Deep Learning-Based Method for Enhancing Natural Earthquake Signals

Abstract

The present invention relates to the field of seismic data processing, and provides a method for enhancing natural earthquake signals based on deep learning, including: constructing a data set; constructing a noise suppression module, the noise suppression module adopts a multi-channel neural network model, and this multi-channel neural network model includes three sets of UNet sub-structures, namely the time-domain UNet, the time-frequency domain UNet and the frequency-domain UNet, so as to simultaneously use the three transform domain features of the time domain, the time-frequency domain and the frequency domain; using the training set to train the multi-channel neural network model; analyzing and processing the actual monitoring data, including actual monitoring data preprocessing and actual monitoring data analysis; judging whether to start the update of the data set and the update of the noise suppression module. The method for enhancing natural earthquake signals based on deep learning of the present invention can identify and suppress the noise of low signal-to-noise ratio earthquake events in seismic monitoring data, enhance the seismic signal, improve the signal-to-noise ratio, and further improve the accuracy of subsequent seismic data processing.

Description

Technical Field

[0001] The present invention relates to the field of seismic data processing, and particularly to a method for enhancing natural seismic signals based on deep learning. It belongs to the field of "processing of seismic data" in "seismology; exploration or detection of seismic or acoustic waves", especially "correlative processing of seismic signals; elimination of the influence of interference energy". Background Art

[0002] Seismic data is inevitably interfered by various noises, and the noise interference will seriously reduce the quality of seismic data, thereby affecting the accuracy of subsequent seismic data processing. In order to suppress the noise interference in natural seismic data and enhance the seismic signals at the same time, various methods have been studied and are currently mainly divided into two categories: conventional methods and emerging deep learning methods.

[0003] The basic idea of the first category of conventional methods is to transform the time-domain waveform into other transform domains to improve the sparsity of seismic data, so as to more easily distinguish seismic signals and noises and separate or suppress them. Such methods mainly include:

[0004] 1. Frequency-domain noise suppression methods. Frequency-domain noise suppression methods include spectral filtering method and spectral subtraction method.

[0005] The spectral filtering method usually uses the Fourier transform to transform the time-domain waveform data into the frequency domain, extracts the spectrum within the frequency range where the seismic event is located, and then suppresses the noise in the seismic data. However, when the frequency ranges of the noise and the seismic event overlap, the noise cannot be suppressed by this method. At the same time, the frequency ranges of different seismic events vary greatly, and corresponding frequency ranges need to be set for each seismic event, and the processing process is relatively cumbersome, making it difficult to achieve automated data processing.

[0006] The spectral subtraction method extracts the seismic event waveform and the noise before the first arrival of the P wave of the seismic event, performs Fourier transform on both of them respectively to obtain the corresponding spectra, squares the spectra to obtain the corresponding energy spectra, and then subtracts the event energy spectrum and the noise energy spectrum to obtain the energy spectrum of the seismic event with noise suppression. However, due to the non-stationarity of the noise, there is a large error in noise suppression by the spectral subtraction method.

[0007] 2. Time-frequency domain noise suppression method. The time-frequency domain noise suppression method uses time-frequency analysis methods to transform the time-domain waveform data into the time-frequency spectrum. Common time-frequency analysis methods include wavelet transform, short-time Fourier transform (STFT), S transform, etc. The time-frequency spectrum within the range where the seismic event is located is extracted in the time-frequency domain, and then the noise in the seismic data is suppressed. The time-frequency domain method increases the sparsity of the data compared with the frequency domain method, and improves the ability to distinguish seismic events from noise. However, it is still necessary to set the time-frequency range for each seismic event, and the processing process is cumbersome, making it difficult to achieve automated data processing.

[0008] The above-mentioned conventional methods face a common difficulty: in the frequency spectrum or time-frequency spectrum, when the ranges where seismic events and noise are located overlap, it is quite challenging to appropriately select the frequency or time-frequency range of seismic events and use an appropriate threshold function to separate seismic events and noise.

[0009] The second category is deep learning methods. As a powerful machine learning technology, deep learning can be used as a tool for mining the characteristics of seismic data. At present, deep learning technology has made great progress in many fields of geoscience. Summary of the Invention

[0010] The purpose of the present invention is to at least partially overcome the defects of the prior art and provide a method for enhancing natural seismic signals based on deep learning.

[0011] The purpose of the present invention is also to provide a method for enhancing natural seismic signals based on deep learning, which can identify and suppress the noise of low signal-to-noise ratio seismic events in seismic monitoring data and improve the signal-to-noise ratio.

[0012] The purpose of the present invention is also to provide a method for enhancing natural seismic signals based on deep learning, which can reduce the uncertainty of separating seismic signals and noise, so as to improve the effects of seismic signal enhancement and noise suppression.

[0013] To achieve the above purpose or one of the purposes, the technical solution of the present invention is as follows:

[0014] A method for enhancing natural seismic signals based on deep learning, the natural seismic signal enhancement method comprising:

[0015] Step A: Construct a data set, the data set including seismic event data of different data types, coherent noise data, and background random noise data;

[0016] Step B: construct a noise suppression module, the noise suppression module adopts a multi-channel neural network model, the multi-channel neural network model includes three sets of UNet substructures, namely time domain UNet, time-frequency domain UNet and frequency domain UNet, so as to use three transform domain features of time domain, time-frequency domain and frequency domain at the same time; the time domain UNet includes an encoder and a decoder, the encoder part is used for feature extraction, and is composed of a convolution layer, an activation function and a pooling layer, the decoder part is used for feature fusion, and is composed of a deconvolution layer, an activation function and an upsampling layer, the convolution operation is performed on a one-dimensional time dimension, has 64 convolution kernels, and the convolution kernel size is 1×3; the time-frequency domain UNet has the same structure as the time domain UNet, the convolution operation is performed on a two-dimensional time-frequency dimension, and the convolution kernel size is 3×3; the frequency domain UNet has the same structure as the time domain UNet, the convolution operation is performed on a one-dimensional frequency dimension, and the convolution kernel size is 3×1;

[0017] Step C: Use the training set to train the multi-channel neural network model;

[0018] Step D: Analyze and process the actual monitoring data, including preprocessing of the actual monitoring data and analysis of the actual monitoring data; wherein the analysis of the actual monitoring data includes: inputting the preprocessed actual monitoring data into the trained multi-channel neural network model, and calling the network model parameters for calculation, obtaining the time domain waveform data after noise suppression, and calculating the signal-to-noise ratio, focal mechanism and moment magnitude of the earthquake event;

[0019] Step E: Determine whether to start updating the data set and the noise suppression module.

[0020] According to a preferred embodiment of the present invention, the data set has the following data size: the ratio of earthquake event data, consistency noise data and background random noise data of different data types in the data set is 5:3:2, and the total amount of data is Nk; the sizes of all types of data are consistent, and all contain three components, each component contains Ns sampling points, and the total size of the data set is 3*Ns*Nk.

[0021] According to a preferred embodiment of the present invention, the seismic event data is obtained through two methods: artificial simulation and actual monitoring.

[0022] According to a preferred embodiment of the present invention, the consistent noise data is taken from the actual monitored seismic data. The time spectrum of the consistent noise data contains non-random, clustered energy. The consistent noise data is selected as follows:

[0023] The actual monitored earthquake data are collected and sorted, and the seismic phase components of the earthquake events in the actual monitored earthquake data are eliminated; the time-frequency spectrum is obtained from the remaining earthquake data using the time-frequency analysis method, and the part containing non-random, clustered energy distribution is selected, and the data segment corresponding to this part is selected as the consistent noise data.

[0024] According to a preferred embodiment of the present invention, the background random noise data is selected from the actual monitored seismic data, and the background random noise data is selected in the following manner:

[0025] Collect and organize the actual monitored earthquake data, and remove the seismic phase components of the earthquake events in the actual monitored earthquake data; use the time-frequency analysis method to obtain the time-frequency spectrum of the remaining earthquake data, and remove the parts containing non-random, clustered energy distribution; in the remaining earthquake data, pick out the continuous part without clustered energy distribution, and select the data segment corresponding to this part as the background random noise data.

[0026] According to a preferred embodiment of the present invention, the step of constructing a data set also includes data augmentation and data preprocessing, wherein the data augmentation includes data amplitude scaling and / or data superposition, and the data preprocessing includes removing instrument response, removing mean, and removing linear trend.

[0027] According to a preferred embodiment of the present invention, the time domain and the time-frequency domain are connected by a cross-attention mechanism. At the same network depth, the time dimension features extracted by the time-domain convolution kernel and the time dimension features of the time-frequency spectrum are mutually constrained; the frequency domain and the time-frequency domain are connected by a cross-attention mechanism. At the same network depth, the frequency dimension features extracted by the frequency domain convolution kernel and the frequency dimension features of the time-frequency spectrum are mutually constrained.

[0028] According to a preferred embodiment of the present invention, the method of using the data set to train the multi-channel neural network model comprises:

[0029] Divide the data set into training set and test set according to the predetermined ratio;

[0030] Set the network model batch size to 400, the total number of training iterations to 1000, the learning rate to 0.0001, and the learning rate to half after every 100 iterations;

[0031] Multi-channel neural network model training on GPU image processing unit.

[0032] According to a preferred embodiment of the present invention, the actual monitoring data preprocessing includes removing instrument response, removing mean, removing linear trend, and picking effective seismic events.

[0033] According to a preferred embodiment of the present invention, the determining whether to start updating the data set and the noise suppression module comprises:

[0034] It is determined whether the earthquake event contains new features according to the earthquake event parameters obtained in the analysis of the actual monitoring data. If the current earthquake event contains new features, the earthquake event is updated to the data set.

[0035] The natural earthquake signal enhancement method based on deep learning of the present invention has the following beneficial effects:

[0036] 1) Suppress the noise in seismic data in real time and efficiently, improve the signal-to-noise ratio, obtain more accurate seismic event waveform information, and further improve the accuracy of subsequent tasks such as seismic source location, seismic source mechanism inversion, and seismic wave velocity inversion.

[0037] 2) Analyze and judge seismic events with new features in real-time monitoring earthquakes in multiple dimensions, supplement these seismic events to the dataset and update the network model parameters, thereby improving the adaptability and accuracy of the network model for noise suppression in real time. Description of the Drawings

[0038] Figure 1 Shows an example of the data type according to an embodiment of the present invention, where (a) is the actually monitored noise, including random noise and coherent noise, (b) is the actually monitored seismic event with a high signal-to-noise ratio; (c) is the low signal-to-noise ratio seismic event synthesized by superimposing (a) and (b);

[0039] Figure 2 Shows the three parameters of the fault plane according to an embodiment of the present invention: dip angle dip (δ), rake angle rake (λ), and strike angle strike (φ);

[0040] Figure 3 Shows the data superposition in data augmentation, which is used to simulate the coincidence of seismic events in actual monitoring. Among them, (a) and (b) are the waveform data of two seismic events respectively, and (c) is the superposition of the waveform data of the two seismic events;

[0041] Figure 4 Is a schematic structural diagram of the multi-channel neural network model in the noise suppression module. Among them, the network model includes three sets of UNet sub-structures (shaded areas), which are the time-domain UNet, the time-frequency domain UNet, and the frequency-domain UNet from top to bottom. It can simultaneously use the three transform domain features of the time domain, the time-frequency domain, and the frequency domain. The weights are connected between different domains through the cross-attention mechanism, and the time-domain waveform and spectrum respectively constrain the time-frequency spectrum;

[0042] Figure 5 Is the data processing module and process of the natural earthquake signal enhancement method based on deep learning according to an embodiment of the present invention. Among them, the data processing module includes two parts: the seismic event noise suppression module and the transfer training module for updating the noise suppression module. The dataset is updated by the trigger threshold of the new feature seismic event, and the model update process is started by the dataset trigger threshold, and the adaptive noise suppression module is updated. Detailed Embodiments

[0043] The exemplary embodiments of the present invention will be described in detail below with reference to the accompanying drawings, where like or similar reference numerals denote like or similar elements. Additionally, in the following detailed description, for the sake of explanation, numerous specific details are set forth to provide a comprehensive understanding of the embodiments of this disclosure. However, it is obvious that one or more embodiments can also be implemented without these specific details. In other cases, well-known structures and devices are illustrated in a schematic manner to simplify the drawings.

[0044] Deep learning technology has been applied in the field of geoscience. The inventor has extended this technology to seismic event noise suppression. By using a neural network model to learn an extremely complex mapping function, the data sparsity is improved, thereby enhancing the ability to suppress seismic data noise. During the research process, the inventor also found that if only the time-frequency domain features of seismic data are used, noise suppression cannot be well achieved. For example, first use the short-time Fourier transform to obtain the time-frequency spectrum of seismic data, and then use a deep network to learn the signal (seismic event) and noise features of the time-frequency spectrum, and construct a non-linear mapping function for separating seismic events and noise. Such a method only uses the time-frequency domain features of seismic data and fails to fully utilize the features and constraints of seismic data in other transform domains.

[0045] The present invention proposes a method for suppressing seismic data noise using a multi-channel neural network model that simultaneously uses the features of three transform domains: time domain, frequency domain, and time-frequency domain. Based on the different sparsities and dimensions of noise and signal features in the three transform domains, the present invention designs a deep neural network to learn and constrain these features, reduce the uncertainty of separating seismic signals and noise, and enhance the effects of noise suppression and seismic signal enhancement.

[0046] The following uses a specific embodiment to illustrate the method for enhancing natural seismic signals based on deep learning of the present invention.

[0047] 1. Construct a dataset

[0048] 1.1 Dataset overview

[0049] The dataset altogether contains three types of data: seismic events (containing clear seismic phases, such as P and / or S phases), coherent noise (such as linear interference), and background random noise, as Figure 1 shown.

[0050] The data size in the dataset: The ratio of the three types of data in the dataset is 5:3:2, and the total amount of data is Nk (100,000); the sizes of all the above types of data are the same, and each contains three components, and each component contains Ns (1,000) sampling points. Therefore, the total size of the dataset is 3 * Ns * Nk sampling points.

[0051] 1.2 Obtaining the three types of data in the dataset

[0052] 1.2.1 Earthquake event acquisition

[0053] Earthquake events are obtained through two methods: artificial simulation and actual monitoring, which are described in detail below.

[0054] (a) Artificially simulated earthquake events: Complete simulation parameters are used during simulation, including epicentral distance, magnitude, focal mechanism, etc., to synthesize earthquake events with good generalization; the artificially simulated earthquake events do not contain noise.

[0055] The operation process and parameter settings for artificially simulated earthquake events are as follows:

[0056] Given the seismic moment and the three fault plane parameters: dip ( ), rake ( ), and strike ( ), as Figure 2 shown. Convert the above parameters into the focal mechanism , ,

[0057] ,

[0058] ,

[0059] ,

[0060] ,

[0061] ,

[0062] ,

[0063] Based on the focal mechanism , obtain the artificially simulated earthquake event :

[0064] ,

[0065] where is the control matrix, an n×6 matrix, where n is the number of observation data, and it contains the information of the epicentral distance .

[0066] In the above artificially simulated earthquake events, the epicentral distance is set to km; the value ranges of the three fault plane parameters are: , , , ( (the maximum magnitude in the history of this region).

[0067] (b) High signal-to-noise ratio seismic events in actual monitoring: Select seismic events with a signal-to-noise ratio (SNR) greater than 10 dB in actual monitoring. The epicentral distance is selected as kilometers. The source parameters (magnitude, focal mechanism, etc.) of actual seismic events depend on the type of seismogenic fault in this region, and the generalization ability is weak; the high signal-to-noise ratio seismic events in actual monitoring contain weak noise. The focal mechanism and moment magnitude of actual monitoring seismic events need to be calculated through P and / or S phase information. These parameters are used to judge and update the dataset, which is described in detail as follows:

[0068] (1) P and S phase identification:

[0069] Use methods such as phase association to identify and confirm the P-wave and S-wave phases of seismic events. These identified phases will be used for the calculation of source parameters (source location, magnitude, focal mechanism, etc.).

[0070] (2) Focal mechanism calculation:

[0071] Given the monitoring data of a seismic event, the focal mechanism of this seismic event can be estimated. Among them, the monitoring seismic data can use P phase, S phase, P and S phases, etc.; decompose the obtained focal mechanism into principal axis components, and convert the principal axis components into three fault plane parameters: dip angle dip ( ), rake angle rake ( ), and strike angle strike ( ).

[0072] (3) Moment magnitude calculation:

[0073] Calculate the seismic moment based on the focal mechanism obtained in (2). Calculate the moment magnitude through the calculated seismic moment .

[0074] 1.2.2 Acquisition of consistent noise data

[0075] The consistent noise is taken from actual monitoring seismic data. The characteristic of consistent noise is that its time-frequency spectrum contains non-random and clustered energy. The selection method is as follows:

[0076] (1) Collect and organize the actually monitored seismic data, and eliminate the seismic phase components of seismic events in the seismic data;

[0077] (2) Use the time-frequency analysis method for the remaining seismic data to obtain the time-frequency spectrum, pick out the part with non-random and clustered energy distribution, and select the corresponding data segment of this part as the coherent noise. In continuously recorded seismic data, the number of seismic events is limited, and most of them are background noise. Therefore, there are more available coherent noise data, and the generalization is better.

[0078] 1.2.3 Acquisition of background random noise data

[0079] The background random noise is selected from the actually monitored seismic data. The characteristic of the background random noise is that there is no obvious clustered energy distribution in its time-frequency spectrum. The selection method is as follows:

[0080] (1) Collect and organize the actually monitored seismic data, and eliminate the seismic phase components of seismic events in the seismic data;

[0081] (2) Use the time-frequency analysis method for the remaining seismic data to obtain the time-frequency spectrum, and eliminate the part with non-random and clustered energy distribution;

[0082] (3) In the above remaining seismic data, pick out the continuous part without clustered energy distribution, and select the corresponding data segment of this part as the background random noise.

[0083] In continuously recorded seismic data, the number of seismic events is limited, and most of them are background noise. Therefore, there are more available background random noise data, and the generalization is better.

[0084] 1.3 Data augmentation of the dataset

[0085] Perform data augmentation on the dataset generated and selected in 1.2 to improve the universality of the dataset. Data augmentation includes the following two methods:

[0086] (a) Data amplitude scaling, and the operation method is as follows:

[0087] (1) For the data in the existing dataset, set a random scaling factor (the range of the scaling factor is set to 0.1 - 10), and multiply the amplitude of the data by the corresponding scaling factor;

[0088] (2) Supplement and update the scaled data to the dataset.

[0089] (b) Data superposition, the purpose of which is to simulate the coincidence phenomenon of multiple seismic events in the actually monitored seismic data, such as Figure 3 , and the method is as follows:

[0090] (1)In the dataset, pick any number of data (the number of data can be set from 2 to 5, and different types of data are acceptable).

[0091] (2)Set a random time delay for each data (the range of random time delay is set from -Ns / 2 to Ns / 2, where Ns is the number of sampling points), and update the time difference of the data.

[0092] (3)Superimpose the multiple data after time difference update.

[0093] (4)Supplement the superimposed data into the dataset.

[0094] 1.4 Data preprocessing in the dataset

[0095] The steps for data preprocessing in the dataset include:

[0096] (1)Remove instrument response;

[0097] (2)Remove the mean;

[0098] (3)Remove the linear trend.

[0099] 2 Construct a noise suppression module:

[0100] 2.1 Multi-channel neural network model architecture, as Figure 4 shown

[0101] This multi-channel neural network model includes three sets of UNet substructures ( Figure 4 shaded areas), namely the time-domain UNet, the time-frequency domain UNet, and the frequency-domain UNet, to simultaneously use the three transform domain features of the time domain, the frequency domain, and the time-frequency domain.

[0102] The time-domain UNet consists of two parts: an encoder and a decoder. The encoder part is used for feature extraction and is composed of a convolutional layer (64 convolutional kernels, with a kernel size of 1×3), an activation function (rectified linear unit, RELU), and a pooling layer (Maxpooling). The decoder part is used for feature fusion and is composed of a transposed convolutional layer, an activation function, and an upsampling layer. The convolutional operation is performed in the 1D time dimension.

[0103] The time-frequency domain UNet has the same structure as the time-domain UNet, and the convolutional operation is performed in the 2D time-frequency dimension with a kernel size of 3×3.

[0104] The frequency-domain UNet has the same structure as the time-domain UNet, and the convolutional operation is performed in the 1D frequency dimension with a kernel size of 3×1.

[0105] The time domain and the time-frequency domain are connected through a cross-attention mechanism. That is, at the same network depth, the time-dimensional features extracted by the time-domain convolution kernel and the time-dimensional features of the time-frequency spectrum are mutually constrained; the frequency domain and the time-frequency domain are connected through a cross-attention mechanism. That is, at the same network depth, the frequency-dimensional features extracted by the frequency-domain convolution kernel and the frequency-dimensional features of the time-frequency spectrum are mutually constrained.

[0106] 2.2 Input and Output of the Multi-Channel Neural Network Model

[0107] During training, the seismic events in the dataset are used as noise-free time-domain waveform data, and any noise data (coherent noise or background random noise) is selected and superimposed on the seismic events to form time-domain noisy waveform data. The input of the multi-channel neural network model is the time-domain noisy waveform data ( Figure 4 left side) and the noise-free time-domain waveform data ( Figure 4 right side) data pair. The time-domain UNet directly uses the time-domain waveform data and constructs the loss function MSE1 (MSE: mean square error) using the error of the waveform data pair; the above time-domain waveform data pairs are respectively subjected to Fourier transform to obtain the corresponding noisy spectrum and noise-free spectrum, and this spectrum data pair constitutes the input of the frequency-domain UNet, and the loss function MSE2 is constructed using the error of the spectrum data pair; the above time-domain waveform data are respectively subjected to time-frequency analysis, such as short-time Fourier transform, wavelet transform, etc., to obtain the corresponding noisy time-frequency spectrum and noise-free time-frequency spectrum, and this time-frequency spectrum data pair constitutes the input of the time-frequency-domain UNet, and the loss function MSE3 is constructed using the error of the time-frequency spectrum data pair. The overall loss function MSE = aMSE1 + bMSE2 + cMSE3 is constructed to update the parameters of the multi-channel neural network model, where a, b, and c are weight factors. The setting method of the weight factors is as follows: Take k noise-free time-domain waveform data (k is generally 1% - 5% of the data volume of the dataset), calculate their spectra and time-frequency spectra respectively, and obtain the root mean square mean aa of the waveform amplitudes, the root mean square mean bb of the spectrum amplitudes, and the root mean square mean cc of the time-frequency spectrum amplitudes, then a = 1 / aa, b = 1 / bb, and c = 1 / cc.

[0108] During testing, the input of the multi-channel neural network model is the time-domain noisy waveform data, and the output is the time-domain waveform data after noise suppression.

[0109] 3 Training the Multi-Channel Neural Network Model with the Training Set

[0110] 3.1 Divide the dataset after data augmentation into a training set and a test set according to a certain ratio, and the ratio between the two is generally set to 8:2 or 7:3.

[0111] 3.2 Set the batch size of the multi-channel neural network model to 400; the total number of training iterations is 1000; the learning rate is 0.0001, and the learning rate is halved after every 100 iterations.

[0112] The termination conditions for network model training include:

[0113] (a) The number of training iterations reaches the preset total number of training iterations.

[0114] (b) The training accuracy or the loss value of the network model no longer improves with training iterations.

[0115] The determination criteria need to meet at least one of the following conditions:

[0116] (1) At the iter-th iteration (where iter is the current iteration number), the average accuracy value from iter to iter - 5 is less than 105% of the average accuracy value from iter - 2 to iter - 7;

[0117] (2) The average loss value from iter to iter - 5 is greater than 95% of the average loss value from iter - 2 to iter - 7.

[0118] 3.3 The training of the multi-channel neural network model is carried out on the GPU image processing unit. The multi-GPU training method is: a method combining model parallelism and data parallelism.

[0119] 4 Actual monitoring data analysis and processing ( Figure 5 )

[0120] 4.1 Actual data preprocessing, the steps include:

[0121] (1) Remove instrument response,

[0122] (2) Remove the mean,

[0123] (3) Remove the linear trend,

[0124] (4) Pick up effective seismic events.

[0125] 4.2 Actual data analysis

[0126] Input the preprocessed actual data into the multi-channel neural network model trained in the above 3, and call the corresponding network model parameters for calculation to obtain the time-domain waveform data (seismic events) after noise suppression of the input noisy seismic events. Calculate the signal-to-noise ratio (SNR), focal mechanism (dip, rake, and strike), and moment magnitude ( ) and other parameters of the signal (seismic event).

[0127] 4.3 Trigger and update the dataset for new feature events

[0128] Judge whether the earthquake event contains new features based on the earthquake event parameters obtained in 4.2: If at least one of the following thresholds is met, it is determined that the current earthquake event contains new features; if the current earthquake event contains new features, update the earthquake event to the dataset constructed in 1 above.

[0129] The thresholds include the following four situations:

[0130] (a) The signal-to-noise ratio (SNR) of the current earthquake event is greater than the 75th percentile of the signal-to-noise ratios of the existing actual earthquake events in the dataset; this triggering condition takes into account the signal-to-noise ratio of the actual earthquake events.

[0131] (b) The average cross-correlation coefficient between the waveform of the current earthquake event and the waveforms of the actual earthquake events in the dataset is less than 0.8; this triggering condition considers the similarity between the current earthquake event and the actual earthquake events in the dataset from the waveform perspective.

[0132] (c) The average value of the differences between the three parameters of the focal mechanism solution of the current earthquake event and the three parameters of the focal mechanism solution of any actual earthquake event in the dataset is greater than 10 degrees; this triggering condition considers the similarity between the current earthquake event and the actual earthquake events in the dataset from the focal mechanism perspective.

[0133] (d) The magnitude of the current earthquake event is greater than the 75th percentile of the magnitudes of the actual earthquake events in the dataset; this triggering condition considers the magnitude difference between the current earthquake event and the actual earthquake events in the dataset from the magnitude perspective. The starting point of this judgment situation is that generally large-magnitude events are accompanied by more small-magnitude aftershocks, and these small-magnitude aftershocks have great similarity with the large-magnitude main shock, thereby improving the analysis ability of the network model for these aftershocks.

[0134] 5 Noise suppression module update mechanism

[0135] 5.1 Threshold setting for triggering the update of the multi-channel neural network model by the dataset

[0136] Based on the updated dataset in 4.3, judge whether to initiate the update of the multi-channel neural network model (i.e., 5.2 Adaptive noise suppression module update): If at least one of the following thresholds is met, initiate the update of the network model.

[0137] The thresholds include the following four situations:

[0138] (a) The number of actual earthquake events updated in 4.3 reaches 5% of the number of original actual monitored earthquake events in the dataset; this triggering condition takes into account the factor of data volume.

[0139] (b)Among the updated actual seismic events in 4.3, the number of seismic events with the average cross - correlation coefficient (see 4.3(b) for details) less than 0.6 with the waveforms of the actual seismic events in the original dataset reaches 50%; this triggering condition starts from the perspective of time - domain waveforms and takes into account the new characteristics of the newly added data in the dataset.

[0140] (c)Among the updated actual seismic events in 4.3, the average value of the differences between the three parameters of the focal mechanism and the three parameters of the focal mechanism solution of any actual seismic event in the dataset (see 4.3(c) for details) is greater than 50 degrees; this triggering condition starts from the perspective of the focal mechanism and takes into account the new characteristics of the newly added data in the dataset.

[0141] (d)Among the updated actual seismic events in 4.3, the magnitude (see 4.3(d) for details) is greater than the 95th percentile of the magnitudes of the actual seismic events in the dataset; this triggering condition starts from the perspective of magnitude and takes into account the new characteristics of the newly added data in the dataset.

[0142] 5.2 Adaptive update method for the noise suppression module ( Figure 5 )

[0143] The update method includes the following two situations:

[0144] (1)When the triggering threshold is the situation (b) in 5.1, taking the current multi - channel neural network model parameters as the initial values, using the current dataset to train and update all the parameters of the network model; this process has a good module update effect, but the update speed is slow.

[0145] (2)When the triggering threshold is the situations (a), (c) and (d) in 5.1, freezing the feature extraction layer of the multi - channel neural network model, keeping the parameters of this part unchanged, and at the same time using the current dataset to train and update the other parameters of the network model; this process has a faster module update speed, but the update effect is relatively weak.

[0146] 5.3 Through the adaptive update of the noise suppression module in 5.2, new multi - channel neural network model parameters are obtained, and these network model parameters are used for subsequent data processing.

[0147] The natural seismic signal enhancement method based on deep learning of the present invention has the following beneficial effects:

[0148] 1) Suppress the noise in seismic data in real - time and efficiently, improve the signal - to - noise ratio, obtain more accurate seismic event waveform information, and further improve the accuracy of subsequent tasks such as focal location, focal mechanism solution inversion, and seismic wave velocity inversion.

[0149] 2) Multi-dimensional analysis and judgment are used to real-time monitor seismic events with new features in earthquakes, supplement these seismic events to the data set and update the network model parameters, thereby real-time improving the adaptability and accuracy of the network model for noise suppression.

[0150] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes may be made therein without departing from the principles and spirit of the invention. The scope of application of the present invention is defined by the appended claims and their equivalents.

Claims

1. A natural earthquake signal enhancement method based on deep learning, characterized in that: The natural earthquake signal enhancement method comprises: Step A: constructing a data set, wherein the data set includes earthquake event data of different data types, consistent noise data and background random noise data; Step B: Construct a noise suppression module, which adopts a multi-channel neural network model. The multi-channel neural network model includes three sets of UNet substructures, namely time domain UNet, time-frequency domain UNet and frequency domain UNet, so as to simultaneously use the three transform domain features of time domain, time-frequency domain and frequency domain: the time domain and time-frequency domain are connected by a cross-attention mechanism. At the same network depth, the time dimension features extracted by the time domain convolution kernel are mutually constrained with the time dimension features of the time-frequency spectrum; the frequency domain and time-frequency domain are connected by a cross-attention mechanism. At the same network depth, the frequency dimension features extracted by the frequency domain convolution kernel are mutually constrained with the frequency dimension features of the time-frequency spectrum. The dimensional features are mutually constrained; the time domain UNet consists of two parts, the encoder and the decoder. The encoder part is used for feature extraction and consists of a convolution layer, an activation function and a pooling layer. The decoder part is used for feature fusion and consists of a deconvolution layer, an activation function and an upsampling layer. The convolution operation is performed on a one-dimensional time dimension, with 64 convolution kernels, and the convolution kernel size is 1×3; the time-frequency domain UNet has the same structure as the time domain UNet, and the convolution operation is performed on a two-dimensional time-frequency dimension, with a convolution kernel size of 3×3; the frequency domain UNet has the same structure as the time domain UNet, and the convolution operation is performed on a one-dimensional frequency dimension, with a convolution kernel size of 3×1; Step C: Use the training set to train the multi-channel neural network model; Step D: Analyze and process the actual monitoring data, including preprocessing of the actual monitoring data and analysis of the actual monitoring data; wherein the analysis of the actual monitoring data includes: inputting the preprocessed actual monitoring data into the trained multi-channel neural network model, and calling the network model parameters for calculation, obtaining the time domain waveform data after noise suppression, and calculating the signal-to-noise ratio, focal mechanism and moment magnitude of the earthquake event; Step E: Determine whether to start updating the data set and the noise suppression module.

2. The natural earthquake signal enhancement method based on deep learning according to claim 1, characterized in that: The earthquake event data are obtained through artificial simulation and actual monitoring.

3. The natural earthquake signal enhancement method based on deep learning according to claim 1, characterized in that: The consistent noise data is taken from the actual monitored seismic data. The time-frequency spectrum of the consistent noise data contains non-random, clustered energy. The consistent noise data is selected as follows: The actual monitored earthquake data are collected and sorted, and the seismic phase components of the earthquake events in the actual monitored earthquake data are eliminated; the time-frequency spectrum is obtained from the remaining earthquake data using the time-frequency analysis method, and the part containing non-random, clustered energy distribution is selected, and the data segment corresponding to this part is selected as the consistent noise data.

4. The natural earthquake signal enhancement method based on deep learning according to claim 1, characterized in that: The background random noise data is selected from the actual monitored earthquake data. The background random noise data is selected in the following way: Collect and organize the actual monitored earthquake data, and remove the seismic phase components of the earthquake events in the actual monitored earthquake data; use the time-frequency analysis method to obtain the time-frequency spectrum of the remaining earthquake data, and remove the parts containing non-random, clustered energy distribution; in the remaining earthquake data, pick out the continuous part without clustered energy distribution, and select the data segment corresponding to this part as the background random noise data.

5. The natural earthquake signal enhancement method based on deep learning according to claim 1, characterized in that: The step of constructing a data set also includes data augmentation and data preprocessing, wherein the data augmentation includes data amplitude scaling and / or data superposition, and the data preprocessing includes removing instrument response, removing mean value and removing linear trend.

6. The natural earthquake signal enhancement method based on deep learning according to claim 1, characterized in that: The method of using the training set to train the multi-channel neural network model includes: Divide the data set into training set and test set according to the predetermined ratio; Set the network model batch size to 400, the total number of training iterations to 1000, the learning rate to 0.0001, and the learning rate to half after every 100 iterations; Multi-channel neural network model training on GPU image processing unit.

7. The natural earthquake signal enhancement method based on deep learning according to any one of claims 1 to 6, characterized in that: The actual monitoring data preprocessing includes removing instrument response, removing mean, removing linear trend, and picking effective seismic events.

8. The natural earthquake signal enhancement method based on deep learning according to any one of claims 1 to 6, characterized in that: The determining whether to start updating the data set and the noise suppression module comprises: It is determined whether the earthquake event contains new features according to the earthquake event parameters obtained in the analysis of the actual monitoring data. If the current earthquake event contains new features, the earthquake event is updated to the data set.

Images