Earthquake low-frequency signal reconstruction method based on deep learning

Abstract

The invention discloses an earthquake low-frequency signal reconstruction method based on deep learning, and relates to the technical field of exploration geophysics. Preprocessing the original seismic data; performing network training; performing network prediction; and performing data recovery. According to the method, mapping from high-frequency data to low-frequency data is realized by building a convolutional neural network framework, and low-frequency signals are recovered; the features of the sample are closer to real data by using a demigration method, and the accuracy of network prediction is improved; and taking the prepared high-frequency sample as the input of the network, taking the low-frequency sample as the output of the network, using the two groups of data sets for network training, and using the trained CNN to predict low-frequency data from high-frequency data. According to the method, the local features in the signals can be automatically extracted, the requirement of artificial feature engineering is reduced, complex modes and relations are effectively captured through a multi-layer structure, the trained neural network can be suitable for a target work area by improving a training sample, and the prediction accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of exploration geophysical technology, and in particular to a seismic low-frequency signal reconstruction method based on deep learning. Background Art

[0002] Low-frequency signals are crucial in seismic data. The effective implementation of many technologies in exploration geophysics depends on low-frequency signals. For example, in the reverse time migration (RTM) imaging method, RTM is one of the commonly used imaging techniques in exploration geophysics. It uses observed seismic records and velocity model information to perform high-precision imaging of underground media. The wider the effective seismic signal band, the richer the underground information it carries, and the more accurately the imaging results can describe the underground medium structure. However, due to the imperfection of related technologies and theories, there is inevitably a problem of missing frequency bands during the acquisition or preprocessing of seismic records; for example, in full-waveform inversion (FWI), falling into local minima is one of the main weaknesses of many existing FWI algorithms. This may be caused by inaccurate initial velocity and lack of low-frequency data, which is the key to low-wavenumber velocity recovery. However, due to acquisition limitations and low-cutoff filters used in seismic data processing, there is insufficient low-frequency signal in seismic data.

[0003] There are many methods for dealing with the problem of frequency band extension of seismic data. Early methods include deconvolution, inverse Q filtering, spectral whitening, etc. The rapid development of artificial intelligence technology in recent years has provided a modern approach to solving traditional geophysical exploration problems. The research on frequency band extension of seismic data is not limited to the field of signal processing. The frequency band extension method of seismic data based on deep learning has opened up new solutions for researchers by training neural networks to predict missing low-frequency signals. However, the generalization ability of the network is an important constraint. Many times, the trained neural network fails to predict when applied to the target work area, and it is impossible to effectively compensate for the low-frequency signal.

[0004] How to effectively compensate for the missing frequency bands of seismic data in the target work area is one of the urgent problems to be solved in exploration geophysics. Summary of the invention

[0005] In order to solve the above technical problems, the present invention discloses a seismic low-frequency signal reconstruction method based on deep learning. The method builds a network framework based on deep learning theory, and builds a convolutional neural network (CNN) framework to realize the mapping from high-frequency data to low-frequency data. The effect of predicting the low-frequency signal missing in the seismic data is achieved by training the neural network. At the same time, the idea of ​​anti-migration is introduced when making training samples to make the characteristics of the samples closer to the real data. Corresponding training samples are made for the target work area, thereby improving the accuracy of network prediction.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A seismic low-frequency signal reconstruction method based on deep learning specifically comprises the following steps:

[0008] Step 1: Prepare training samples

[0009] Based on the post-stack reverse time migration imaging profile, broadband seismic records are obtained by using the reverse migration method. According to the actual frequency band of seismic data, broadband seismic records are divided into low-frequency and high-frequency training samples by filtering.

[0010] Step 2: Preprocess the original seismic data

[0011] On the basis of energy balance, the seismic data is made into smaller data sliders, and a suitable sliding window size and a suitable sliding step length are set. Several data sliders are slid on the seismic records in sequence, and the data sliders are used as training data. High-frequency block pairs and corresponding low-frequency block pairs are selected for training;

[0012] Step 3: Network training

[0013] Input the high-frequency data blocks and low-frequency data blocks preprocessed in step 2 into the CNN network architecture for network training;

[0014] Step 4: Perform network prediction

[0015] The low-frequency reconstruction neural network is trained to generate the low-frequency extension artificial intelligence neural network seismic data mapping relationship, and then the neural network is used for data prediction. The seismic data with limited frequency band is input into the trained network, and the low-frequency extension seismic data can be output, thereby recovering the weak reflection signal of the earthquake;

[0016] Step 5: Data recovery

[0017] The predicted low-frequency signal is restored and preprocessed, and the data mentioned in the data preprocessing is inversely transformed to ensure the fidelity of the predicted seismic data.

[0018] Furthermore, in step 2, the energy balance uses automatic gain control to balance the amplitude, and the formula is as follows:

[0019] ;

[0020] In the formula, is the jth sampling point of the ith seismic trace, Represents the data after corresponding energy balance, 2L+1 and 2K+1 are the horizontal and vertical lengths of the calculation window respectively.

[0021] Furthermore, in step 2, a sliding window size of 128×128 and a sliding step size of 16 are set, and a plurality of data sliders are obtained by sliding on the seismic record in sequence.

[0022] Furthermore, in step three, the CNN network architecture includes an input layer, seven convolutional layers, two activation functions and an output layer. The size of the input data block is 128×128. Each convolutional layer uses 32 convolution kernels of a fixed size of 9×9 to capture the feature map. Except for the output layer, all other layers use rectified linear units, where the tanh function is used to fit the data to the [-1, 1] interval. At the same time, data normalization is introduced in some layers to accelerate convergence.

[0023] The beneficial effect of the present invention is that a convolutional neural network (CNN) framework is constructed to realize the mapping from high-frequency data to low-frequency data so as to restore the low-frequency signal. At the same time, the idea of ​​anti-drift is introduced when preparing training samples so that the characteristics of the samples are closer to the real data, thereby improving the accuracy of network prediction. The prepared high-frequency samples are used as input and the corresponding low-frequency samples are used as output. The characteristics between the high-frequency signal and the low-frequency signal are learned through the neural network. The trained network can be used for low-frequency prediction to predict seismic data lacking low frequency. Finally, a simulation experiment is carried out using a pushover model to prove the accuracy of the method of the present invention.

[0024] This method uses a neural network to automatically extract local features from the signal, reducing the need for manual feature engineering and effectively capturing complex patterns and relationships through a multi-layer structure. In addition, by improving the training samples, the trained neural network can be applied to the target work area, improving prediction accuracy and solving the problem of lack of low-frequency signals in seismic records. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic diagram of the process of the present invention;

[0026] Figure 2 For the implementation of the present invention, the example lacks low-frequency seismic records;

[0027] Figure 3 is a broadband sample obtained by de-migrating in an embodiment of the present invention;

[0028] Figure 4A Filter 5-10 Hz samples for the demistified samples in the embodiment of the present invention;

[0029] Figure 4B Filter 0-5 Hz samples for the anti-offset samples in the embodiment of the present invention;

[0030] Figure 5A This is the result after 5-10 Hz sample preprocessing in the embodiment of the present invention;

[0031] Figure 5B This is the result after 0-5Hz sample preprocessing in the embodiment of the present invention;

[0032] Fig. 6A This is the sliding block division after the 5-10 Hz sample preprocessing in the embodiment of the present invention;

[0033] Figure 6B This is the sliding block division after the 0-5Hz samples are preprocessed in the embodiment of the present invention;

[0034] Figure 7 A neural network architecture diagram in an embodiment of the present invention;

[0035] Figure 8 is the de-migration shot gather prediction result that is not involved in the training in the embodiment of the present invention;

[0036] Fig. 9 The prediction result of the frequency-deficient earthquake record of the pushover model in the embodiment of the present invention;

[0037] Fig.10 A is the low frequency prediction result in the embodiment of the present invention;

[0038] Fig.10 B is the preprocessing of restoring the prediction results in the embodiment of the present invention;

[0039] Fig.10 C is the real low-frequency result in the embodiment of the present invention;

[0040] Fig.11A The spectrum diagram of the embodiment of the present invention lacks low-frequency earthquake records;

[0041] Fig. 11B It is a spectrum diagram of the seismic record after broadband reconstruction in an embodiment of the present invention. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] The present invention discloses a method for reconstructing low-frequency seismic signals based on deep learning, such as Figure 1 As shown, the following steps are included:

[0044] S1. Prepare training samples

[0045] The key to deep learning neural network signal reconstruction lies in the selection of network training samples. The traditional method is to obtain broadband seismic records as training samples by forward modeling. However, when the seismic records forward modeled by the model are significantly different from the seismic records to be reconstructed, it is usually difficult to achieve low-frequency prediction, which affects the generalization ability of the network.

[0046] In order to ensure that the low-frequency reconstruction can be applied to the target work area, the broadband seismic records are obtained by using the reverse time migration (RTM) imaging section based on the post-stack Figure 3 As shown, the broadband de-migrated samples are filtered to obtain relatively high frequency samples of 5-10 Hz ( Figure 4A ) and low frequency samples from 0-5Hz ( Figure 4B ).

[0047] S2. Preprocessing of original seismic data

[0048] Data preprocessing is a key step that affects network training and generalization capabilities. The amplitude of seismic data decreases with increasing propagation distance, and the received reflected wave signal becomes weaker with increasing depth. Therefore, it is necessary to preprocess the original seismic data to balance the energy and enhance weak signals.

[0049] In order to prevent the high-energy direct waves from shielding the reflected wave information during the training process, the direct waves were removed when preparing the samples. Considering the large changes in seismic records, even after removing the direct waves, the energy of the refracted waves is much greater than the energy of the reflected waves. In this case, if the normalized seismic data is directly input into the network, the contribution of the reflected waves to the network training will be too small, thus affecting the network performance and the final imaging results. Therefore, automatic gain control (AGC) is used to balance the amplitude, and the formula is as follows:

[0050] ;

[0051] In the formula, is the jth sampling point of the ith seismic trace, It represents the data after energy balance, 2L+1 and 2K+1 are the horizontal and vertical lengths of the calculation window respectively. Figure 5A A and B show the effects of 5-10Hz and 0-5Hz samples after energy balance.

[0052] Smaller data blocks can better learn the features in them. Based on energy balance, the seismic data is made into smaller data sliders. A sliding window size of 128×128 and a sliding step of 16 are set to slide on the seismic records in sequence to obtain several data sliders, such as Fig. 6A , 6BRepresent high-frequency sample sliders and low-frequency sample sliders respectively. Data sliders can be regarded as the basic elements of a gun set. Therefore, they can be used as training data to enable us to obtain good network generalization capabilities in many applications. The data block method is introduced to divide the training data into high-frequency and low-frequency block pairs for training CNN.

[0053] S3. Perform network training

[0054] The proposed CNN architecture is as follows: Figure 7 As shown, it includes an input layer, 7 convolutional layers, two activation functions and an output layer. The size of the input data block is 128×128. Each convolutional layer uses 32 convolution kernels of fixed size 9×9 to capture the feature map. Except for the last output layer, all layers use rectified linear units (ReLU), where the tanh function is used to fit the data to the [-1, 1] interval, and data normalization is introduced in some layers to accelerate convergence. The high-frequency data blocks and low-frequency data blocks preprocessed in step S2 are input into the CNN network architecture to start training the network.

[0055] S4. Perform network prediction

[0056] In step S1, a total of 100 shot reverse migration seismic shot data samples are generated, 50 of which are used for network training, and the remaining 50 are used for network prediction. In step S3, a trained convolutional neural network (CNN) can be obtained, which can predict low-frequency signals for band-limited seismic records. The 50 shot seismic shot data that are not involved in the training are filtered to filter out the high-frequency components of 5-10 Hz as band-limited seismic records to predict their low-frequency signals of 0-5 Hz. The prediction results are as follows: Figure 8 As shown in the figure, the four sub-graphs in the figure represent the 5-10Hz high-frequency signal, the predicted low-frequency result, the actual low-frequency result, and the residual of the predicted low-frequency and the actual low-frequency from left to right. It can be seen from the residual graph that the errors of the two are within the acceptable range, and the network training effect is good. Then, the trained network is used to predict the frequency-deficient earthquake records of the overthrow model.

[0057] Similarly, the seismic records obtained by the pushover model are also preprocessed in the same way and then the 5-10 Hz frequency components are filtered to predict the 0-5 Hz low-frequency signal. The prediction results are as follows: Fig. 9 As shown in the figure, it can be seen from the residual graph that the trained network also achieves good results on the pushover model.

[0058] S5. Perform data recovery

[0059] The prediction result obtained in step S4 actually includes various operations of data preprocessing in step S2. In order to ensure the fidelity of the predicted seismic data, it is necessary to perform recovery preprocessing on the predicted low-frequency signal and perform inverse transformation operations on the data mentioned in the data preprocessing, including inverse transformation energy balance correction. Fig.10 A shows the predicted low-frequency signal. Fig.10 B is the result after the prediction result is restored and preprocessed. Fig.10 C is the real low-frequency signal.

[0060] In order to visually see the effect of low-frequency signal reconstruction, the spectrum of the original earthquake record lacking low-frequency signals and the spectrum of the earthquake record reconstructed to include low-frequency signals are plotted as follows: Fig.11A , B, from Fig. 11B It can be seen that the signal is effectively restored in the frequency range of 0-5Hz.

[0061] This method builds a network framework based on deep learning theory, develops a low-frequency signal reconstruction method based on deep learning, and builds a convolutional neural network (CNN) framework to realize the idea of ​​mapping from high-frequency data to low-frequency data. At the same time, the idea of ​​anti-offset is introduced when making training samples to make the characteristics of the samples closer to the target work area data. Corresponding training samples are made for the target work area, thereby improving the accuracy of network prediction and solving the problem of poor migration of neural network prediction of low frequency between different data.

[0062] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A seismic low-frequency signal reconstruction method based on deep learning, characterized in that: Specifically include the following steps: Step 1: Prepare training samples Based on the post-stack reverse time migration imaging profile, broadband seismic records are obtained by using the reverse migration method. According to the actual frequency band of seismic data, broadband seismic records are divided into low-frequency and high-frequency training samples by filtering. Step 2: Preprocess the original seismic data On the basis of energy balance, the seismic data is made into smaller data sliders, and a suitable sliding window size and a suitable sliding step length are set. Several data sliders are slid on the seismic records in sequence, and the data sliders are used as training data. High-frequency block pairs and corresponding low-frequency block pairs are selected for training; Step 3: Network training Input the high-frequency data blocks and low-frequency data blocks preprocessed in step 2 into the CNN network architecture for network training; Step 4: Perform network prediction The low-frequency reconstruction neural network is trained to generate the low-frequency extension artificial intelligence neural network seismic data mapping relationship, and then the neural network is used for data prediction. The seismic data with limited frequency band is input into the trained network, and the low-frequency extension seismic data can be output, thereby recovering the weak reflection signal of the earthquake; Step 5: Data recovery The predicted low-frequency signal is restored and preprocessed, and the data mentioned in the data preprocessing are inversely transformed to ensure the fidelity of the predicted seismic data.

2. A method for reconstructing seismic low-frequency signals based on deep learning as claimed in claim 2, characterized in that: In step 2, energy balance uses automatic gain control to balance the amplitude, the formula is as follows: ； In the formula, is the jth sampling point of the ith seismic trace, Represents the data after energy balance, 2L+1 and 2K+1 are the horizontal and vertical lengths of the calculation window respectively.

3. A method for reconstructing seismic low-frequency signals based on deep learning as claimed in claim 2, characterized in that: In step 2, a sliding window size of 128×128 and a sliding step size of 16 are set, and a number of data sliders are obtained by sliding on the seismic records in sequence.

4. A method for reconstructing seismic low-frequency signals based on deep learning as claimed in claim 3, characterized in that: In step three, the CNN network architecture includes an input layer, seven convolutional layers, two activation functions and an output layer. The size of the input data block is 128×128. Each convolutional layer uses 32 convolution kernels of a fixed size of 9×9 to capture the feature map. All layers except the output layer use rectified linear units, where the tanh function is used to fit the data to the [-1, 1] interval. At the same time, data normalization is introduced in some layers to accelerate convergence.

Images