Dual-channel denoising network model training method and seismic data denoising method and device

Abstract

The invention discloses a two-channel denoising network model training method and a seismic data denoising method and device. The training method comprises the following steps: constructing a sample data set based on obtained noisy seismic data and corresponding label data; the sample data set is input into a constructed dual-channel denoising network model for model training, the network model comprises a frequency domain channel network, a time domain channel network and a shared network, the frequency domain channel network and the time domain channel network carry out learning training on the sample data in a frequency domain and a time domain respectively, and the frequency domain channel network and the time domain channel network carry out data processing; the first de-noised data and the second de-noised data are respectively output as input data of a shared network, and the shared network outputs third de-noised data; and calculating a preset loss function based on the third de-noised data output by the shared network and the sample data set, and obtaining a trained network model after a preset network model training termination condition is satisfied. The network model can process seismic data containing noise with high efficiency and high precision.

Description

Technical Field

[0001] The present invention relates to the technical field of seismic data processing, and in particular to a dual-channel denoising network model training method, a seismic data denoising method and a device. Background Art

[0002] In recent years, as the exploration environment becomes increasingly complex and the demand for fine exploration continues to increase, the demand for efficient and fidelity denoising of seismic data has become increasingly strong. The types of noise in seismic data are complex and diverse. For example, according to the classification of noise morphology, seismic noise can be roughly divided into random noise and coherent noise. For random noise, existing noise suppression algorithms include fx domain prediction filtering algorithm, wavelet transform algorithm, curvelet transform algorithm and other methods; for coherent noise such as surface waves and linear interference, existing algorithms include FK filtering (frequency-wavenumber filtering) algorithm, FKK domain filtering (frequency-wavenumber-wavenumber filtering) algorithm, abnormal amplitude attenuation algorithm, KL (Karhunen-Loeve) transform algorithm, dip filter algorithm and other algorithms. Summary of the invention

[0003] Existing noise suppression algorithms abstract information such as noise propagation laws, physical properties, and morphological characteristics into mathematical language, and complete the separation of signal and noise in a certain data space (such as frequency domain, curve domain, etc.); and in the process of mathematical abstraction, theoretical assumptions are usually required. Although these assumptions can characterize certain characteristics of seismic data to a certain extent, they usually require mathematical approximation, which in turn causes limitations of denoising algorithms; in addition, conventional denoising algorithms are usually designed for a certain specific noise. In actual production, multiple denoising algorithms are required to jointly denoise seismic data for various types of noise in seismic data, and in the process of seismic data denoising, each denoising algorithm requires manual parameter debugging, resulting in a strong subjectivity in the noise suppression process and low efficiency when processing massive seismic data. The inventors of this application have found that the use of deep network learning technology can well solve the problems of low efficiency and poor accuracy in existing noise suppression algorithms.

[0004] In view of the above problems, the present invention is proposed to provide a dual-channel denoising network model training method, seismic data denoising method and device that overcome the above problems or at least partially solve the above problems.

[0005] In a first aspect, an embodiment of the present invention provides a dual-channel denoising network model training method, comprising:

[0006] Constructing a sample data set based on the acquired noisy seismic data and its corresponding label data, wherein the label data is the seismic data after denoising;

[0007] Inputting the sample data of the sample data set into the constructed dual-channel denoising network model for model training, the dual-channel denoising network model includes a frequency domain channel network, a time domain channel network and a shared network; wherein the frequency domain channel network converts the sample data into frequency domain data, performs learning and training on the frequency domain data, and outputs first denoised data; the time domain channel network performs learning and training on the sample data, and outputs second denoised data; the first denoised data and the second denoised data are used as input data of the shared network, and the shared network outputs third denoised data;

[0008] Based on the third denoising data output by the shared network and the sample data set, a preset loss function is calculated. If the loss function meets the preset convergence condition or the number of training iterations reaches the preset threshold, a trained dual-channel denoising network model is obtained; otherwise, the network parameters of the dual-channel denoising network model are adjusted and then the steps of continuing model training are returned.

[0009] In an optional embodiment, obtaining seismic data and label data corresponding to the seismic data includes:

[0010] Using actual seismic data as acquired seismic data containing noise, and using data processed by the selected noise suppression technology as label data;

[0011] and / or,

[0012] The geological structure and geological body in the field are made into a physical model according to a preset simulation similarity ratio, and the physical model is forward simulated to obtain physical model data as the label data;

[0013] The data after adding noise to the label data is used as noisy seismic data.

[0014] In an optional embodiment, the frequency domain channel network includes: a Fourier transform layer, a deep neural network layer and an inverse Fourier transform layer; the frequency domain channel network converts the sample data into frequency domain data, performs learning and training on the frequency domain data, and outputs first denoised data; including:

[0015] The Fourier transform layer performs Fourier transform on the noisy seismic data and the label data to obtain frequency domain data;

[0016] Using the real part and the imaginary part of the frequency domain data as input data of the deep neural network layer, training the deep neural network layer and obtaining frequency domain denoised data;

[0017] The frequency domain denoised data is used as input data of an inverse Fourier transform layer, and the inverse Fourier transform layer performs an inverse Fourier transform on the frequency domain denoised data and outputs the result to the first denoised data.

[0018] In an optional embodiment, the loss function is calculated as follows:

[0019]

[0020] In the formula, K is the number of samples in the sample data set, x k is the earthquake data, y k is the label data, Θ represents the network parameters that the network needs to train and learn; represents the third denoised data output by the shared network; SIM() represents the structural similarity loss; ||·|| F represents the Frobenius norm; |·| represents the 1 norm; μ 1 、u 2 Both are penalty coefficients.

[0021] In an optional embodiment, the network parameters of the dual-channel denoising network model are adjusted through a back-propagation algorithm.

[0022] In a second aspect, an embodiment of the present invention provides a dual-channel seismic data denoising method, comprising:

[0023] Acquiring seismic data to be processed, and performing trace equalization processing on the seismic data to be processed to obtain seismic data after trace equalization processing;

[0024] Input the seismic data after trace equalization into the pre-trained dual-channel denoising network model, and output the seismic data after noise suppression through the dual-channel denoising network model;

[0025] Performing inverse channel equalization processing on the noise-suppressed seismic data to obtain a denoising result of the seismic data;

[0026] The dual-channel denoising network model is trained using the dual-channel denoising network model training method described above.

[0027] In an optional embodiment, performing trace equalization processing on the seismic data to be processed to obtain seismic data after trace equalization processing includes:

[0028] Selecting the maximum value of the seismic data of each seismic channel in the seismic data to be processed as the channel equalization factor in the seismic channel;

[0029] The seismic data in each seismic channel is compared with its corresponding channel equalization factor to obtain the seismic data after channel equalization processing.

[0030] In an optional embodiment, the seismic data after the trace equalization processing is input into a pre-trained dual-channel denoising network model, and the seismic data after the noise is suppressed is output through the dual-channel denoising network model, including:

[0031] Using the Fourier transform layer of the frequency domain channel network, the input seismic data after the trace equalization processing is Fourier transformed to obtain frequency domain data, using the deep neural network layer of the frequency domain channel network to calculate the frequency domain data to obtain frequency domain denoised data, and using the inverse Fourier transform layer to perform Fourier inverse transform on the frequency domain denoised data to obtain first denoised data;

[0032] The seismic data after the trace equalization processing is calculated by using the time domain channel network, and the second denoised data is output by the time domain channel network;

[0033] The first denoised data and the second denoised data are input as input data into the shared network for calculation, and the shared network outputs the noise-suppressed seismic data.

[0034] In an optional embodiment, performing inverse equalization processing on the noise-suppressed seismic data to obtain a denoising result of the seismic data includes:

[0035] The noise-suppressed seismic data is multiplied by the trace equalization factor corresponding to the seismic data to obtain the denoising result of the seismic data.

[0036] In an optional embodiment, the above seismic data denoising method further includes:

[0037] A difference operation is performed on the seismic data to be processed and the denoising result of the seismic data to obtain the noise data of the seismic data to be processed.

[0038] In a third aspect, an embodiment of the present invention further provides a dual-channel denoising network model training device, comprising:

[0039] The first acquisition module: constructs a sample data set based on the acquired noisy seismic data and its corresponding label data, wherein the label data is the seismic data after denoising;

[0040] Training module: inputting the sample data of the sample data set into the constructed dual-channel denoising network model for model training, wherein the dual-channel denoising network model includes a frequency domain channel network, a time domain channel network and a shared network; wherein the frequency domain channel network converts the sample data into frequency domain data, performs learning and training on the frequency domain data, and outputs first denoised data; the time domain channel network performs learning and training on the sample data, and outputs second denoised data; the first denoised data and the second denoised data are used as input data of the shared network, and the shared network outputs third denoised data;

[0041] Model adjustment module: Based on the third denoising data output by the shared network and the sample data set, a preset loss function is calculated. If the loss function meets the preset convergence condition or the number of training iterations reaches the preset threshold, a trained dual-channel denoising network model is obtained; otherwise, the network parameters of the dual-channel denoising network model are adjusted and then the model training step is continued.

[0042] In a fourth aspect, an embodiment of the present invention further provides a dual-channel seismic data denoising device, comprising:

[0043] The second acquisition module is used to acquire the seismic data to be processed, and to perform trace equalization processing on the seismic data to obtain seismic data after trace equalization processing;

[0044] Processing module: input the seismic data after channel equalization into the pre-trained dual-channel denoising network model, and output the seismic data after noise suppression through the dual-channel denoising network model;

[0045] The result acquisition module performs inverse equalization processing on the noise-suppressed seismic data to obtain a denoising result of the seismic data.

[0046] In a fifth aspect, an embodiment of the present invention further provides a computer storage medium, in which computer executable instructions are stored. When the computer executable instructions are executed by a processor, the above-mentioned dual-channel denoising network model training method and the above-mentioned dual-channel seismic data denoising method are implemented.

[0047] In a sixth aspect, an embodiment of the present invention further provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the dual-channel denoising network model training method and the dual-channel seismic data denoising method are implemented.

[0048] The beneficial effects of the above technical solution provided by the embodiment of the present invention include at least:

[0049] The dual-channel denoising network model training method provided by the embodiment of the present invention constructs a sample data set based on the acquired noisy seismic data and its corresponding label data, uses the sample data of the sample data set as the training data of the dual-channel denoising network model, and trains the constructed dual-channel denoising network model. After the loss function meets the preset convergence condition or the number of training iterations reaches the preset number threshold, a trained dual-channel denoising network model is obtained; when training the dual-channel denoising network model, the frequency domain channel network and the time domain channel network included in the dual-channel denoising network model are trained using the sample data, wherein the frequency domain channel network uses the sample data as the training data of the dual-channel denoising network model. The data is converted into frequency domain data, and the frequency domain data is trained. The time domain channel network is used to learn and train the sample data. The time domain denoising results output by the frequency domain channel network and the time domain channel network are used as the input data of the shared network. The denoising results are output after fusion by the shared network. That is, the dual-channel denoising network model is trained from the two channels of the frequency domain and time domain of the sample data. Compared with the method of training the model on the sample data separately from the frequency domain or the time domain, the training result of the dual-channel denoising network model is more accurate. When the network model is used to denoise the seismic data, the denoising result obtained is also more accurate. When denoising noisy seismic data using the trained dual-channel denoising network model, the denoised seismic data can be directly output for the input noisy seismic data. Compared with the existing denoising methods for seismic data, there is no need to design denoising algorithms corresponding to specific noises for different types of noises contained in seismic data. Therefore, when denoising seismic data, there is no need to perform separate parameter debugging on the denoising algorithm corresponding to the specific noise, which improves the noise suppression efficiency of seismic data. In addition, there is no need to make theoretical assumptions when performing denoising, which effectively improves the accuracy of noise identification.

[0050] The dual-channel seismic data denoising method of the present invention, when processing seismic data, first performs channel equalization processing on the seismic data of each seismic channel to make the seismic data more balanced and concentrated, which can avoid the problem that some data cannot be processed due to large differences in data between different seismic channels, and can further improve the accuracy of seismic data processing.

[0051] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings.

[0052] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0054] Figure 1 Flow chart of a dual-channel denoising network model training method according to an embodiment of the present invention;

[0055] Figure 2 A schematic diagram of the structure of a deep neural network in an embodiment of the present invention;

[0056] Figure 3 Flow chart of a dual-channel seismic data denoising method according to an embodiment of the present invention;

[0057] Figure 4 is the noisy seismic data simulated in the embodiment of the present invention;

[0058] Figure 5 For the embodiment of the present invention Figure 4 De-noised seismic data simulated in ;

[0059] Figure 6 The dual-channel seismic data denoising method according to the embodiment of the present invention is Figure 4 The denoising result obtained after data processing;

[0060] Figure 7 For the embodiment of the present invention Figure 4 The real noise of the noisy seismic data simulated in ;

[0061] Figure 8 The dual-channel seismic data denoising method of the present invention is identified in the embodiment of the present invention. Figure 4 Noise;

[0062] Fig. 9 It is a single shot seismic data of a certain exploration block in an embodiment of the present invention;

[0063] Fig.10 In the embodiment of the present invention, conventional composite process is used to Fig. 9 The result after denoising the seismic data;

[0064] Fig.11 The dual-channel seismic data denoising method according to the embodiment of the present invention is Fig. 9 The result after denoising the seismic data;

[0065] Fig.12 In the embodiment of the present invention, the conventional composite process is used to identify Fig. 9 Noise in seismic data;

[0066] Fig.13The dual-channel seismic data denoising method according to the embodiment of the present invention identifies Fig. 9 Noise in seismic data;

[0067] Fig.14 Schematic diagram of a dual-channel denoising network model training device in an embodiment of the present invention;

[0068] Fig.15 Schematic diagram of the process of the dual-channel seismic data denoising method in an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0070] As a machine learning algorithm that has developed rapidly in recent years, deep network learning has achieved revolutionary success in image classification, speech recognition and other fields, demonstrating its powerful nonlinear representation and potential for deep data mining. In the field of seismic exploration, deep learning technology has been rapidly applied in the fields of seismic fault identification and first arrival picking.

[0071] In order to solve the problems existing in the prior art, an embodiment of the present invention provides a dual-channel denoising network model training method, a seismic data denoising method and a device, wherein the noisy seismic data and label data in the sample data mentioned in the embodiment of the present invention, and the noise-suppressed seismic data obtained by the dual-channel denoising network model are all time domain data.

[0072] The dual-channel denoising network model training method provided by the embodiment of the present invention has a flowchart referring to Figure 1 As shown, the following steps are included:

[0073] Step S101: constructing a sample data set based on the acquired noisy seismic data and its corresponding label data, where the label data is the seismic data after denoising;

[0074] Step S102: inputting the sample data of the sample data set into the constructed dual-channel denoising network model for model training, the dual-channel denoising network model includes a frequency domain channel network, a time domain channel network and a shared network; wherein the frequency domain channel network converts the sample data into frequency domain data, performs learning and training on the frequency domain data, and outputs first denoised data; the time domain channel network performs learning and training on the sample data, and outputs second denoised data; the first denoised data and the second denoised data are used as input data of the shared network, and the shared network outputs third denoised data;

[0075] Step S103: Based on the shared network output of the third denoising data and the sample data set, calculate the preset loss function. If the loss function meets the preset convergence condition or the number of training iterations reaches the preset threshold, obtain the trained dual-channel denoising network model; otherwise, adjust the network parameters of the dual-channel denoising network model and return to continue the model training step.

[0076] The dual-channel denoising network model training method provided by the embodiment of the present invention uses noisy seismic data and its corresponding label data as sample data to construct a sample data set; wherein the noisy seismic data can be seismic data containing multiple noises, and the label data is seismic data after denoising (i.e., seismic signal); the constructed dual-channel denoising network model is trained by using the noisy seismic data and the denoised seismic data, and when training the network model, the sample data is used to train both the frequency domain channel network and the time domain channel network included in the dual-channel denoising network model, wherein the frequency domain channel network converts the sample data into frequency domain data, and the frequency domain data is trained for learning, and the time domain channel network is trained for learning. The frequency domain channel network is used to learn and train the sample data, and the time domain denoising results output by the frequency domain channel network and the time domain channel network are used as the input data of the shared network. The denoising results are output after being fused by the shared network. That is, the dual-channel denoising network model is trained from the frequency domain and time domain channels of the sample data. Compared with the method of training the model on the sample data separately from the frequency domain or the time domain, the training result of the dual-channel denoising network model is more accurate. When the network model is used to denoise the seismic data, the denoising result obtained is also more accurate. After meeting the preset training conditions, a trained dual-channel denoising network model for denoising the seismic data is obtained. When the noisy seismic data is denoised using the trained dual-channel denoising network model, the denoised seismic data can be directly output for the input noisy seismic data. Compared with the existing seismic data denoising methods, there is no need to design a denoising algorithm corresponding to the specific noise for the different types of noise contained in the seismic data. Therefore, when denoising the seismic data, there is no need to perform separate parameter debugging on the denoising algorithm corresponding to the specific noise. The model has strong adaptability and improves the noise suppression efficiency of the seismic data. There is no need to make theoretical assumptions when performing denoising, which effectively improves the accuracy of noise recognition.

[0077] Furthermore, the dual-channel denoising network model training method, in step S101, obtains seismic data and label data corresponding to the seismic data; including:

[0078] The actual seismic data is used as the acquired seismic data containing noise, and the data processed by the actual seismic data using the selected noise suppression technology is used as the label data;

[0079] and / or,

[0080] The geological structures and geological bodies in the field are made into physical models according to the preset simulation similarity ratio, and the physical models are forward simulated to obtain physical model data as label data;

[0081] The labeled data is processed with noise added and used as noisy seismic data.

[0082] Specifically, when constructing a sample data set, noisy seismic data and its corresponding label data can be obtained in the following two ways:

[0083] Method 1: De-noising the actual seismic data through the selected seismic noise suppression technology, using the seismic data before de-noising as the noisy seismic data in the sample data set, and using the seismic data after de-noising as the label data in the sample data set; wherein, the actual seismic data can be obtained in previous exploration operations; the selected seismic noise suppression technology can be: abnormal amplitude suppression, FKK domain filtering, dip filtering and other existing combination technologies for de-noising seismic data.

[0084] Method 2: In the laboratory, the geological structures and geological bodies in the field are made into physical models according to a certain simulation similarity ratio (for example, the simulation similarity ratio can be 1:10000), and the field seismic exploration method is forward simulated by ultrasonic or laser ultrasonic methods to obtain seismic data with a high signal-to-noise ratio (i.e., physical model data); random noise, linear noise, etc. are added to the physical model data through numerical simulation methods as the noisy seismic data of the sample data set, and the physical model data is used as its corresponding label data.

[0085] When training a dual-channel denoising network model, the noisy seismic data obtained by method 1 and its corresponding label data can be used alone to construct a sample data set; the noisy seismic data obtained by method 2 and its corresponding label data can also be used to construct a sample data set; the noisy seismic data obtained by method 1 and method 2 and their corresponding label data can also be used simultaneously to construct a sample data set; the embodiment of the present invention does not specifically limit this. Preferably, the noisy seismic data obtained by methods 1 and 2 and their corresponding label data are used together to construct a sample data set. The advantages of constructing the sample data set in this way are as follows: Method 1 uses actual seismic data as the noisy seismic data in the sample data set, and uses the existing seismic noise suppression technology to denoise the noisy seismic data as the label data. This method cannot guarantee that the noise in the noisy seismic data is completely removed when obtaining the label data; and Method 2 uses the physical model data as the label data after obtaining the physical model data through physical simulation, and adds noise to the physical model data as the noisy seismic data. Therefore, the label data is the seismic data after the noise is completely removed, but the physical simulation cannot completely simulate the real seismic data; therefore, if the data obtained by one of the methods is used alone to construct the sample data set, the sample data in the sample data set will be relatively limited; and using the data obtained by the two methods to jointly construct the sample data set to train the dual-channel denoising network model can not only improve the generalization of the dual-channel denoising network model, but also improve the denoising accuracy of the network model.

[0086] Optionally, in order to continuously optimize the trained dual-channel denoising network model, new sample data may be continuously acquired and input into the model for training, so that the dual-channel denoising network model may be continuously optimized.

[0087] Among them, before training the constructed dual-channel denoising network model, it can also include: randomly initializing the network parameters of the dual-channel denoising network model, and when training the dual-channel denoising network model, inputting the sample data of the sample data set into the dual-channel denoising network model after the network parameters are initialized for training.

[0088] Further, the frequency domain channel network includes: a Fourier transform layer, a deep neural network layer and an inverse Fourier transform layer; in step S102, the frequency domain channel network converts the sample data into frequency domain data, performs learning and training on the frequency domain data, and outputs the first denoised data, including:

[0089] The Fourier transform layer performs Fourier transform on the noisy seismic data and label data to obtain frequency domain data;

[0090] The real part and imaginary part of the frequency domain data are used as input data of the deep neural network layer, the deep neural network layer is trained and the frequency domain denoised data is obtained;

[0091] The frequency domain denoised data is used as input data of the inverse Fourier transform layer, and the inverse Fourier transform layer performs inverse Fourier transform on the frequency domain denoised data and outputs the data to the first denoised data.

[0092] Among them, the first denoised data output by the frequency domain channel network of the embodiment of the present invention and the second denoised data output by the time domain channel network are both time domain data. After the first denoised data and the second denoised data are input into the shared network, the first denoised data and the second denoised data are fused in the time domain through the shared network and the third denoised data is output. The third denoised data is the final denoised data obtained by processing the noisy seismic data using the dual-channel network model.

[0093] The structural diagram of the dual-channel denoising network model of the embodiment of the present invention is shown in FIG. Figure 2 As shown in the figure, the dual-channel denoising network model consists of three branch networks: frequency domain channel network, time domain channel network and shared network. The frequency domain channel network branch network consists of FT layer (Fourier transform layer), deep neural network layer and IFT layer (inverse Fourier transform layer), among which the deep neural network layer consists of FE layer (feature extraction layer) and FR layer (feature reconstruction layer); the time domain channel network consists of FE layer (feature extraction layer) and FR layer (feature reconstruction layer); the shared network consists of ATT layer (self-attention module), FM layer (feature mapping layer) and OUT layer (output layer).

[0094] Among them, the FT layer has an input channel number of 1 and an output channel number of 2. Its function is to transform the input time domain seismic data into frequency domain data, and output the real and imaginary parts of the frequency domain data; the real and imaginary parts of the frequency domain data are used as input data of the deep neural network layer. The IFT layer has an input channel number of 2 and an output channel number of 1. Its function is to transform the input frequency domain seismic data into the time domain, and output the time domain data.

[0095] Specifically, in the dual-channel denoising network model of the embodiment of the present invention, the FE layer is composed of five groups of convolutional layers with 128 channels and a filter size of 3 and one group of convolutional layers with 128 channels, a filter size of 3 and a step size of 2, and each group of convolutional layers includes a ReLU activation function; the FR layer is composed of five groups of convolutional layers with 128 channels and a filter size of 3 and one group of deconvolutional layers with 128 channels, a filter size of 3 and a step size of 2, and each group of convolutional layers includes a ReLU activation function; the FM layer is composed of five groups of convolutional layers with 128 channels and a filter size of 3 and one group of deconvolutional layers with 128 channels, a filter size of 3 and a step size of 2, and each group of convolutional layers includes a ReLU activation function; The network is composed of 3 convolutional layers, each of which contains a ReLU activation function; the ATT layer is a channel self-attention module, which includes a group of pooling layers, a group of fully connected layers, a group of ReLU activation functions, a group of fully connected layers, and a group of Sigmoid activation functions; the output layer is composed of a group of convolutional layers with an input channel number of 128, an output channel number of 1, and a filter size of 3; in addition, in order to avoid the loss of features of seismic data during the denoising process, skip connections are added between feature maps of the same scale in the frequency domain channel network and the time domain channel network.

[0096] The network model training method of the embodiment of the present invention has a loss function calculation formula as follows:

[0097]

[0098] In the formula, K is the number of samples in the sample data set, x k is the noisy seismic data, y k is the label data, Θ represents the network parameters that the network needs to train and learn; represents the third denoised data output by the shared network; SIM() represents the structural similarity loss; ||·|| F represents the Frobenius norm; |·| represents the 1 norm; μ 1 、u 2 are penalty coefficients, where μ 1 、u 2 The specific value of can be selected according to actual needs, and the embodiment of the present invention does not make any specific limitation on this; for example, the default values ​​thereof can be 0.1 and 0.01 respectively.

[0099] In the process of training the constructed dual-channel denoising network model, if the loss function does not meet the preset convergence conditions or the number of training iterations does not reach the preset number threshold, the current network parameters of the dual-channel denoising network model can be adjusted through the back propagation algorithm, and the step of inputting the noisy seismic data in the sample data set into the dual-channel denoising network model for model training is returned to continue. At this time, the dual-channel denoising network model is a network model after the network parameters are adjusted, until the maximum number of training times is reached or the loss function meets the preset convergence conditions, so as to obtain a trained dual-channel denoising network model.

[0100] Among them, the default value of the preset number threshold of training iterations can be set to 50 times. The back propagation algorithm of the dual-channel denoising network is a process of obtaining the optimal network parameters by minimizing the loss function. Specifically, the minimization process of the loss function can be achieved through the Adam optimization algorithm.

[0101] In some optional embodiments, the obtained sample data can be divided into a test data set and a sample data set according to a preset ratio. Optionally, when setting the preset ratio of the test data set to the sample data set, the default value can be set to 10% of the total number of track sets for the test data set.

[0102] Among them, the dual-channel denoising network model is trained using a sample data set. During the training process, whether to stop the training of the dual-channel denoising network model is determined based on whether the calculated loss function meets the preset convergence condition or whether the number of training iterations reaches the preset number threshold. After the model meets the conditions for stopping training, the noisy seismic data of the test data set can be input to test the dual-channel denoising network model whose training is stopped, and the accuracy of the dual-channel denoising network model is determined by comparing the output result of the dual-channel denoising network model with the label data. If the test result does not reach the preset accuracy threshold, other parameters of the dual-channel denoising network model, such as the learning rate, etc., can be adjusted, and the dual-channel denoising network model with the adjusted parameters can be retrained using the sample data set. Alternatively, the sample data set can be reconstructed, and the dual-channel denoising network model can be trained using the reconstructed sample data set, etc., so that the output result of the dual-channel denoising network model can meet the accuracy requirements.

[0103] Optionally, if the preset number of training times is reached during model training, but the loss function has not converged, the dual-channel denoising network model can be retrained by adjusting other parameters of the dual-channel denoising network model or reconstructing the sample data set.

[0104] In view of the problems of low noise recognition accuracy, poor adaptability and low efficiency in the existing seismic noise suppression technology, an embodiment of the present invention provides a dual-channel denoising network model for seismic data noise suppression, and uses seismic data processed by conventional seismic noise suppression technology and seismic data collected based on physical models to jointly construct a denoising sample data set, and uses the sample data set to train the above dual-channel denoising network model, so that the trained dual-channel denoising network model is used for noise suppression of seismic data. The dual-channel denoising network model includes: a frequency domain channel network, a time domain channel network and a shared network, and the frequency domain channel network and the time domain channel network are used to complete the encoding and decoding of seismic data features in the frequency domain and the time domain respectively, thereby realizing the denoising processing of seismic data, and the denoising data output from the frequency domain channel network and the time domain channel network are merged by the shared network to obtain the final denoising data. Compared with conventional noise suppression technology, the present invention realizes the suppression of multiple seismic noises, and the dual-channel denoising network model has strong adaptability when performing noise processing of seismic data, and effectively improves the noise recognition accuracy on the basis of significantly improving the noise suppression efficiency.

[0105] In addition, when designing a dual-channel denoising network model, the embodiment of the present invention uses the noisy seismic data and its corresponding denoised seismic data as sample data to construct a sample data set; and uses the noisy seismic data and its corresponding noise data as sample data to construct a sample data set, and uses the constructed sample data to train the constructed network model for denoising of the noisy seismic data. After comparing the denoising results, it is found that the method of using the noisy seismic data and its corresponding denoised seismic data in the embodiment of the present invention as sample data to construct a sample data set to train the model, when performing denoising of seismic data, the accuracy of the processing result obtained is higher.

[0106] Based on the same inventive concept, the embodiment of the present invention also provides a dual-channel seismic data denoising method, referring to Figure 3 As shown, including:

[0107] Step S201: acquiring the seismic data to be processed, and performing trace equalization processing on the seismic data to be processed to obtain seismic data after trace equalization processing;

[0108] Step S202: inputting the seismic data after trace equalization into a pre-trained dual-channel denoising network model, and outputting the seismic data after noise suppression through the dual-channel denoising network model;

[0109] Step S203: performing inverse channel equalization processing on the noise-suppressed seismic data to obtain a denoising result of the seismic data (ie, a seismic signal).

[0110] Among them, the dual-channel denoising network model is trained using the above-mentioned dual-channel denoising network model training method.

[0111] In step S201, the seismic data to be processed is subjected to trace equalization processing to obtain seismic data after trace equalization processing, including:

[0112] Selecting the maximum value of the seismic data of each seismic channel in the seismic data to be processed as the channel equalization factor of the seismic data in the seismic channel;

[0113] The seismic data in each seismic channel is compared with its corresponding channel equalization factor to obtain the seismic data after channel equalization processing.

[0114] The formula for calculating the channel equalization process is as follows:

[0115]

[0116] Where, d i is the seismic data of the ith seismic channel, max() is the maximum value of the seismic data of the ith seismic channel, d i ′ is the seismic data after trace equalization, and g is the trace equalization factor.

[0117] In step S202, the seismic data after the trace equalization processing is input into the pre-trained dual-channel denoising network model, and the seismic data after the noise suppression is output through the dual-channel denoising network model, including:

[0118] Using the Fourier transform layer of the frequency domain channel network, the input seismic data after the trace equalization processing is Fourier transformed to obtain frequency domain data, using the deep neural network layer of the frequency domain channel network to calculate the frequency domain data to obtain frequency domain denoised data, and using the inverse Fourier transform layer to perform Fourier inverse transform on the frequency domain denoised data to obtain first denoised data;

[0119] The seismic data after the trace equalization processing is calculated by using the time domain channel network, and the second denoised data is output by the time domain channel network;

[0120] The first denoised data and the second denoised data are input as input data into a shared network for calculation, and the shared network outputs noise-suppressed seismic data.

[0121] That is, the seismic data after noise suppression

[0122] In step S203, the noise-suppressed seismic data is subjected to inverse equalization processing to obtain a denoising result of the seismic data, including:

[0123] The noise-suppressed seismic data is multiplied by the trace equalization factor corresponding to the seismic data to obtain the denoising result of the seismic data.

[0124] The calculation formula for inverse equalization processing of noise-suppressed seismic data is as follows:

[0125] Denoising results of seismic data

[0126] In an optional embodiment, after obtaining the denoising result of the seismic data (ie, the seismic signal), the seismic data denoising method can also obtain noise data based on the difference between the seismic data to be processed and the denoising result.

[0127] An embodiment of the present invention provides a dual-channel seismic data denoising method. When processing seismic data, the seismic data of each seismic channel is firstly subjected to channel balancing processing to make the seismic data more balanced and concentrated. This can avoid the problem that some data cannot be processed due to large differences in data between different seismic channels, and can further improve the accuracy of seismic data processing.

[0128] In an optional embodiment, Figure 4 As an example, the simulated seismic data shown in FIG. 1 includes simulated effective seismic signal data and real noise data, wherein the simulated effective seismic signal is shown in FIG. Figure 5 As shown, the real noise data is Figure 7 As shown, Figure 4 After the seismic data shown in FIG. 1 is processed by the dual-channel seismic data denoising method provided by the embodiment of the present invention, the denoising result obtained is as follows: Figure 6 As shown, according to the denoising result and the seismic data, the noise identified by the dual-channel seismic data provided by the embodiment of the present invention can be obtained as follows: Figure 8 As shown in the figure, it can be seen that the difference between the simulated effective seismic signal data and the denoising result (i.e., seismic signal) obtained after being processed by the denoising method provided by the embodiment of the present invention is very small, which can prove the effectiveness and accuracy of the denoising method of the present embodiment in suppressing noise; at the same time, the noise and Figure 7 Compared with the real noise in the image, the difference between the two is very small, and there is almost no effective seismic signal in the noise identified by the denoising method of this embodiment, which can prove that the denoising method provided by the embodiment of the present invention has high amplitude preservation while effectively suppressing noise.

[0129] In an optional embodiment, the seismic data denoising method of the embodiment of the present invention is further described by taking an application example of a certain exploration block. The noise of the exploration block is mainly 50Hz interference, abnormal amplitude and surface wave. The signal-to-noise ratio of the post-stack data of various noises is low, and the denoising is difficult, which affects the further processing and interpretation of the seismic data. The schematic diagram is shown in FIG. Fig. 9 shown.

[0130] Fig.10This is the composite denoising result after denoising using multiple conventional denoising technologies such as FKK filtering and abnormal amplitude suppression. When using conventional composite processes to denoise seismic data, experts are required to design denoising processes based on experience and set different denoising modules for different noise types. Each denoising module needs to repeatedly debug parameters, so the efficiency is relatively low. At the same time, for surface waves with severe scattering, although conventional methods can remove most of the noise, there are still obvious residues of multiple refraction waves. Fig.11 This is the denoising result of the seismic data denoising method according to the embodiment of the present invention. As can be seen from the figure, the denoising effect of the seismic data denoising method according to the embodiment of the present invention is significantly improved, and no parameter debugging and process design are required during the denoising process, and the denoising efficiency is higher.

[0131] in addition, Fig.12 and Fig.13 The noises identified by the conventional composite process and the method are respectively, and there is no obvious signal residue in both of them. This shows that the method of the present invention can identify noise efficiently and accurately, and has a high amplitude preservation performance.

[0132] It can also be seen from the above embodiments that the dual-channel seismic data denoising method of the embodiment of the present invention, compared with the conventional noise suppression technology, can suppress a variety of seismic noises, and effectively improve the noise identification accuracy on the basis of significantly improving the noise suppression efficiency.

[0133] In an optional embodiment, the present invention provides a dual-channel denoising network model training device, referring to Fig.14 As shown, including:

[0134] The first acquisition module 11: constructs a sample data set based on the acquired noisy seismic data and its corresponding label data, where the label data is the seismic data after denoising;

[0135] Training module 12: inputting sample data of the sample data set into the constructed dual-channel denoising network model for model training, wherein the dual-channel denoising network model includes a frequency domain channel network, a time domain channel network and a shared network; wherein the frequency domain channel network converts the sample data into frequency domain data, performs learning and training on the frequency domain data, and outputs first denoised data; the time domain channel network performs learning and training on the sample data, and outputs second denoised data; the first denoised data and the second denoised data are used as input data of the shared network, and the shared network outputs third denoised data;

[0136] Model adjustment module 13: Based on the third denoising data and sample data set output by the shared network, calculate the preset loss function. If the loss function meets the preset convergence condition or the number of training iterations reaches the preset threshold, obtain the trained dual-channel denoising network model; otherwise, adjust the network parameters of the dual-channel denoising network model and return to continue the model training step.

[0137] Regarding the dual-channel denoising network model training device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0138] In an optional embodiment, the present invention provides a dual-channel seismic data denoising device, referring to Fig.15 As shown, including:

[0139] The second acquisition module 21 is used to acquire the seismic data to be processed, and to perform trace equalization processing on the seismic data to obtain the seismic data after trace equalization processing;

[0140] Processing module 22: inputting the seismic data after trace equalization into a pre-trained dual-channel denoising network model, and outputting the seismic data after noise suppression through the dual-channel denoising network model;

[0141] The result acquisition module 23 performs inverse channel equalization processing on the noise-suppressed seismic data to obtain a denoising result of the seismic data.

[0142] Regarding the dual-channel seismic data denoising device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0143] Based on the same inventive concept, an embodiment of the present invention also provides a computer storage medium, which stores computer executable instructions. When the computer executable instructions are executed by a processor, the above-mentioned dual-channel denoising network model training method and the above-mentioned dual-channel seismic data denoising method are implemented.

[0144] An embodiment of the present invention also provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the dual-channel denoising network model training method and the dual-channel seismic data denoising method are implemented.

[0145] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0146] Unless otherwise specifically stated, terms such as processing, computing, calculating, determining, displaying, etc. may refer to the actions and / or processes of one or more processing or computing systems, or similar devices, which operate and convert data represented as physical (e.g., electronic) quantities within registers or memories of a processing system into other data similarly represented as physical quantities within memories, registers, or other such information storage, transmission, or display devices of the processing system. Information and signals may be represented using any of a variety of different techniques and methods. For example, data, instructions, commands, information, signals, bits, symbols, and chips mentioned throughout the above description may be represented by voltages, currents, electromagnetic waves, magnetic fields or particles, light fields or particles, or any combination thereof.

[0147] It should be understood that the specific order or hierarchy of steps in the disclosed process is an example of an exemplary method. Based on design preferences, it should be understood that the specific order or hierarchy of steps in the process can be rearranged without departing from the scope of the present invention. The attached method claims present the elements of the various steps in an exemplary order and are not intended to be limited to the specific order or hierarchy described.

[0148] In the above detailed description, various features are grouped together in a single embodiment to simplify the invention. This method of disclosure should not be interpreted as reflecting an intention that the embodiments of the claimed subject matter require more features than are clearly stated in each claim. On the contrary, as reflected in the appended claims, the invention is in a state of having less than all the features of the disclosed individual embodiments. Therefore, the appended claims are hereby expressly incorporated into the detailed description, with each claim standing alone as a separate preferred embodiment of the invention.

[0149] Those skilled in the art will also appreciate that the various illustrative logic blocks, modules, circuits, and algorithmic steps described in conjunction with the embodiments herein may all be implemented as electronic hardware, computer software, or a combination thereof. In order to clearly illustrate the interchangeability between hardware and software, various illustrative components, blocks, modules, circuits, and steps are generally described above around their functions. Whether such functions are implemented as hardware or software depends on the specific application and the design constraints imposed on the entire system. A skilled person may implement the described functions in an alternative manner for each specific application, but such implementation decisions should not be interpreted as departing from the scope of protection of the present invention.

[0150] The steps of the method or algorithm described in conjunction with the embodiments herein may be directly embodied as hardware, a software module executed by a processor, or a combination thereof. The software module may be located in a RAM memory, a flash memory, a ROM memory, an EPROM memory, an EEPROM memory, a register, a hard disk, a mobile disk, a CD-ROM, or any other form of storage medium well known in the art. An exemplary storage medium is connected to the processor so that the processor can read information from the storage medium and can write information to the storage medium. Of course, the storage medium may also be an integral part of the processor. The processor and the storage medium may be located in an ASIC. The ASIC may be located in a user terminal. Of course, the processor and the storage medium may also be present in a user terminal as discrete components.

[0151] For software implementation, the techniques described in this application can be implemented with modules (e.g., procedures, functions, etc.) that perform the functions described in this application. These software codes can be stored in a memory unit and executed by a processor. The memory unit can be implemented within the processor or outside the processor. In the latter case, it is coupled to the processor in a communication manner via various means, which are well known in the art.

[0152] The above description includes examples of one or more embodiments. Of course, it is impossible to describe all possible combinations of components or methods for the purpose of describing the above embodiments, but it should be recognized by those skilled in the art that the various embodiments may be further combined and arranged. Therefore, the embodiments described herein are intended to cover all such changes, modifications and variations that fall within the scope of protection of the appended claims. In addition, with respect to the term "comprising" used in the specification or claims, the word is covered in a manner similar to the term "including", just as "including," is explained as a transitional word in the claims. In addition, any term "or" used in the specification of the claims is intended to mean "non-exclusive or".

Claims

1. A dual-channel denoising network model training method, characterized in that: include: Constructing a sample data set based on the acquired noisy seismic data and its corresponding label data, wherein the label data is the seismic data after denoising; Inputting the sample data of the sample data set into the constructed dual-channel denoising network model for model training, the dual-channel denoising network model includes a frequency domain channel network, a time domain channel network and a shared network; wherein the frequency domain channel network converts the sample data into frequency domain data, performs learning and training on the frequency domain data, and outputs first denoised data; the time domain channel network performs learning and training on the sample data, and outputs second denoised data; the first denoised data and the second denoised data are used as input data of the shared network, and the shared network outputs third denoised data; Based on the third denoising data output by the shared network and the sample data set, a preset loss function is calculated. If the loss function meets the preset convergence condition or the number of training iterations reaches the preset threshold, a trained dual-channel denoising network model is obtained; otherwise, the network parameters of the dual-channel denoising network model are adjusted and then the step of continuing the model training is returned.

2. The dual-channel denoising network model training method according to claim 1, characterized in that: Acquiring seismic data and label data corresponding to the seismic data; including: Using actual seismic data as acquired seismic data containing noise, and using data processed by the actual seismic data using the selected noise suppression technology as label data; and / or, The geological structure and geological body in the field are made into a physical model according to a preset simulation similarity ratio, and the physical model is forward simulated to obtain physical model data as the label data; The data after adding noise to the label data is used as noisy seismic data.

3. The dual-channel denoising network model training method according to claim 1, characterized in that: The frequency domain channel network includes: a Fourier transform layer, a deep neural network layer and an inverse Fourier transform layer; the frequency domain channel network converts the sample data into frequency domain data, performs learning and training on the frequency domain data, and outputs first denoised data; including: The Fourier transform layer performs Fourier transform on the noisy seismic data and the label data to obtain frequency domain data; Using the real part and the imaginary part of the frequency domain data as input data of the deep neural network layer, training the deep neural network layer and obtaining frequency domain denoised data; The frequency domain denoised data is used as input data of an inverse Fourier transform layer, and the inverse Fourier transform layer performs an inverse Fourier transform on the frequency domain denoised data and outputs the result to the first denoised data.

4. The model training method according to claim 1, characterized in that: The calculation formula of the loss function is as follows: In the formula, K is the number of samples in the sample data set, x k is the earthquake data, y k is the label data, Θ represents the network parameters that the network needs to train and learn; represents the third denoised data output by the shared network; SIM() represents the structural similarity loss; ||·|| F represents the Frobenius norm; |·| represents the 1 norm; μ1 and u2 are penalty coefficients.

5. The model training method according to any one of claims 1 to 4, characterized in that: The network parameters of the dual-channel denoising network model are adjusted through the back-propagation algorithm.

6. A dual-channel seismic data denoising method, characterized in that: include: Acquiring seismic data to be processed, and performing trace equalization processing on the seismic data to be processed to obtain seismic data after trace equalization processing; Input the seismic data after trace equalization into the pre-trained dual-channel denoising network model, and output the seismic data after noise suppression through the dual-channel denoising network model; Performing inverse channel equalization processing on the noise-suppressed seismic data to obtain a denoising result of the seismic data; Wherein, the dual-channel denoising network model is trained using the dual-channel denoising network model training method as described in any one of claims 1-5.

7. The seismic data denoising method according to claim 6, characterized in that: Performing trace equalization processing on the seismic data to be processed to obtain seismic data after trace equalization processing, including: Selecting the maximum value of the seismic data of each seismic channel in the seismic data to be processed as the channel equalization factor in the seismic channel; The seismic data in each seismic channel is compared with its corresponding channel equalization factor to obtain the seismic data after channel equalization processing.

8. The seismic data denoising method according to claim 7, characterized in that: The seismic data after channel equalization processing is input into the pre-trained dual-channel denoising network model, and the seismic data after noise suppression is output through the dual-channel denoising network model, including: Using the Fourier transform layer of the frequency domain channel network, the input seismic data after the trace equalization processing is Fourier transformed to obtain frequency domain data, using the deep neural network layer of the frequency domain channel network to calculate the frequency domain data to obtain frequency domain denoised data, and using the inverse Fourier transform layer to perform Fourier inverse transform on the frequency domain denoised data to obtain first denoised data; The seismic data after the trace equalization processing is calculated by using the time domain channel network, and the second denoised data is output by the time domain channel network; The first denoised data and the second denoised data are input as input data into the shared network for calculation, and the shared network outputs the noise-suppressed seismic data.

9. The seismic data denoising method according to claim 8, characterized in that: Performing inverse equalization processing on the noise-suppressed seismic data to obtain a denoising result of the seismic data includes: The noise-suppressed seismic data is multiplied by the trace equalization factor corresponding to the seismic data to obtain the denoising result of the seismic data.

10. The seismic data denoising method according to any one of claims 6 to 9, characterized in that: Also includes: A difference operation is performed on the seismic data to be processed and the denoising result of the seismic data to obtain the noise data of the seismic data to be processed.

11. A dual-channel denoising network model training device, characterized in that: include: The first acquisition module: constructs a sample data set based on the acquired noisy seismic data and its corresponding label data, wherein the label data is the seismic data after denoising; Training module: inputting the sample data of the sample data set into the constructed dual-channel denoising network model for model training, wherein the dual-channel denoising network model includes a frequency domain channel network, a time domain channel network and a shared network; wherein the frequency domain channel network converts the sample data into frequency domain data, performs learning and training on the frequency domain data, and outputs first denoised data; the time domain channel network performs learning and training on the sample data, and outputs second denoised data; the first denoised data and the second denoised data are used as input data of the shared network, and the shared network outputs third denoised data; Model adjustment module: Based on the third denoising data output by the shared network and the sample data set, a preset loss function is calculated. If the loss function meets the preset convergence condition or the number of training iterations reaches the preset threshold, a trained dual-channel denoising network model is obtained; otherwise, the network parameters of the dual-channel denoising network model are adjusted and then the model training step is continued.

12. A dual-channel seismic data denoising device, characterized in that: include: The second acquisition module is used to acquire the seismic data to be processed, and to perform trace equalization processing on the seismic data to obtain seismic data after trace equalization processing; Processing module: input the seismic data after channel equalization into the pre-trained dual-channel denoising network model, and output the seismic data after noise suppression through the dual-channel denoising network model; The result acquisition module performs inverse equalization processing on the noise-suppressed seismic data to obtain a denoising result of the seismic data.

13. A computer storage medium, characterized in that: The computer storage medium stores computer executable instructions, which, when executed by a processor, implement the dual-channel denoising network model training method described in any one of claims 1-5 and the dual-channel seismic data denoising method described in any one of claims 6-10.

14. A computer device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the dual-channel denoising network model training method described in any one of claims 1 to 5 and the dual-channel seismic data denoising method described in any one of claims 6 to 10 are implemented.

Images