from strong background noise, and the time domain and frequency domain dimensions of the signal in the microseismic signal profile denoising task are different. The existing technology lacks a microseismic profile signal denoising method that combines time domain and frequency domain.

[0005] To solve the above problems, the present invention establishes a time-frequency domain joint microseismic profile denoising network model. By performing discrete wavelet transform on the signal, the maximum pooling operation in the general UNet network is replaced. Due to the time-frequency characteristics of the wavelet transform, this method can also take into account the frequency characteristics of the signal; three loss functions are designed and combined, namely, the time series information function, the frequency domain information function and the time-frequency fusion information function, to ensure that the network can effectively fuse time-frequency information in the same dimension; a spatial self-attention module STAB based on the Transformer encoding structure is designed to enhance the feature extraction capability and replace the jump connection of the ordinary UNet network, as well as the adaptive attention fusion module AAFB for downsampling and upsampling features to improve the feature fusion effect, thereby enhancing the signal-to-noise separation capability of low signal-to-noise ratio microseismic signals.

[0006] In order to achieve the above technical effects, the present invention adopts the following technical solutions:

[0007] In the WNet subnetwork, discrete wavelet transform is used to decompose the time series signal and inverse wavelet transform is used to reconstruct the up-sampled result to reduce the dimension of the time series signal; the spatial transformer attention module STAB is used to extract frequency domain features and time domain features; the adaptive attention fusion module AAFB is used to fuse frequency domain features and time domain features;

[0008] The microseismic profile denoising method based on the joint time-frequency domain includes the following steps:

[0009] S1: In the selected exploration terrain, simulated microseismic signals are collected to obtain data set D1, noise of the real work area is added to data set D1 to obtain data set D2, and after preprocessing data set D2, it is divided into training set, validation set and test set according to a fixed ratio;

[0010] S2: constructing a microseismic signal denoising model, the microseismic signal denoising model comprising: a TNet sub-network for extracting time domain features and denoising, a WNet sub-network for extracting frequency domain features and denoising, and a Fusion sub-network for fusing time domain features and frequency domain features;

[0011] S3: Build TNet sub-network, WNet sub-network and Fusio respectively n The loss function of the sub-network;

[0012] S4: training and testing the microseismic signal denoising model using the training set, validation set and test set obtained in step S1 until the trained microseismic signal denoising model meets the accuracy requirement;

[0013] S5: Input the real microseismic signal profile to be denoised into the trained microseismic signal denoising model to obtain the denoised microseismic signal profile.

[0014] Furthermore, in step S1, the data set D1 uses a finite difference wave equation forward modeling method to generate simulation signals under different main frequencies and formation velocity models.

[0015] Preferably, in step S3, the time series signal is B one-dimensional signals of length L, and the tensor is (B, 1, L). It is decomposed by discrete wavelet transform, including decomposing 1 layer using Dobesi wavelet function and using periodic extension to obtain tensor (B, 2, 2 / L) of low-frequency approximation coefficients and high-frequency detail coefficients as input of WNet.

[0016] Preferably, the calculation formula for discrete wavelet transform decomposition is:

[0017] S input =DWT(X input );

[0018] Where DWT stands for discrete wavelet transform; S input represents the input of WNet, including high-frequency information and low-frequency information; X input Indicates the timing signal input.

[0019] Furthermore, in step S3, the WNe subnetwork includes a downsampling layer, a convolutional layer, an upsampling layer, a spatial self-attention module STAB, and an adaptive attention fusion module AAFB;

[0020] The calculation formula for the downsampling layer is:

[0021] W down =DWT(COnv(S input ));

[0022] Where W down Represents the time-frequency information S input Down sampling is performed, Conv means convolution;

[0023] The calculation formula of the convolution layer is:

[0024] W bottleneck =Conv(W down );

[0025] Where W bottleneck Represents the output after convolution;

[0026] The calculation formula of the spatial self-attention module STAB is:

[0027] W extract =STAB(W down );

[0028] Where W extract represents the output of feature extraction, STAB represents the spatial self-attention module;

[0029] Upsampling calculation formula and adaptive attention fusion module AAFB:

[0030] S up =AAFB(W extract ,UpSampling(W down ));

[0031] In the formula, S up Represents the output of the adaptive attention fusion module, UpSampling means reconstructing the features using upsampling and convolution, and AAFB is the adaptive attention fusion module.

[0032] Preferably, in step S3, the TNe subnetwork includes a downsampling layer, a feature extraction layer, an upsampling layer, a spatial self-attention module STAB, and an adaptive attention fusion module AAFB.

[0033] The downsampling calculation formula is:

[0034] T down =Conv(Maxpooling(X input ));

[0035] Where, T down It represents downsampling of time series signals, and Maxpooling represents maximum pooling;

[0036] The calculation formula of the spatial self-attention module STAB is:

[0037] T extract =STAB(T down );

[0038] Where, T down represents the spatial self-attention module feature extraction, STAB represents the Transformer-based spatial attention feature extraction module;

[0039] Upsampling calculation formula and adaptive attention fusion module AAFB:

[0040] X up =AAFB(T extract ,UpSampling(T down ));

[0041] Where, X up It represents the output of the fusion stage, UpSampling represents the reconstruction of features using upsampling and convolution, and AAFB represents the module for adaptively fusing the two parts of features;

[0042] Furthermore, in step S3, the Fusion sub-network fuses the two sub-networks WNet and TNet, including performing inverse wavelet transform on the sampling result on the WNet sub-network, reconstructing the inverse transform on the sampling result on the TNet sub-network, fusing features and concatenating features;

[0043] 1) The calculation formula for the inverse wavelet transform of the sampling result on the WNet sub-network is:

[0044] W up =IDWT(S up );

[0045] Where W up represents the output of WNet sub-network fusion, and IDWT represents the inverse wavelet transform of the up-sampled result;

[0046] 2) Perform inverse transformation on the sampled result of the TNet sub-network to reconstruct the calculation formula:

[0047] T up =UpSampling(X up );

[0048] Where, T up Represents the output of TNet sub-network fusion;

[0049] 3) The calculation formula of fusion features is:

[0050] F o =UpSampling(AAFB(Wup , T up ));

[0051] In the formula, F o It represents the output after fusing the two parts of features and then upsampling and convolution. AAFB fuses the output of the WNet sub-network fusion stage with the output of the TNet sub-network in the fusion stage.

[0052] 4) The calculation formula of splicing features is:

[0053] F output =UpSampling(Conv(CAT(F o ,AAFB(W up , T up ))));

[0054] F output It represents the final output after the fusion of the two sub-networks, UpSampling represents the reconstruction of the features using upsampling and convolution, and CAT represents the concatenation of the two parts of the features.

[0055] Preferably, the feature extraction module STAB in step S3 is composed of a one-dimensional spatial self-attention module, normalization and a multi-layer perceptron, including position encoding embedding, normalization and spatial self-attention operations on the feature output, and the calculation formula is:

[0056] F e =Add(Emb(f),SSA(LN(Emb(f))));

[0057] F ex =Add(F e ,MLP(LN(F e )));

[0058] Where f is the input of STAB; Add represents the skip connection operation; Emb is the positional encoding embedding operation; LN is layer normalization; SSA is spatial self-attention; F e represents the output after skip connection; MLP represents multi-layer perceptron; F ex Represents the output of the STAB module.

[0059] Preferably, the skip connection expression is:

[0060]

[0061] Where Skipi represents the calculation method of the i-th skip connection, The dimension of Skipi is L represents the specific length of the input, i=1, 2, 3.

[0062] Furthermore, C1, C2, and C3 are set as the convolution kernel size of each downsampling layer;

[0063] X s =SSA(Q, K, V);

[0064] Q=W q X;

[0065] K=W k X;

[0066] V=W v X:

[0067] Where, X s It represents the spatial self-attention feature map output by SSA processing. SSA represents spatial self-attention. Q, K, and V are the inputs of SSA, respectively, through the learned weight matrix W q , W k , W v Calculated with input X.

[0068] Furthermore, the calculation formula of spatial self-attention is:

[0069]

[0070] Y s represents the feature output extracted by spatial self-attention. Q and K T are the transpose of the quires matrix and the keys matrix of spatial self-attention, and V is the values ​​matrix of spatial self-attention. The activation function softmax represents the normalization of the input vector, d k Indicates the magnitude of a vector.

[0071] Furthermore, in the adaptive attention fusion module AAFB in step S3, the time series signal is divided into two parts after downsampling, one part is subjected to convolution and upsampling operations, and the other part is subjected to feature extraction operation by the STAB module. The features of the two parts of the output are subjected to weighted attention fusion using the AAFB module, and the calculation formula is:

[0072] G=Mul(D,Sigmoid(AdaptConv(GAP(D,U))));

[0073] In the formula, G represents the weight output for feature fusion, Mul represents the multiplication of D and Sigmoid, Sigmoid represents a nonlinear activation function, AdaptConv represents an adaptive convolution operation, GAP represents global average pooling, D represents the result after feature extraction by the STAB module, and U represents the result after feature fusion using AAFB in the upsampling stage.

[0074] Preferably, step S3 also includes a loss function of the training process in S3, including a coefficient loss function, a timing loss function, and a time-frequency fusion loss function. The coefficient loss function, the timing loss function, and the time-frequency fusion loss function are used to constrain the WNet subnetwork, the TNet subnetwork, and the Fusion subnetwork, respectively. The loss function calculation formula is:

[0075] Loss = λ1loss w +λ2loss t +λ3loss tw ;

[0076] In the formula, loss w 、loss t 、loss tw They are coefficient loss function, time series loss function and time-frequency fusion loss function respectively.

[0077] Compared with the prior art, the beneficial effects of the present invention include:

[0078] 1) The present invention proposes a microseismic profile denoising method based on the joint time-frequency domain, and constructs a microseismic signal denoising model including a WNet subnetwork, a TNet subnetwork and a Fusion subnetwork. The collected seismic time series signals are subjected to feature extraction by the WNet subnetwork and the TNet subnetwork respectively for different characteristics of the microseismic signals, and then the features are extracted by the spatial self-attention module STAB, the features are fused by the adaptive attention fusion module AAFB, the time-frequency information is fused and jump-connected, so as to obtain a trained microseismic signal denoising model. Finally, the real microseismic signal profile to be denoised is input into the trained microseismic signal denoising model to obtain the denoised microseismic profile.

[0079] 2) The present invention performs discrete wavelet transform on the signal to replace the maximum pooling operation in the general UNet network, thereby ensuring that the network model of the present invention can effectively fuse time-frequency information in the same dimension.

[0080] 3) The spatial self-attention module STAB based on the Transformer encoding structure of the present invention enhances the feature extraction capability.

[0081] 4) The adaptive attention fusion module AAFB for downsampling and upsampling features of the present invention improves the feature fusion effect and enhances the signal-to-noise separation capability of low signal-to-noise ratio microseismic signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0083] Figure 1 A schematic diagram of a microseismic profile denoising method according to an embodiment of the present invention;

[0084] Figure 2 It is a structural diagram of a microseismic profile denoising model based on time-frequency domain combination according to an embodiment of the present invention;

[0085] Figure 3 It is a structural diagram of a feature extraction module STAB according to an embodiment of the present invention;

[0086] Figure 4 is a structural diagram of an AAFB module according to an embodiment of the present invention;

[0087] Figure 5 A microseismic signal diagram according to an embodiment of the present invention;

[0088] Figure 6 This is a graph of the Unet network denoising results of an embodiment of the present invention;

[0089] Figure 7 This is a denoising result diagram of the microseismic signal denoising model according to an embodiment of the present invention;

[0090] Figure 8 It is a time-frequency diagram of a microseismic signal according to an embodiment of the present invention;

[0091] Fig. 9 This is a time-frequency diagram after denoising of the Unet network in an embodiment of the present invention;

[0092] Fig.10 It is a time-frequency diagram of the denoising result of the microseismic signal denoising model according to an embodiment of the present invention;

[0093] Fig.11 A comparison diagram of the spectrum energy of a single-channel signal after denoising by Clean, Unet and the present invention in an embodiment of the present invention;

[0094] Fig.12 A cross-sectional diagram of a noisy microseismic signal according to an embodiment of the present invention;

[0095] Fig.13 This is a cross-sectional diagram of a microseismic signal after Unet denoising according to an embodiment of the present invention;

[0096] Fig.14 This is a cross-sectional view of the microseismic signal after denoising using the method of the present invention. DETAILED DESCRIPTION

[0097] like Figure 1 As shown in FIG. 1 , the microseismic profile denoising method based on the time-frequency domain combination includes the following steps:

[0098] S1: In the selected exploration terrain, simulated microseismic signals are collected to obtain data set D1, noise of the real work area is added to data set D1 to obtain data set D2, and after preprocessing data set D2, it is divided into training set, validation set and test set according to a fixed ratio;

[0099] S2: constructing a microseismic signal denoising model, the microseismic signal denoising model comprising: a TNet sub-network for extracting time domain features and denoising, a WNet sub-network for extracting frequency domain features and denoising, and a Fusion sub-network for fusing time domain features and frequency domain features;

[0100] S3: Construct the loss functions of TNet sub-network, WNet sub-network and Fusion sub-network respectively;

[0101] S4: training and testing the microseismic signal denoising model using the training set, validation set and test set obtained in step S1 until the trained microseismic signal denoising model meets the accuracy requirement;

[0102] S5: Input the real microseismic signal profile to be denoised into the trained microseismic signal denoising model to obtain the denoised microseismic signal profile.

[0103] The network structure of microseismic profile denoising based on time-frequency domain combination is shown in the figure. Figure 2 shown.

[0104] Furthermore, in step S1, the data set D1 uses a finite difference wave equation forward modeling method to generate simulation signals under different main frequencies and formation velocity models.

[0105] Preferably, in step S2, the time series signal is B one-dimensional signals of length L, and the tensor is (B, 1, L). It is decomposed by discrete wavelet transform, including decomposing 1 layer using Dobesi wavelet function and using periodic extension to obtain a tensor of (B, 2, 2 / L) low-frequency approximation coefficients and high-frequency detail coefficients as the input of WNet.

[0106] Preferably, the calculation formula for discrete wavelet transform decomposition is:

[0107] S input =DWT(X input );

[0108] Where DWT stands for discrete wavelet transform; S input represents the input of WNet, including high-frequency information and low-frequency information; X input Indicates the timing signal input.

[0109] Further, such as Figure 2As shown, in step S3, the WNe subnetwork includes 3 downsampling layers, each of which includes 2 convolutional layers; 3 upsampling layers, each of which includes 2 convolutional layers; 3 spatial self-attention modules STAB and 3 adaptive attention fusion modules AAFB;

[0110] The calculation formula for the downsampling layer is:

[0111] W down,i =DWT(Conv(S input ));

[0112] Where W down，i Represents the time-frequency information S for the i-th time input Downsampling, i=1,2,3; Conv means convolution;

[0113] The calculation formula of the spatial self-attention module STAB is:

[0114] W extract,i =STAB(W down，i );

[0115] Where W extract represents the output of the i-th feature extraction, i = 1, 2, 3; STAB represents the spatial self-attention module;

[0116] The calculation formula of upsampling and adaptive attention fusion module AAFB is:

[0117] S up，i =AAFB(W extract，i ,UpSampling(W down,i ));

[0118] In the formula, S up，i represents the output of the i-th adaptive attention fusion module, i = 1, 2, 3; UpSampling represents the upsampling;

[0119] Preferably, in step S3, the TNe subnetwork includes 3 downsampling layers, each of which includes 2 convolutional layers; 3 upsampling layers, each of which includes 2 convolutional layers; 3 spatial self-attention modules STAB and 3 adaptive attention fusion modules AAFB;

[0120] The downsampling calculation formula is:

[0121] T down,i =Conv(Maxpooling(X input ));

[0122] Where, T down represents the i-th downsampling of the time series signal, i = 1, 2, 3; Maxpooling represents maximum pooling;

[0123] The calculation formula of the spatial self-attention module STAB is:

[0124] T extract,i =STAB(T down，i );

[0125] Where, T down represents the feature extraction of the i-th spatial self-attention module, and STAB represents the spatial self-attention module;

[0126] The calculation formula of upsampling and adaptive attention fusion module AAFB is:

[0127] X up，i =AAFB(T extract，i ,UpSampling(T down,i ));

[0128] Where, X up，i represents the i-th adaptive attention fusion output, i = 1, 2, 3; AAFB represents the adaptive attention fusion module, and UpSampling represents upsampling;

[0129] Furthermore, in step S3, the Fusion sub-network fuses the two sub-networks WNet and TNet, including three wavelet inverse transforms on the sampling results on the WNet sub-network, three inverse transform reconstructions on the sampling results on the TNet sub-network, and three fusion features and splicing features;

[0130] 1) The calculation formula for the inverse wavelet transform of the sampling result on the WNet sub-network is:

[0131] W up，i =IDWT(S up，i );

[0132] Where W up，i represents the output of the i-th fusion of the WNet sub-network, i = 1, 2, 3; IDWT represents the inverse wavelet transform of the up-sampled result;

[0133] 2) Perform an inverse transformation on the sampling result of the TNet sub-network to reconstruct the calculation formula:

[0134] T up，i =UpSampling(X up，i );

[0135] Where, T up represents the output of the i-th fusion stage of the TNet sub-network, i = 1, 2, 3; UpSampling represents the use of upsampling and convolution to achieve feature reconstruction;

[0136] 3) The calculation formula of fusion features is:

[0137] F o =UpSampling(AAFB(W up，i , T up，i ));

[0138] In the formula, F o It represents the output after fusing the two parts of features and then upsampling and convolution. AAFB fuses the output of the WNet sub-network fusion stage with the output of the TNet sub-network in the fusion stage.

[0139] 4) The calculation formula of splicing features is:

[0140] F output =UpSampling(Conv(CAT(F o ,AAFB(W up , T up ))));

[0141] F output represents the final output after the fusion of the two sub-networks, UpSampling represents the reconstruction of features using upsampling and convolution, and CAT represents the reconstruction of F o And the feature concatenation after the fusion of the two sub-networks.

[0142] Preferably, Figure 3 As shown, the feature extraction module STAB in step S3 consists of a one-dimensional spatial self-attention module, normalization and a multi-layer perceptron, including three operations of position encoding embedding, normalization and spatial self-attention on the feature output. The calculation formula is:

[0143] F e，i =Add(Emb(f i ), SSA(LN(Emb(f i ))));

[0144] F ex，i =Add(F e，i ,MLP(LN(F e，i )));

[0145] In the formula, f i is the i-th input of STAB; Add represents the skip connection operation; Emb is the positional encoding embedding operation; LN is layer normalization; SSA is spatial self-attention; F e，i represents the output after the i-th skip connection, i = 1, 2, 3; MLP represents multi-layer perceptron; F ex，i Represents the output of the STAB module after the i-th time, i=1.2,3.

[0146] Preferably, the skip connection expression is:

[0147]

[0148] Where Skipi represents the calculation method of the i-th skip connection, The dimension of Skipi is L represents the specific length of the input, i=1, 2, 3.

[0149] Furthermore, C1=64, C2=128, and C3=256 are set as the convolution kernel size of each downsampling layer;

[0150] Y s =SSA(Q, K, V);

[0151] Q=W q X;

[0152] K=W k X;

[0153] V=W v X:

[0154] Where Y s represents the feature output of spatial self-attention extraction, SSA represents spatial self-attention, Q, K, V are the inputs of SSA, respectively, through the learned weight matrix W q , W k , W v Calculated with input X, where X is the input of the STAB module;

[0155] Furthermore, the calculation formula of spatial self-attention is:

[0156]

[0157] Y s represents the feature output extracted by spatial self-attention. Q and K T are the transpose of the quires matrix and the keys matrix of spatial self-attention, and V is the values ​​matrix of spatial self-attention. The activation function softmax represents the normalization of the input vector, d k Indicates the magnitude of a vector.

[0158] Further, such as Figure 4 As shown, in the adaptive attention fusion module AAFB in step S3, the time series signal is divided into two parts after downsampling, one part is subjected to convolution and upsampling operations, and the other part is subjected to feature extraction operation by the STAB module. The features of the two parts of the output are weighted attention fused using the AAFB module, and the calculation formula is:

[0159] G(i)=Mul(D,Sigmoid(AdaptConv(GAP(D i , U i ))));

[0160] Where G(i) represents the weight output of the i-th feature fusion, i = 1, 2, 3; Mul represents the weight of D i Multiply the result with Sigmoid; Sigmoid represents a nonlinear activation function; AdaptConv represents an adaptive convolution operation; GAP represents global average pooling; D i represents the result after the feature extraction by STAB module for the i-th time, i=1, 2, 3; U i It represents the result of feature fusion after using AAFB for the i-th time in the upsampling stage, i=1.2,3.

[0161] Preferably, step S3 also includes a loss function of the training process in S3, including a coefficient loss function, a timing loss function, and a time-frequency fusion loss function. The coefficient loss function, the timing loss function, and the time-frequency fusion loss function are used to constrain the WNet subnetwork, the TNet subnetwork, and the Fusion subnetwork, respectively. The loss function calculation formula is:

[0162] Loss = λ1loss w +λ2loss t +λ3loss tw ;

[0163] In the formula, loss w 、loss t 、loss tw They are coefficient loss function, time series loss function and time-frequency fusion loss function respectively.

[0164] In order to test the effectiveness of the method proposed in this invention, simulated microseismic signals were collected under real geological conditions and exploration backgrounds, and real work area noise was selected and added to obtain noise signals, such as Figure 5 As shown; the result diagrams of denoising by Unet network and denoising by the microseismic signal denoising model of the present invention are respectively as shown Figure 6 and Figure 7 As shown;

[0165] In order to verify the performance of the present invention in the frequency domain, the time-frequency diagram of the microseismic signal was collected, such as Figure 8 As shown; the result diagrams of denoising by Unet network and denoising by the microseismic signal denoising model of the present invention are respectively as shown Fig. 9 and Fig.10 As shown;

[0166] Through the above comparison, the microseismic signal denoising model of the present invention is superior to the UNet network in terms of both time sequence and frequency for microseismic signal denoising.

[0167] like Fig.11 As shown, the single-channel signal spectrum energy diagram of the UNet network still retains some noise at around 50Hz, and some effective signals are removed; the single-channel signal spectrum energy diagram of the present invention can be close to the clean signal in frequency and amplitude;

[0168] For profile signals containing a lot of noise, such as Fig.12 As shown, the microseismic signal profiles of the UNet network and the microseismic signal denoising model of the present invention are respectively as shown in Fig.13 and Fig.14 As shown, the microseismic signal denoising model of the present invention recovers more effective signals from the background noise and maintains the continuity and integrity of the microseismic signal.

Claims

1. A microseismic profile denoising method based on the joint time-frequency domain, characterized in that: In the WNet sub-network, discrete wavelet transform is used to decompose the time series signal and inverse wavelet transform is used to reconstruct the up-sampled result to reduce the dimension of the time series signal. Use the spatial transformer attention module STAB to extract frequency domain features and time domain features; Use the adaptive attention fusion module AAFB to fuse frequency domain features and time domain features; S1: In the selected exploration terrain, simulated microseismic signals are collected to obtain a data set D1, noise of the real work area is added to the data set D1 to obtain a data set D2, and after preprocessing the data set D2, it is divided into a training set, a validation set, and a test set according to a fixed ratio; S2: constructing a microseismic signal denoising model, the microseismic signal denoising model comprising: a TNet sub-network for extracting time domain features and denoising, a WNet sub-network for extracting frequency domain features and denoising, and a Fusion sub-network for fusing time domain features and frequency domain features; S3: Construct the loss functions of TNet sub-network, WNet sub-network and Fusion sub-network respectively; S4: training and testing the microseismic signal denoising model using the training set, validation set and test set obtained in step S1 until the trained microseismic signal denoising model meets the accuracy requirement; S5: Input the real microseismic signal profile to be denoised into the trained microseismic signal denoising model to obtain the denoised microseismic signal profile.

2. The microseismic profile denoising method based on time-frequency domain combination according to claim 1 is characterized in that: In step S1, the data set D1 uses a finite difference wave equation forward modeling method to generate simulation signals under different main frequencies and formation velocity models.

3. The microseismic profile denoising method based on time-frequency domain combination according to claim 2 is characterized in that: In step S3, the time series signal is decomposed by discrete wavelet transform, including decomposition using Dobesi wavelet function, and periodic extension is adopted to obtain low-frequency approximate coefficients and high-frequency detail coefficients as input of the WNet sub-network.

4. The microseismic profile denoising method based on time-frequency domain combination according to claim 3 is characterized in that: The calculation formula of the discrete wavelet transform decomposition is: S input =DWT(X input ); Where DWT stands for discrete wavelet transform; S input represents the input of WNet, including high-frequency information and low-frequency information; X input Indicates the timing signal input.

5. The microseismic profile denoising method based on time-frequency domain combination according to claim 4 is characterized in that: In step S3, the WNe subnetwork includes a downsampling layer, a convolutional layer, an upsampling layer, a spatial self-attention module STAB, and an adaptive attention fusion module AAFB; 1) The calculation formula of the downsampling layer is: W down =DWT(Conv(S input )); Where W down Represents the time-frequency information S input Down sampling is performed, Conv means convolution; 2) The calculation formula of the convolution layer is: W bottleneck =Conv(W down ); Where W bottleneck Represents the output after convolution; 3) The calculation formula of the spatial self-attention module STAB is: W extract =STAB(W down ); Where W extract represents the output of feature extraction, STAB represents the spatial self-attention module; 4) Upsampling and adaptive attention fusion module AAFB calculation formula: S up =AAFB(W extract ,UpSampling(W down )); Where W extract represents the output of feature extraction, UpSampling represents feature reconstruction using upsampling and convolution, and W down Represents the time-frequency information S input Downsampling is performed, and AAFB represents the adaptive attention fusion module.

6. The microseismic profile denoising method based on time-frequency domain combination according to claim 5 is characterized in that: In step S3, the TNe subnetwork includes a downsampling layer, a feature extraction layer, an upsampling layer, a spatial self-attention module STAB, and an adaptive attention fusion module AAFB; 1) The downsampling calculation formula is: T down =Conv(Maxpooling(X input )); Where, T down It represents downsampling of time series signals, and Maxpooling represents maximum pooling; 2) The calculation formula of the spatial self-attention module STAB is: T extract =STAB(T down ); Where, T down represents the spatial self-attention module feature extraction, STAB represents the Transformer-based spatial attention feature extraction module; 3) Upsampling and adaptive attention fusion module AAFB calculation formula: X up =AAFB(T extract ,UpSampling(T down )); Where, X up It represents the output of the fusion stage, AAFB represents the module that adaptively fuses the two parts of features, and UpSampling represents the reconstruction of features using upsampling and convolution.

7. The microseismic profile denoising method based on time-frequency domain combination according to claim 6 is characterized in that: In step S3, the Fusion sub-network fuses the extracted features of the WNet and TNet sub-networks, including performing inverse wavelet transform on the sampling results of the WNet sub-network, reconstructing the features of the TNet sub-network by upsampling, fusing features, and concatenating features; The calculation formula for the inverse wavelet transform of the sampling results on the WNet sub-network is: W up =IDWT(S up ); Where W up Represents the output of the WNet sub-network fusion stage; The calculation formula for feature reconstruction using upsampling on TNet is: T up =UpSampling(X up ); Where, T up represents the output of the fusion stage; The fusion feature calculation formula is: F o =UpSampling(AAFB(W up ,T up )); In the formula, F o It represents the output after fusing the two parts of features and then upsampling and convolution. The adaptive attention fusion module AAFB fuses the output of the WNet sub-network fusion stage and the output of the TNet sub-network in the fusion stage; The calculation formula of splicing features is: F output =UpSampling(Conv(CAT(F o ,AAFB(W up ,T up )))); F output It represents the final output after the fusion of the two sub-networks, UpSampling represents the reconstruction of the features using upsampling and convolution, and CAT represents the concatenation of the two parts of the features.

8. The microseismic profile denoising method based on time-frequency domain combination according to claim 7 is characterized in that: Step S2 includes a spatial transformer attention module STAB, and the feature extraction module STAB is composed of a one-dimensional spatial self-attention module, normalization and a multi-layer perceptron, including position encoding embedding, normalization, and spatial self-attention operations on the feature output.

9. The microseismic signal denoising method based on time-frequency domain combination according to claim 8, characterized in that: Step S2 includes an adaptive attention fusion module AAFB. After downsampling, the time series signal is divided into two parts for output. One part undergoes convolution and upsampling operations, and the other part undergoes feature extraction operations using the STAB module. The features of the two parts of the output are fused with weighted attention using the AAFB module.

10. The microseismic profile denoising method based on time-frequency domain combination according to claim 9, characterized in that: Step S3 also includes the loss function of the training process in S3, including coefficient loss function, timing loss function and time-frequency fusion loss function. The coefficient loss function, timing loss function and time-frequency fusion loss function are used to constrain the WNet sub-network, TNet sub-network and Fusion sub-network respectively. The loss function calculation formula is: Loss=λ1Loss w +λ2loss t +λ3loss tw ; In the formula, loss w 、loss t 、loss tw They are coefficient loss function, time series loss function and time-frequency fusion loss function respectively.

Images