A method for enhancing virtual shot signals from a few seismic sources based on convolutional neural networks.

Abstract

This invention relates to a method for enhancing virtual shot signals from a small number of seismic sources based on a convolutional neural network (CNN). The CNN identifies and suppresses coherent noise and spurious phase axes in seismic data. Virtual shot records with a small number of passive seismic sources are used as training data, while those with a larger number of sources are used as training labels. The characteristics of the effective signal are learned from these labels to suppress coherent noise and spurious phase axes, and to restore waveforms in areas of discontinuity. For virtual shot records with uneven source distribution, those with a wider distribution are used as labels. In this case, the network's task is not only to suppress coherent noise and spurious phase axes and restore waveform continuity and extension, but also to restore intersecting linear phase axes to hyperbolic phase axes. This reduces the impact of passive source acquisition on seismic records, ultimately achieving better results and improving the applicability of passive source seismic exploration.

Description

Technical Field

[0001] This invention belongs to the field of seismic exploration technology, specifically relating to a method for processing passive source virtual shot records using deep learning, and particularly to a method for enhancing virtual shot signals from a small number of seismic sources based on convolutional neural networks. Background Technology

[0002] In seismic exploration, underground noise often poses a significant challenge for researchers. Its signals are weak but widely distributed and chaotic, typically requiring various methods to eliminate. However, in recent years, it has been discovered that its propagation patterns in the subsurface medium are similar to conventional seismic waves, and it carries a wealth of real-world subsurface information. If properly utilized, it can serve as a substitute for active source seismic sources. In scenarios where active source excitation is not feasible, such as near cities or protected areas, blasting is unnecessary; simply deploying detectors to receive passive source signals is sufficient for data acquisition, saving on blasting costs. Furthermore, because underground noise sources possess a rich frequency range, including much low-frequency information, passive sources offer a greater advantage over active source seismic exploration in large-scale, deep-seismic exploration.

[0003] Passive source signals are generally classified into transient sources and noisy sources. For transient sources, it is usually necessary to extract the sampling time period containing the signal for seismic interferometry reconstruction, and then superimpose multiple reconstructed records to obtain a reconstructed record with a higher signal-to-noise ratio. For noisy sources, after recording for a certain period of time, seismic interferometry reconstruction is often performed directly on the seismic record, and then a virtual shot record is reconstructed according to requirements.

[0004] However, when reconstructing passive source seismic data using seismic interferometry, coherent noise and spurious phase axes of varying intensities inevitably occur. Some of these interferences originate from underground noise, while others arise from the sparse or unevenly distributed underground seismic sources during seismic interferometry reconstruction. Removing the interference from coherent noise and spurious phase axes has been a long-standing goal for geophysicists. Yilmaz proposed FK filtering to suppress coherent noise, but when the coherent signal overlaps with the effective signal, the processing may damage the effective wave. Douglas et al., Mauricio et al., and Akerberg et al. have all attempted to suppress coherent noise in the Radon domain, but the problem of low resolution remains. Rabiner et al. created the median filtering denoising method, and many derivative algorithms have emerged, such as weighted median filtering and multi-level median filtering. These can operate on pixels in the image, but small windows are ineffective against coherent noise and spurious phase axes, while large windows blur the effective signal.

[0005] While the above methods have some effect on removing coherent noise, none can achieve completely satisfactory results. Furthermore, their calculations are complex and time-consuming, placing high demands on computing equipment. Parameter selection requires considerable experience from the personnel involved, introducing a degree of subjectivity. Moreover, they are ineffective in dealing with spurious in-phase axes and recovering valid waveforms.

[0006] Deep learning is a branch of machine learning that trains neural networks to learn deep features from information such as images. LeCun et al. first invented convolutional neural networks, which achieved good results in handwritten digit recognition. Hinton et al.'s AlexNet achieved excellent results in image recognition. Various deep learning networks such as VggNet, ResNet, FCN, DnCNN, and UNet have achieved good results in classification problems and image segmentation. While maintaining accuracy comparable to manual recognition and conventional methods, they are also more efficient and have been widely developed in various fields.

[0007] In recent years, thanks to the upgrading of computer hardware, deep learning algorithms have been widely used in the field of geophysics. Kong et al. used neural networks to distinguish seismic signals from active noise. Waldeland et al. used convolutional neural networks to interpret seismic profiles. Qian et al. used a DCAE network to identify pre-stack seismic phases. Wu et al. used convolutional neural networks to pick out the in-phase axes of microseismic signals. Di proposed a deconvolutional network (DCNN), which effectively identified and interpreted features in seismic images. Mandelli et al. used an autoencoder to reconstruct missing seismic traces. In terms of seismic data denoising, Zhu et al. created a Deep Denoiser using convolutional neural networks, which can still effectively suppress noise even in high-noise backgrounds. Sun et al. converted coherent noise from marine mining into random noise and used convolutional neural networks to suppress it. Song et al. used a UNet network to more precisely suppress multiple wave interference. Yu et al. used CNN to remove various types of noise, demonstrating the powerful data processing capabilities and applicability of convolutional neural networks.

[0008] In summary, if a passive source seismic exploration data processing method can be developed, utilizing convolutional neural networks to identify and suppress coherent noise and spurious phase axes in seismic data, this would be a novel approach that could effectively reduce the impact on seismic records and improve the applicability of passive source seismic exploration, thus possessing broad application value. Summary of the Invention

[0009] The purpose of this invention is to provide a method for enhancing virtual shot signals with a small number of passive source earthquakes based on convolutional neural networks, in order to solve the problems that virtual shot records contain a large amount of coherent noise and false phase axes, as well as poor continuity of phase axes and morphology that does not conform to the laws of physics when the number of passive source earthquakes is small and the distribution of earthquakes is uneven.

[0010] The objective of this invention is achieved through the following technical solutions:

[0011] A method for enhancing virtual shot signals from a small number of seismic sources based on convolutional neural networks is implemented using the Tensorflow-GPU framework. The method involves reconstructing virtual shot records by cross-correlation of the original passive source observation data. The core idea is to input the reconstructed virtual shot records as a dataset and labels into the neural network, allowing it to iteratively update parameters, learn the characteristics of the effective signal, preserve them, and perform a certain degree of signal enhancement; and learn the characteristics of coherent noise and spurious phase axes, removing them. This method not only reduces the influence of the processor's subjective experience on data processing, but also accurately removes coherent noise and spurious phase axes, while preserving the effective signal. It is not only more effective but also less time-consuming.

[0012] A method for enhancing virtual shot signals from a small number of seismic sources based on convolutional neural networks includes the following steps:

[0013] a. Use forward modeling to collect original passive source earthquake records for different velocity models, including those with a small number of earthquake sources, those with uneven earthquake source distribution, and those with a large number of earthquake sources and wide distribution.

[0014] b. Using cross-correlation seismic interferometry, a virtual shot record is constructed from the collected passive source raw records. The construction equation is as follows:

[0015] R(x B ,x A ,t)+R(x B ,x A ,-t)=δ(x H,B ,x H,A )δ(t)-T(x A ,-t)*T(x B ,t)

[0016] Where R(x) B ,x A ,t) represents the seismic reflection record of one shot relative to another shot, i.e., x A Fire blasting at location x B The earthquake records obtained by receiving them at the location; R(x) B ,x A ,-t) represents its non-causal part; δ() represents the Dirac function; T(x A,-t) and T(x B ,t) represent x respectively A and x B The transmitted wave response received by the two detectors; "*" represents convolution operation;

[0017] c. Extract virtual shot records with an appropriate number of sampling points to create training and testing sets, along with their corresponding labels, and then normalize them.

[0018]

[0019] Where x represents a pixel in the seismic reconstruction record, x min x represents the value of the smallest pixel in the seismic reconstruction record where x is located. max This represents the value of the largest pixel in the seismic reconstruction record where x is located.

[0020] d. Design the model framework of the convolutional neural network and set reasonable hyperparameters;

[0021] e. Input the prepared training set and its corresponding labels into the designed network, train the convolutional neural network, and iteratively update the weight parameters.

[0022] f. Input the test set data into the trained network to obtain a virtual shot record with low coherent noise level, fewer false phase axes, and good waveform continuity, and restore the linear phase axis to a hyperbolic phase axis.

[0023] Furthermore, step c employs single-record normalization, using the maximum and minimum values ​​of each seismic record for normalization processing to ensure high contrast and horizontal continuity of pixels in each record.

[0024] Compared with existing technologies, the advantages of this invention are as follows: This invention proposes a method for enhancing virtual shot signals based on convolutional neural networks. When there are few underground noise sources, the virtual shot record reconstructed from passive sources will exhibit certain coherent noise and spurious phase axes. By simply inputting the virtual shot record into the network, we can obtain a virtual shot record with better signal quality with high efficiency. Specifically:

[0025] By utilizing a convolutional neural network model and training data, the network is able to effectively suppress coherent noise, thus effectively suppressing the coherent noise generated in the virtual shot record due to the limited number of seismic sources or the earthquake interferometric calculation process.

[0026] By utilizing convolutional neural network models and training data, the network can be equipped with a good ability to identify spurious in-phase axes, which can be used to identify and eliminate spurious in-phase axes that do not conform to the laws of physics during actual processing.

[0027] By utilizing convolutional neural network models and training data, the network can be made to have a good ability to recover the in-phase axis. When the continuity of the in-phase axis is poor or the lateral continuity is poor, it can be recovered into an in-phase axis with better continuity.

[0028] The method of virtual shot signal enhancement based on convolutional neural networks is easy to implement and can be achieved with a well-designed network model. While ensuring high efficiency, it can better handle coherent noise and spurious phase axes, improve the continuity of effective signals, significantly improve the signal-to-noise ratio of virtual shot records, effectively expand the applicability of passive source seismic exploration, and is very suitable for data enhancement in passive source seismic exploration. Attached Figure Description

[0029] Figure 1 shows the forward velocity model, where... Figure 1a Training the model, Figure 1b Test model;

[0030] Figure 2 shows the location of the earthquake focal points. Figure 2a A schematic diagram showing the distribution of a small number of earthquake sources. Figure 2b Schematic diagram of non-uniform seismic source distribution; Figure 2c A schematic diagram showing a large number of widely distributed earthquake sources;

[0031] Figure 3 Source wavelet (random noise);

[0032] Figure 4 shows the virtual seismic source record. Figure 4a A small number of hypothetical seismic sources were recorded. Figure 4b Non-uniformly distributed virtual source records Figure 4c The number of hypothetical earthquake sources is large and they are widely distributed;

[0033] Figure 5 Convolutional Neural Networks;

[0034] Figure 6 shows the processing results of a small number of hypothetical earthquake source records and their corresponding labels. Figure 6a A small number of hypothetical seismic sources were recorded. Figure 6b Processing results of a small number of hypothetical seismic source records. Figure 6c Numerous hypothetical earthquakes were recorded with widely distributed hypothetical sources.

[0035] Figure 7 shows the processing results of virtual source records with unevenly distributed seismic sources and their corresponding labels. Figure 7a Records of virtual seismic sources with unevenly distributed sources; Figure 7b Processing results of virtual seismic source records with unevenly distributed seismic sources; Figure 7c Numerous hypothetical earthquake sources are widely distributed in the record of virtual earthquake sources. Detailed Implementation

[0036] This invention utilizes a convolutional neural network to identify and suppress coherent noise and spurious phase axes in seismic data. Virtual shot records with a small number of passive source sources are used as training data, while those with a larger number of sources are used as training labels. The network learns the characteristics of the effective signal from these labels to suppress coherent noise and spurious phase axes, and to restore waveforms in areas of discontinuity. For virtual shot records with uneven source distribution, this invention uses those with a wider distribution as labels. In this case, the network's task is not only to suppress coherent noise and spurious phase axes and restore waveform continuity and extension, but also to restore intersecting linear phase axes to hyperbolic phase axes. Although these are two tasks, this invention uses the same network model to implement them, and employs a large number of evenly distributed virtual shot records as training labels for both tasks. Therefore, when passive source acquisition occurs, this invention can reduce the impact of these two undesirable conditions on the seismic record, ultimately achieving better results and improving the applicability of passive source seismic exploration.

[0037] The present invention provides a method for enhancing virtual shot signals from a small number of seismic sources based on convolutional neural networks, comprising the following steps: 1. Setting up the MATLAB installation environment and installing the MATLAB Parallel Computing Toolbox;

[0038] 2. Using forward modeling techniques, we collected original passive source earthquake records for different velocity models, including those with a small number of earthquake sources, uneven distribution of earthquake sources, and a large number of earthquake sources with uniform distribution.

[0039] 3. Using the cross-correlation seismic interferometry method, a virtual shot record is constructed from the collected passive source raw records. The construction equation is as follows:

[0040] R(x B ,x A ,t)+R(x B ,x A ,-t)=δ(x H,B ,x H,A )δ(t)-T(x A ,-t)*T(x B ,t)

[0041] Where R(x) B ,x A ,t) represents the seismic reflection record of one shot relative to another shot, i.e., x A Fire blasting at location x B The earthquake records obtained by receiving them at the location; R(x) B ,x A ,t) represents its non-causal part; δ() represents the Dirac function; T(xA ,-t) and T(x B (,-t) represent x respectively A and x B The transmitted wave response received by the two detectors;

[0042] 4. Extract virtual shot records with an appropriate number of sampling points to create training and testing sets, along with their corresponding labels. To prevent gradient update issues during training, normalization is employed.

[0043]

[0044] Here, x represents a pixel in the seismic reconstruction record. min x represents the value of the smallest pixel in the seismic reconstruction record where x is located. max This represents the value of the largest pixel in the seismic reconstruction record where x is located. This invention does not use global extreme values ​​for normalization of all data to prevent differences between different seismic records from causing insufficient contrast in some records; nor does it use the commonly used method of taking extreme values ​​per trace for normalization to prevent the processed data from exhibiting a Venetian blind-like imbalance due to differences between traces. This invention employs single-record normalization, using the maximum and minimum values ​​of each seismic record for normalization processing, ensuring high contrast and horizontal continuity of pixels in each record.

[0045] 5. Design the model framework of the convolutional neural network and set reasonable hyperparameters;

[0046] 6. Input the prepared training set and its corresponding labels into the designed network, train the convolutional neural network, and iteratively update the weight parameters;

[0047] 7. Input the test set data into the trained network to obtain virtual shot records with low coherent noise levels, fewer false phase axes, and good waveform continuity, and restore the linear phase axis to a hyperbolic phase axis.

[0048] This invention solves the problems of removing coherent noise and false phase axes from virtual shots with few seismic sources, restoring waveform continuity, and suppressing coherent noise and false phase axes to restore waveform continuity and extensibility when the seismic source distribution is uneven. It also restores the linear intersecting phase axis to a hyperbolic phase axis.

[0049] Example 1

[0050] According to the exploration requirements, Parallel Computing Toolbox and MATLAB DistributedComputing Server were installed on Windows 10 Professional Edition system to build a MATLAB parallel platform.

[0051] Install Anaconda on a Windows 10 Professional system and configure the toolkits required for deep learning (CUDA9, CuDnn9.0).

[0052] The forward velocity model (64×128) and source distribution model are set up as shown in Figure 1. The grid spacing is set to 2m, and the velocity range is between 250m / s and 650m / s. Seismic detectors are placed on the model surface, with each grid point serving as a detector, and the detectors are spaced 2m apart. The number and distribution of seismic sources are shown in Figure 2. The source wavelet is selected as follows: Figure 3 The random noise sequence shown has a frequency between 4 Hz and 20 Hz. The sampling interval is 0.001 s.

[0053] The obtained noise records, after cross-correlation, yielded virtual source records, as shown in Figure 4. It can be seen that when the number of sources is small, the reconstructed virtual source seismic records exhibit coherent noise and spurious phase axes, as well as sections with discontinuous waveforms. Virtual shot records with unevenly distributed sources not only show coherent noise and spurious phase axes with poor waveform extension, but also exhibit linear intersecting phase axes that do not conform to physical laws. In contrast, virtual source records with a large number of evenly distributed sources not only have continuous waveforms and good extension, but also exhibit a hyperbolic shape that conforms to physical laws. Although some noise and spurious phase axes still exist, their distribution location, shape, and intensity do not completely overlap with the two types of virtual source records mentioned above.

[0054] In this embodiment, a small number of virtual shot records and non-uniformly distributed virtual shot records under a portion of the velocity model are used as training data, while virtual shot records under a large number of widely distributed sources are used as training labels; another small number of virtual shot records and non-uniformly distributed virtual shot records under a portion of the velocity model are used as test data, while virtual shot records under a large number of widely distributed sources are used as test labels.

[0055] In this embodiment, the prepared training data is used to train the designed data, such as... Figure 5 The convolutional neural network model shown was trained and iterated 1000 times. The trained network was then used to process the test data, and the results are as follows.

[0056] Figure 6 shows the reconstructed hypothetical seismic records from a small number of sources. After processing by our neural network, we can observe that coherent noise and spurious phase axes are effectively suppressed, and areas with insufficient waveform continuity are also effectively repaired. Compared with the test labels, the waveform of our processed data is cleaner, and the noise level and energy of spurious phase axes are lower.

[0057] Figure 7 shows the reconstructed virtual source seismic record from the non-uniform source. It can be observed that not only are the coherent noise and false phase axes effectively suppressed, but the phase axes with intersecting straight lines are also restored to hyperbolic phase axes with good continuity.

Claims

1. A method for enhancing virtual shot signals from a small number of seismic sources based on convolutional neural networks, characterized in that, Includes the following steps: a. Use forward modeling to collect original passive source earthquake records for different velocity models, including those with a small number of earthquake sources, uneven distribution of earthquake sources, and a large number of earthquake sources with wide distribution. b. Using cross-correlation seismic interferometry, a virtual shot record is constructed from the collected passive source raw records. The construction equation is as follows: R(x B ,x A ,t)+R(x B ,x A ,-t)=δ(x H,B ,x H,A )δ(t)-T(x A ,-t)*T(x B ,t) Where R(x) B ,x A ,t) represents the seismic reflection record of one shot relative to another shot, i.e., x A Fire blasting at location x B The earthquake records obtained by receiving them at the location; R(x) B ,x A ,-t) represents its non-causal part; δ() represents the Dirac function; T(x A ,-t) and T(x B ,t) represent x respectively A and x B The transmitted wave response received by the two detectors; "*" represents convolution operation; c. Extract virtual shot records with an appropriate number of sampling points to create training and testing sets, along with their corresponding labels, and then normalize them. Where x represents a pixel in the seismic reconstruction record, x min x represents the value of the smallest pixel in the seismic reconstruction record where x is located. max This represents the value of the largest pixel in the seismic reconstruction record where x is located. d. Design the model framework of the convolutional neural network and set reasonable hyperparameters; e. Input the prepared training set and its corresponding labels into the designed network, train the convolutional neural network, and iteratively update the weight parameters. f. Input the test set data into the trained network to obtain a virtual shot record with low coherent noise level, fewer false phase axes, and good waveform continuity, and restore the linear phase axis to a hyperbolic phase axis.

2. The method for enhancing virtual shot signals from a small number of seismic sources based on a convolutional neural network according to claim 1, characterized in that: Step c uses single-record normalization, which uses the maximum and minimum values ​​of each seismic record for normalization to ensure high contrast and horizontal continuity of pixels in each record.

Images