Seismic random noise attenuation method and device based on convolutional long short-term memory network
Through the autoencoder model based on the convolutional long and short-term memory network, seismic data is processed in blocks and iteratively denoising, which solves the problem of relying on clean label data in supervised deep learning, and achieves efficient denoising effect and good robustness without relying on labels.
Patent Information
- Application Number
- CN202311655432.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-05
- Publication Date
- 2025-06-06
AI Technical Summary
Supervised deep learning denoising methods rely on clean seismic data as training labels, but it is difficult for explorers to obtain this data, resulting in a lot of effort to create label samples.
The seismic random noise decay method based on convolutional long and short-term memory network is adopted. By constructing a convolutional long and short-term memory network autoencoder model with a bidirectional connection structure, the original seismic data is blocked and iteratively denoised. The autoencoder model is used to mine useful information from the seismic data itself.
The denoising effect without relying on label data is achieved, and the difficulty of making label samples in supervised learning is solved. At the same time, it shows good denoising robustness at different noise levels, and basically does not damage the effective signal.
Smart Images
Figure CN120103435A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of seismic data processing in oil and gas geophysical exploration, and more specifically, to a seismic random noise attenuation method and device based on a convolutional long short-term memory network. Background Art
[0002] As the exploration environment becomes increasingly complex, seismic data collected in the field are often accompanied by noise. The presence of noise reduces the quality of seismic profiles and has an adverse impact on seismic interpretation, seismic imaging, seismic inversion, etc.
[0003] In the past few decades, a large number of seismic noise suppression methods have been proposed by researchers. Denoising methods based on prediction and denoising methods based on transform domain are two classic denoising methods. Denoising methods based on prediction use the characteristics that effective waves can be predicted but random noise cannot be predicted to suppress random noise, such as the fx deconvolution denoising method. Denoising methods based on transform domain use the good difference between effective waves and random noise in the transform domain to remove random noise by designing denoising thresholds in the transform domain, such as wavelet transform and curvelet transform. The above traditional methods have the disadvantages of insufficient denoising strength when removing random noise, resulting in residual noise or excessive denoising strength, resulting in damage to the effective signal.
[0004] In recent years, deep learning, as an emerging technology, has achieved good results in speech recognition, image processing, and autonomous driving. Deep learning can be divided into supervised deep learning and unsupervised deep learning according to whether there are training labels or not. At present, supervised learning is the main method for suppressing random noise in seismic data. The denoising effect of supervised deep learning depends on clean seismic data as training labels, but it is difficult for explorers to obtain clean seismic data, which requires a lot of effort to produce labeled samples when applying supervised learning to denoise. Summary of the invention
[0005] In view of this, the present invention discloses a seismic random noise attenuation method based on a convolutional long short-term memory network, which can solve the problem that the supervised deep learning denoising effect depends on clean seismic data as training labels, while it is difficult for explorers to obtain clean seismic data, which makes it necessary to spend a lot of effort to produce label samples when applying supervised learning denoising.
[0006] According to one aspect of the present invention, a seismic random noise attenuation method based on a convolutional long short-term memory network is proposed, the method comprising:
[0007] Step 1, block processing is performed on the original seismic data to obtain the seismic data after block processing;
[0008] Step 2, construct a convolutional long short-term memory network autoencoder model with a bidirectional connection structure;
[0009] Step 3, inputting the block-processed seismic data into the convolutional long short-term memory network autoencoder model with a bidirectional connection structure to obtain reconstructed seismic data;
[0010] Step 4: Iterative denoising is performed by taking the error between each reconstructed seismic data and the original input seismic data as the convergence cost.
[0011] In some embodiments, the step of constructing a convolutional long short-term memory network autoencoder model with a bidirectional connection structure includes:
[0012] Under the preset encoding framework and decoding framework, a convolutional long short-term memory network autoencoder model with a bidirectional connection structure is constructed through a preset number of convolutional long short-term network bidirectional connection structure layers and pooling layers.
[0013] In some embodiments, the step of inputting the block-processed seismic data into the convolutional long short-term memory network autoencoder model with a bidirectional connection structure to obtain reconstructed seismic data includes:
[0014] The block-processed seismic data is compressed at the encoding layer through a convolutional long-short term network bidirectional connection structure layer and a pooling layer with a preset number of layers;
[0015] In the decoding layer, the feature map is expanded through a preset number of convolutional long-short term network bidirectional connection structure layer and an upsampling layer, and the expanded feature map is input into a three-dimensional convolutional layer to obtain reconstructed seismic data.
[0016] In some embodiments, in the preset number of convolutional long short-term network bidirectional connection structure layers, each layer of bidirectional convolutional long short-term network includes a forward layer and a reverse layer.
[0017] In some embodiments, the step of inputting the block-processed seismic data into the convolutional long short-term memory network autoencoder model with a bidirectional connection structure to obtain reconstructed seismic data includes:
[0018] According to the formula s = σ(W 1 x+b 1 ) and z=σ(W 2 x+b 2 ) obtains reconstructed seismic data, wherein σ represents a nonlinear activation function, x represents the input seismic data after the block processing, Z represents the obtained reconstructed seismic data, and W and b represent a trainable weight matrix and a bias vector, respectively.
[0019] In some embodiments, the step of iteratively denoising by using the error between each reconstructed seismic data and the original input seismic data as a convergence cost comprises:
[0020] By presetting the root mean square error function, the error between each reconstructed seismic data and the original input seismic data is used as the convergence cost for iterative denoising.
[0021] In some implementations, the step of obtaining data after the original seismic data is processed into blocks includes:
[0022] The original seismic data is divided into blocks with a block size of 16×48×48 to obtain seismic data after block processing.
[0023] According to one aspect of the present invention, a seismic random noise attenuation device based on a convolutional long short-term memory network is also proposed, the device comprising:
[0024] An acquisition module is used to process the original seismic data in blocks and obtain the seismic data after the block processing;
[0025] A building block for building a convolutional long short-term memory network autoencoder model with a bidirectional connection structure;
[0026] A reconstruction module, used for inputting the block-processed seismic data into the convolutional long short-term memory network autoencoder model with a bidirectional connection structure to obtain reconstructed seismic data;
[0027] The denoising module is used to iteratively denoise the error between each reconstructed seismic data and the original input seismic data as the convergence cost.
[0028] According to another aspect of the present invention, an electronic device is also provided, the electronic device comprising:
[0029] A memory storing executable instructions;
[0030] A processor runs the executable instructions in the memory to implement the seismic random noise attenuation method based on convolutional long short-term memory network as described above.
[0031] According to another aspect of the present invention, a computer-readable storage medium is also proposed, which stores a computer program. When the computer program is executed by a processor, the seismic random noise attenuation method based on convolutional long short-term memory network described above is implemented.
[0032] The technical solution has at least the following advantages: since the unsupervised autoencoder model based on the convolutional long short-term memory network of the present invention suppresses three-dimensional seismic random noise, useful information can be mined from the seismic data itself, and its denoising effect does not depend on the label data, which solves the defect that supervised deep learning denoising requires the production of a large amount of clean label data; at the same time, the present invention uses a bidirectional connection structure in the network model design, so the present invention can take the noise conditions of adjacent seismic profiles into account when attenuating the random noise of the current seismic profile, thereby ensuring the suppression effect of seismic random noise while basically not damaging the effective signal; and the method provided by the present invention has good denoising robustness, and has good denoising effect under different noise levels.
[0033] The methods and apparatus of the present invention have other features and advantages that will be apparent from or will be described in detail in the accompanying drawings and subsequent detailed descriptions incorporated herein, which together serve to explain the specific principles of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] The above and other objects, features and advantages of the present invention will become more apparent through a more detailed description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, wherein like reference numerals generally represent like components throughout the exemplary embodiments of the present invention.
[0035] Figure 1 A flow chart of a seismic random noise attenuation method based on a convolutional long short-term memory network according to an embodiment of the present invention is shown.
[0036] Figure 2 A schematic diagram of a seismic random noise attenuation device based on a convolutional long short-term memory network according to another embodiment of the present invention is shown;
[0037] Figure 3 A schematic diagram of a seismic random noise attenuation electronic device based on a convolutional long short-term memory network according to another embodiment of the present invention is shown;
[0038] Figure 4 A schematic diagram of a computer-readable storage medium for attenuating seismic random noise based on a convolutional long short-term memory network according to another embodiment of the present invention is shown;
[0039] Figure 5 This is a schematic diagram of the convolutional long-term short-term network model structure;
[0040] Figure 6 This is a schematic diagram of the bidirectional connection structure of the convolutional long-term and short-term network model;
[0041] Figure 7This is a schematic diagram of the convolutional long short-term memory network autoencoder denoising model;
[0042] Figure 8a Schematic diagram of clean synthetic seismic recording;
[0043] Figure 8b This is a schematic diagram of noise-added synthetic seismic records;
[0044] Figure 8c This is a schematic diagram of MSSA denoising results;
[0045] Figure 8d This is a schematic diagram of denoising results according to an embodiment of the present invention;
[0046] Fig. 9 Schematic diagram of the comparison of average amplitude spectra before and after denoising;
[0047] Fig.10 Schematic diagram of the comparison of denoising results under different noise levels;
[0048] Fig.11a This is a schematic diagram of actual earthquake data;
[0049] Fig.11b This is a schematic diagram of MSSA denoising results;
[0050] Fig.11c Schematic diagram of denoising results according to an embodiment of the present invention;
[0051] Fig.12a This is a schematic diagram of actual earthquake data;
[0052] Figure 12b This is a schematic diagram of MSSA denoising results;
[0053] Fig.12c This is the denoising result of the present invention;
[0054] Fig.12d is the MSSA residual profile;
[0055] Fig.12e This is the residual profile of the embodiment of the present invention. DETAILED DESCRIPTION
[0056] The preferred embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the preferred embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.
[0057] The present invention proposes a seismic random noise attenuation method based on a convolutional long short-term memory network, the method comprising:
[0058] Step 1, block processing is performed on the original seismic data to obtain the seismic data after block processing;
[0059] Step 2, construct a convolutional long short-term memory network autoencoder model with a bidirectional connection structure;
[0060] Step 3, inputting the block-processed seismic data into the convolutional long short-term memory network autoencoder model with a bidirectional connection structure to obtain reconstructed seismic data;
[0061] Step 4: Iterative denoising is performed by taking the error between each reconstructed seismic data and the original input seismic data as the convergence cost.
[0062] In some embodiments, the step of constructing a convolutional long short-term memory network autoencoder model with a bidirectional connection structure includes:
[0063] Under the preset encoding framework and decoding framework, a convolutional long short-term memory network autoencoder model with a bidirectional connection structure is constructed through a preset number of convolutional long short-term network bidirectional connection structure layers and pooling layers.
[0064] In some embodiments, the step of inputting the block-processed seismic data into the convolutional long short-term memory network autoencoder model with a bidirectional connection structure to obtain reconstructed seismic data includes:
[0065] The block-processed seismic data is compressed at the encoding layer through a convolutional long-short term network bidirectional connection structure layer and a pooling layer with a preset number of layers;
[0066] In the decoding layer, the feature map is expanded through a preset number of convolutional long-short term network bidirectional connection structure layer and an upsampling layer, and the expanded feature map is input into a three-dimensional convolutional layer to obtain reconstructed seismic data.
[0067] In some embodiments, in the preset number of convolutional long short-term network bidirectional connection structure layers, each layer of bidirectional convolutional long short-term network includes a forward layer and a reverse layer.
[0068] In some embodiments, the step of inputting the block-processed seismic data into the convolutional long short-term memory network autoencoder model with a bidirectional connection structure to obtain reconstructed seismic data includes:
[0069] According to the formula s = σ(W 1 x+b 1 ) and z=σ(W 2 x+b 2) obtains reconstructed seismic data, wherein σ represents a nonlinear activation function, x represents the input seismic data after the block processing, Z represents the obtained reconstructed seismic data, and W and b represent a trainable weight matrix and a bias vector, respectively.
[0070] In some embodiments, the step of iteratively denoising by using the error between each reconstructed seismic data and the original input seismic data as a convergence cost comprises:
[0071] By presetting the root mean square error function, the error between each reconstructed seismic data and the original input seismic data is used as the convergence cost for iterative denoising.
[0072] In some implementations, the step of obtaining data after the original seismic data is processed into blocks includes:
[0073] The original seismic data is divided into blocks with a block size of 16×48×48 to obtain seismic data after block processing.
[0074] The present invention also proposes a seismic random noise attenuation device based on a convolutional long short-term memory network, the device comprising:
[0075] An acquisition module is used to process the original seismic data in blocks and obtain the seismic data after the block processing;
[0076] A building block for building a convolutional long short-term memory network autoencoder model with a bidirectional connection structure;
[0077] A reconstruction module, used for inputting the block-processed seismic data into the convolutional long short-term memory network autoencoder model with a bidirectional connection structure to obtain reconstructed seismic data;
[0078] The denoising module is used to iteratively denoise the error between each reconstructed seismic data and the original input seismic data as the convergence cost.
[0079] The present invention further provides an electronic device, comprising:
[0080] A memory storing executable instructions;
[0081] A processor runs the executable instructions in the memory to implement the seismic random noise attenuation method based on convolutional long short-term memory network as described above.
[0082] The present invention also proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the seismic random noise attenuation method based on convolutional long short-term memory network mentioned above.
[0083] Example 1
[0084] Figure 1 The flowchart of the seismic random noise attenuation method based on the convolutional long short-term memory network according to an embodiment of the present invention is shown. As shown in the figure, the method includes steps 1 to 4.
[0085] Step 1: block-process the original seismic data to obtain the seismic data after block processing.
[0086] Specifically, the original seismic data is processed by block processing with a block size of 16×48×48 to obtain seismic data after block processing.
[0087] Step 2: Build a convolutional long short-term memory network autoencoder model with a bidirectional connection structure.
[0088] Specifically, under the preset encoding framework and decoding framework, a convolutional long short-term memory network autoencoder model with a bidirectional connection structure is constructed through a preset number of convolutional long short-term network bidirectional connection structure layers and pooling layers.
[0089] In order to suppress the random noise of three-dimensional seismic data, the present invention proposes a convolutional long short-term memory network autoencoder denoising model. The network structure is as follows: Figure 7 As shown in the figure, it consists of two parts, the encoding framework and the decoding framework. In the encoding stage, three Biconvlstm layers and pooling are used to compress the seismic data. The decoding layer first uses three Biconvlstm layers and upsampling layers to expand the feature map, and then feeds the expanded feature map to the three-dimensional convolution layer to obtain the final reconstructed data. The Biconvlstm layer uses 12 3×3 convolution kernels. From this network, the input data size is 16×48×48.
[0090] It should be noted that the autoencoder is an unsupervised learning algorithm that uses the back-propagation algorithm to reconstruct the input as much as possible. The autoencoder consists of two stages: encoding and decoding. The encoding stage can convert the original data into a compressed expression. The whole process can be expressed as:
[0091] s=σ(W 1 x+b 1 ) (1)
[0092] Where σ represents a nonlinear activation function and x represents input data. The decoding stage can map the compressed expression to the original data. The whole process can be expressed as:
[0093] z=σ(W 2 x+b 2 ) (2)
[0094] Z represents the reconstructed output, W and b represent the trainable weight matrix and bias vector respectively. The long short-term memory network has been proven to capture long-term dependencies, but because its input is two-dimensional, it is difficult to process three-dimensional data. In order to solve this problem, the convolutional long short-term memory network was proposed. The main structure of the long short-term memory network and the convolutional long short-term memory network is the same. The difference is that the former uses full connection and the latter introduces convolution operation during information transmission. Therefore, the convolutional long short-term memory network can extract the spatiotemporal characteristics of the data very well. The internal structure of the convolutional long short-term memory network is shown in the figure. Its hidden unit state h t Contains cell state c t , forget gate f t , input gate i t , output gate o t . Cell state c t It carries information about the past moment, and its information transmission is controlled by three gates. When new input arrives, the forget gate f t Determines whether old information is removed from the cell state c t Clear, the input gate determines whether new information is added to the cell state c t . Cell state c t After the information is updated, it is passed to the hidden state h through the output gate t The entire internal model information transfer process can be expressed as follows:
[0095] f t =σ(W f *[h t-1 ,x t ,c t-1 ]+b f ) (3)
[0096] i t =σ(W i *[h t-1 ,x t ,c t-1 ]+b i ) (4)
[0097]
[0098] o t =σ(W o *[h t-1 ,x t ,c t-1 ]+b o ) (6)
[0099]
[0100] Furthermore, replace formulas (3)-(7) with the following formula:
[0101] (h t ,c t )=H(x t ,h t-1 ,c t-1 ) (8)
[0102] In order to better capture the spatiotemporal characteristics of three-dimensional seismic data, the present invention adopts a bidirectional connection structure. Figure 5 and Figure 6 As shown, each bidirectional Convlstm layer contains two unidirectional Convlstm layers: a forward layer and a reverse layer. The output values of these two unidirectional layers are aggregated together to produce the output value y t , so the output value y at a certain moment t Depending on the previous state, the state at the future moment can also be used. The entire information transmission process of the bidirectional connection structure layer is as follows:
[0103]
[0104]
[0105]
[0106] Step 3: input the block-processed seismic data into the convolutional long short-term memory network autoencoder model with a bidirectional connection structure to obtain reconstructed seismic data.
[0107] Specifically, the block-processed seismic data is compressed in the encoding layer through a preset number of convolutional long-term and short-term network bidirectional connection structure layers and a pooling layer; in the decoding layer, the feature map is enlarged through a preset number of convolutional long-term and short-term network bidirectional connection structure layers and an upsampling layer, and the enlarged feature map is input into the three-dimensional convolution layer to obtain the reconstructed seismic data. Further, according to the formula s=σ(W 1 x+b 1 ) and z=σ(W 2 x+b 2 ) obtains reconstructed seismic data, wherein σ represents a nonlinear activation function, x represents the input seismic data after the block processing, Z represents the obtained reconstructed seismic data, and W and b represent a trainable weight matrix and a bias vector, respectively.
[0108] Step 4: Iterative denoising is performed by taking the error between each reconstructed seismic data and the original input seismic data as the convergence cost.
[0109] Specifically, by presetting a root mean square error function, the error between each reconstructed seismic data and the original input seismic data is used as a convergence cost for iterative denoising.
[0110] It should be noted that Figure 8a The synthetic clean seismic data is shown, which contains complex geological structures: linear phase axis, curved phase axis, and fault. Random noise is added to the clean synthetic seismic record to make it noisy seismic data with SNR = 0.52dB, and then it is denoised by multi-channel singular value denoising method MSSA and the present invention respectively. The denoising results are shown in FIG. Figure 8b to Figure 8d As shown. It can be seen that the denoising result of MSSA still contains a lot of residual noise, while the denoising result of the present invention does not show obvious residual noise. In order to quantitatively compare the two denoising methods, the signal-to-noise ratio of the denoising results is calculated. The signal-to-noise ratio SNR of the MSSA denoising result is 9.88dB, and the signal-to-noise ratio SNR of the denoising result of the present invention is 13.53dB. It can be seen that the performance of the denoising result of the present invention is greatly improved. Fig. 9 The comparison of the average amplitude spectra of the noise-free data, noisy data and denoised data implemented by the present invention is shown. It can be seen that the average amplitude spectra before and after denoising are basically the same, which illustrates the effectiveness of the denoising ability of the present invention.
[0111] Furthermore, the denoising robustness of the present invention is tested by a series of synthetic examples with different noise levels. Fig.10 As shown. It can be seen that the higher the signal-to-noise ratio of the original data, the better the denoising result. Secondly, under different noise levels, the denoising performance of the present invention is always better than that of MSSA.
[0112] The method provided in Example 1 of the present invention can mine useful information from the seismic data itself because the unsupervised autoencoder model based on the convolutional long short-term memory network suppresses the three-dimensional seismic random noise. Its denoising effect does not depend on the label data, which solves the defect that supervised deep learning denoising requires the production of a large amount of clean label data. At the same time, the present invention uses a bidirectional connection structure in the network model design. Therefore, when attenuating the random noise of the current seismic profile, the present invention can take the noise conditions of the adjacent seismic profiles into account, thereby ensuring the suppression effect of the seismic random noise while basically not damaging the effective signal. In addition, the method provided by the present invention has good denoising robustness and has a good denoising effect at different noise levels.
[0113] Example 2
[0114] According to one embodiment of the present invention, a seismic random noise attenuation device based on a convolutional long short-term memory network is provided. Figure 2 As shown, the device includes: an acquisition module 21, a construction module 22, a reconstruction module 23, and a denoising module 24;
[0115] An acquisition module 21 is used to process the original seismic data in blocks and obtain the seismic data after the block processing;
[0116] A construction module 22 is used to construct a convolutional long short-term memory network autoencoder model with a bidirectional connection structure;
[0117] A reconstruction module 23, used for inputting the block-processed seismic data into the convolutional long short-term memory network autoencoder model with a bidirectional connection structure to obtain reconstructed seismic data;
[0118] The denoising module 24 is used for iterative denoising by taking the error between each reconstructed seismic data and the original input seismic data as the convergence cost.
[0119] The device provided in Example 2 of the present invention can suppress three-dimensional seismic random noise by using an unsupervised autoencoder model based on a convolutional long short-term memory network, and can mine useful information from the seismic data itself. Its denoising effect does not depend on label data, which solves the defect that supervised deep learning denoising requires the production of a large amount of clean label data. At the same time, the present invention uses a bidirectional connection structure in the network model design, so the present invention can take the noise conditions of adjacent seismic profiles into account when attenuating the random noise of the current seismic profile, thereby ensuring the suppression effect of seismic random noise while basically not damaging the effective signal. In addition, the method provided by the present invention has good denoising robustness and has good denoising effect at different noise levels.
[0120] Example 3
[0121] According to another aspect of the present invention, an electronic device is also provided. Figure 3 As shown, the electronic device includes:
[0122] The memory 31 stores executable instructions:
[0123] A processor 32 is used to execute the executable instructions in the memory to implement the seismic random noise attenuation method based on convolutional long short-term memory network according to the present invention.
[0124] The method comprises the following steps:
[0125] Step 1, block processing is performed on the original seismic data to obtain the seismic data after block processing;
[0126] Step 2, construct a convolutional long short-term memory network autoencoder model with a bidirectional connection structure;
[0127] Step 3, inputting the block-processed seismic data into the convolutional long short-term memory network autoencoder model with a bidirectional connection structure to obtain reconstructed seismic data;
[0128] Step 4: Iterative denoising is performed by taking the error between each reconstructed seismic data and the original input seismic data as the convergence cost.
[0129] Furthermore, the step of constructing a convolutional long short-term memory network autoencoder model with a bidirectional connection structure includes:
[0130] Under the preset encoding framework and decoding framework, a convolutional long short-term memory network autoencoder model with a bidirectional connection structure is constructed through a preset number of convolutional long short-term network bidirectional connection structure layers and pooling layers.
[0131] Furthermore, the step of inputting the block-processed seismic data into the convolutional long short-term memory network autoencoder model with a bidirectional connection structure to obtain reconstructed seismic data includes:
[0132] The block-processed seismic data is compressed at the encoding layer through a convolutional long-short term network bidirectional connection structure layer and a pooling layer with a preset number of layers;
[0133] In the decoding layer, the feature map is expanded through a preset number of convolutional long-short term network bidirectional connection structure layer and an upsampling layer, and the expanded feature map is input into a three-dimensional convolutional layer to obtain reconstructed seismic data.
[0134] Furthermore, in the preset number of convolutional long-short-term network bidirectional connection structure layers, each layer of bidirectional convolutional long-short-term network includes a forward layer and a reverse layer.
[0135] Furthermore, the step of inputting the block-processed seismic data into the convolutional long short-term memory network autoencoder model with a bidirectional connection structure to obtain reconstructed seismic data includes:
[0136] According to the formula s = σ(W 1 x+b 1 ) and z=σ(W 2 x+b 2 ) obtains reconstructed seismic data, wherein σ represents a nonlinear activation function, x represents the input seismic data after the block processing, Z represents the obtained reconstructed seismic data, and W and b represent a trainable weight matrix and a bias vector, respectively.
[0137] Furthermore, the step of performing iterative denoising by taking the error between each reconstructed seismic data and the original input seismic data as the convergence cost comprises:
[0138] By presetting the root mean square error function, the error between each reconstructed seismic data and the original input seismic data is used as the convergence cost for iterative denoising.
[0139] Furthermore, the step of obtaining data after the original seismic data is processed into blocks includes:
[0140] The original seismic data is divided into blocks with a block size of 16×48×48 to obtain seismic data after block processing.
[0141] The electronic device provided in Example 3 of the present invention can mine useful information from the seismic data itself because the unsupervised autoencoder model based on the convolutional long short-term memory network suppresses three-dimensional seismic random noise. Its denoising effect does not depend on the label data, which solves the defect that supervised deep learning denoising requires the production of a large amount of clean label data. At the same time, the present invention uses a bidirectional connection structure in the network model design. Therefore, when attenuating the random noise of the current seismic profile, the present invention can take the noise conditions of the adjacent seismic profiles into account, thereby ensuring the suppression effect of the seismic random noise while basically not damaging the effective signal. In addition, the method provided by the present invention has good denoising robustness and has a good denoising effect at different noise levels.
[0142] Example 4
[0143] According to another aspect of the present invention, a computer readable storage medium is also provided. Figure 4 As shown, the computer-readable storage medium 41 stores a computer program, and when the computer program is executed by a processor, the seismic random noise attenuation method based on a convolutional long short-term memory network according to the present invention is implemented.
[0144] The method comprises the following steps:
[0145] Step 1, block processing is performed on the original seismic data to obtain the seismic data after block processing;
[0146] Step 2, construct a convolutional long short-term memory network autoencoder model with a bidirectional connection structure;
[0147] Step 3, inputting the block-processed seismic data into the convolutional long short-term memory network autoencoder model with a bidirectional connection structure to obtain reconstructed seismic data;
[0148] Step 4: Iterative denoising is performed by taking the error between each reconstructed seismic data and the original input seismic data as the convergence cost.
[0149] Furthermore, the step of constructing a convolutional long short-term memory network autoencoder model with a bidirectional connection structure includes:
[0150] Under the preset encoding framework and decoding framework, a convolutional long short-term memory network autoencoder model with a bidirectional connection structure is constructed through a preset number of convolutional long short-term network bidirectional connection structure layers and pooling layers.
[0151] Furthermore, the step of inputting the block-processed seismic data into the convolutional long short-term memory network autoencoder model with a bidirectional connection structure to obtain reconstructed seismic data includes:
[0152] The block-processed seismic data is compressed at the encoding layer through a convolutional long-short term network bidirectional connection structure layer and a pooling layer with a preset number of layers;
[0153] In the decoding layer, the feature map is expanded through a preset number of convolutional long-short term network bidirectional connection structure layer and an upsampling layer, and the expanded feature map is input into a three-dimensional convolutional layer to obtain reconstructed seismic data.
[0154] Furthermore, in the preset number of convolutional long-short-term network bidirectional connection structure layers, each layer of bidirectional convolutional long-short-term network includes a forward layer and a reverse layer.
[0155] Furthermore, the step of inputting the block-processed seismic data into the convolutional long short-term memory network autoencoder model with a bidirectional connection structure to obtain reconstructed seismic data includes:
[0156] According to the formula s = σ(W 1 x+b 1 ) and z=σ(W 2 x+b 2 ) obtains reconstructed seismic data, wherein σ represents a nonlinear activation function, x represents the input seismic data after the block processing, Z represents the obtained reconstructed seismic data, and W and b represent a trainable weight matrix and a bias vector, respectively.
[0157] Furthermore, the step of performing iterative denoising by taking the error between each reconstructed seismic data and the original input seismic data as the convergence cost comprises:
[0158] By presetting the root mean square error function, the error between each reconstructed seismic data and the original input seismic data is used as the convergence cost for iterative denoising.
[0159] Furthermore, the step of obtaining data after the original seismic data is processed into blocks includes:
[0160] The original seismic data is divided into blocks with a block size of 16×48×48 to obtain seismic data after block processing.
[0161] The computer-readable storage medium provided in Example 4 of the present invention can mine useful information from the seismic data itself because the unsupervised autoencoder model based on the convolutional long short-term memory network suppresses three-dimensional seismic random noise. Its denoising effect does not depend on the label data, which solves the defect that supervised deep learning denoising requires the production of a large amount of clean label data. At the same time, the present invention uses a bidirectional connection structure in the network model design. Therefore, when attenuating the random noise of the current seismic profile, the present invention can take the noise conditions of the adjacent seismic profiles into account, thereby ensuring the suppression effect of the seismic random noise while basically not damaging the effective signal. In addition, the method provided by the present invention has good denoising robustness and has a good denoising effect at different noise levels.
[0162] Example 5
[0163] In order to verify the effect of the seismic random noise attenuation scheme based on convolutional long short-term memory network according to the present invention, the test is conducted on actual seismic data seriously polluted by random noise, such as Figure 11a-Figure 11c The denoising results of the MSSA method and the present invention are as follows: Fig.11b and Fig.11c It can be seen that the continuity of the MSSA event axis is not as good as that of the present invention, which is caused by the MSSA method damaging some effective signals. A random section is selected to further compare the denoising effects of the two. Fig.12a , Figure 12b and Fig.12c The denoising results of the MSSA method and the present invention are shown in Figure 1. It can be seen that the denoising results of the present invention have richer details. Fig.12d is the residual profile between the denoising result of the MSSA method and the actual seismic data, Fig.12e This is the residual section of the denoising result of the present invention and the actual seismic data. It can be seen that the MSSA residual section has obvious effective signal residues, while no obvious effective signal residues can be seen in the residual section of the present invention.
[0164] This example fully illustrates that the present invention can mine useful information from the seismic data itself, and its denoising effect does not depend on the labeled data, which solves the defect that supervised deep learning denoising requires the production of a large amount of clean labeled data; at the same time, the present invention uses a bidirectional connection structure in the network model design, so the present invention can take the noise conditions of adjacent seismic profiles into account when attenuating the random noise of the current seismic profile, thereby ensuring the suppression effect of seismic random noise while basically not damaging the effective signal; and the method provided by the present invention has good denoising robustness, and has good denoising effect under different noise levels.
[0165] For other detailed descriptions of this exemplary embodiment, reference may be made to the corresponding descriptions in the aforementioned embodiments, which will not be repeated here.
[0166] The embodiments of the present invention have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The terms used herein are selected to best explain the principles of the embodiments, practical applications, or technical improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.
Claims
1. A seismic random noise attenuation method based on convolutional long short-term memory network, It is characterized in that The method comprises: Step 1, block processing is performed on the original seismic data to obtain the seismic data after block processing; Step 2, construct a convolutional long short-term memory network autoencoder model with a bidirectional connection structure; Step 3, inputting the block-processed seismic data into the convolutional long short-term memory network autoencoder model with a bidirectional connection structure to obtain reconstructed seismic data; Step 4: Iterative denoising is performed by taking the error between each reconstructed seismic data and the original input seismic data as the convergence cost.
2. A seismic random noise attenuation method based on convolutional long short-term memory network according to claim 1, It is characterized in that The steps of constructing a convolutional long short-term memory network autoencoder model with a bidirectional connection structure include: Under the preset encoding framework and decoding framework, a convolutional long short-term memory network autoencoder model with a bidirectional connection structure is constructed through a preset number of convolutional long short-term network bidirectional connection structure layers and pooling layers.
3. A seismic random noise attenuation method based on convolutional long short-term memory network according to claim 2, It is characterized in that The step of inputting the block-processed seismic data into the convolutional long short-term memory network autoencoder model with a bidirectional connection structure to obtain reconstructed seismic data includes: The block-processed seismic data is compressed at the encoding layer through a convolutional long-short term network bidirectional connection structure layer and a pooling layer with a preset number of layers; In the decoding layer, the feature map is expanded through a preset number of convolutional long-short term network bidirectional connection structure layer and an upsampling layer, and the expanded feature map is input into a three-dimensional convolutional layer to obtain reconstructed seismic data.
4. A seismic random noise attenuation method based on convolutional long short-term memory network according to any one of claims 1 to 3, It is characterized in that In the preset number of convolutional long-short-term network bidirectional connection structure layers, each layer of bidirectional convolutional long-short-term network includes a forward layer and a reverse layer.
5. According to claim 3, a seismic random noise attenuation method based on convolutional long short-term memory network, It is characterized in that The step of inputting the block-processed seismic data into the convolutional long short-term memory network autoencoder model with a bidirectional connection structure to obtain reconstructed seismic data includes: According to the formula s = σ(W 1 x+b 1 ) and z=σ(W 2 x+b 2 ) obtains reconstructed seismic data, wherein σ represents a nonlinear activation function, x represents the input seismic data after the block processing, Z represents the obtained reconstructed seismic data, and W and b represent a trainable weight matrix and a bias vector, respectively.
6. A seismic random noise attenuation method based on convolutional long short-term memory network according to claim 1, It is characterized in that The step of iteratively denoising by taking the error between each reconstructed seismic data and the original input seismic data as the convergence cost comprises: By presetting the root mean square error function, the error between each reconstructed seismic data and the original input seismic data is used as the convergence cost for iterative denoising.
7. The seismic random noise attenuation method based on convolutional long short-term memory network according to claim 1, It is characterized in that The step of obtaining data after block processing of the original seismic data comprises: The original seismic data is divided into blocks with a block size of 16×48×48 to obtain seismic data after block processing.
8. A seismic random noise attenuation device based on convolutional long short-term memory network, It is characterized in that The device comprises: An acquisition module is used to process the original seismic data in blocks and obtain the seismic data after the block processing; A building block for building a convolutional long short-term memory network autoencoder model with a bidirectional connection structure; A reconstruction module, used for inputting the block-processed seismic data into the convolutional long short-term memory network autoencoder model with a bidirectional connection structure to obtain reconstructed seismic data; The denoising module is used to iteratively denoise the error between each reconstructed seismic data and the original input seismic data as the convergence cost.
9. An electronic device, It is characterized in that The electronic device comprises: A memory storing executable instructions; A processor, wherein the processor runs the executable instructions in the memory to implement the method according to any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, wherein the computer program implements the method according to any one of claims 1 to 7 when executed by a processor.