Deep learning model for picking up seismic phase from seismic signal with low signal-to-noise ratio

Abstract

The invention discloses a deep learning model for picking up a seismic phase from a seismic signal with a low signal-to-noise ratio. The deep learning model comprises a feature extraction trunk, a bidirectional time sequence-channel attention module (BTCA) and a multi-scale dilated convolutional layer module (DSCN). According to the feature extraction trunk, a cascaded LiteMobileBlock module is used, shallow high-resolution features are extracted from an original waveform step by step, and an initial feature map is generated; the bidirectional time sequence-channel attention module integrates the time sequence modeling capability of the bidirectional LSTM and the spectrum sensing capability (emphasizing importance of different sensors / directions) of a channel attention mechanism, and outputs fusion features; and the multi-scale cavity convolution layer module utilizes parallel cavity convolution modules with different expansion coefficients to synchronously capture a local mutation and global oscillation mode of a waveform to generate multi-scale enhancement features, further enhance the capture capability of the model on long-range time dependence in seismic signals through LSTM, and output time sequence enhancement features. According to the model, a multi-scale cavity convolution module, a bidirectional time-frequency attention mechanism module and an LSTM enhanced sequence modeling module are fused, so that the feature extraction and time sequence modeling capability of a seismic signal with a low signal-to-noise ratio is improved, and robust pickup of a seismic phase is realized.

Description

Technical Field

[0001] The present invention relates to the field of seismic signal processing technology, and in particular to a deep learning model for picking up seismic phases from low signal-to-noise ratio seismic signals, which is suitable for fields such as seismic exploration and geological disaster monitoring. Background Art

[0002] Seismic phase picking—the identification of the arrival times of P and S waves in earthquake records—is a fundamental task in seismological observation and research. Accurate phase picking is crucial for earthquake detection, epicenter location, earthquake onset time and magnitude determination, and the detection of underground velocity structures. Specifically, the arrival times of P and S waves are the primary input parameters for earthquake source location algorithms, directly affecting the calculation of the location and depth of the earthquake source. Furthermore, phase picking data is used in focal mechanism analysis, such as using the polarity of the first motion to determine the fault geometry of an earthquake. Rapidly picking the earliest arriving P waves is crucial in earthquake early warning systems, as rapid P-wave detection allows for rapid estimation of the epicenter location and magnitude, enabling early warnings to be issued before the arrival of more destructive S and surface waves. Therefore, reliable P- and S-wave picking is crucial for timely warnings, accurate earthquake location, earthquake hazard assessment, and subsequent seismological analysis.

[0003] At present, the application of deep learning models in seismic phase picking mainly includes PhaseNet, EQTransformer, U-Net, EPick, etc. The main defects of existing deep learning models are: (1) The picking is unstable under low signal-to-noise ratio conditions. Seismic stations located near cities or in severe weather conditions often record waveform data containing a large amount of transient interference and low signal-to-noise ratio, which will cause errors in deep learning models; (2) The generalization ability of the model is poor. The model trained on a specific area or data set has differences in frequency distribution and amplitude of seismic waveforms in different regions due to differences in local geological conditions, source distance and other factors. This may cause the model to mispick data outside the training distribution; (3) The model is not lightweight enough and difficult to implement real-time or edge deployment. Summary of the Invention

[0004] The purpose of the present invention is to provide a lightweight deep learning model for robustly picking up seismic phases from low signal-to-noise ratio seismic signals.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A deep learning model for extracting seismic phases from low signal-to-noise ratio seismic signals, including a feature extraction backbone, a bidirectional temporal-channel attention module, and a multi-scale dilated convolutional layer module, where:

[0007] The feature extraction backbone uses cascaded “LiteMobileBlock” modules to gradually extract shallow high-resolution features from the original waveform to generate the initial feature map ;

[0008] The bidirectional time-channel attention module, including the time-series modeling module and the channel-aware attention module, integrates the time-series modeling capability of the bidirectional LSTM and the spectrum perception capability of the channel attention mechanism (focusing on the importance of different sensors / directions), and outputs fusion features. ;

[0009] The multi-scale dilated convolution layer module uses parallel dilated convolution modules with different expansion coefficients to synchronously capture the local mutation and global oscillation mode of the waveform to generate multi-scale enhanced features. LSTM is used to further enhance the model's ability to capture long-term temporal dependencies in seismic signals and output time series enhancement features. .

[0010] Preferably, the feature extraction backbone adopts a lightweight module "LiteMobileBlock" based on depth-separable convolution. The backbone network consists of three levels of "LiteMobileBlock", and features at each level are transmitted to the decoder through jump connections.

[0011] Preferably, the time series modeling module converts the feature Input the bidirectional LSTM network, capture the waveform changes before and after the initial arrival phase through forward and backward propagation, and then obtain the global temporal context features The hidden state of the forward LSTM at time t is , the hidden state of the backward LSTM at time t is Finally, the bidirectional features are concatenated to obtain .

[0012] Preferably, the channel-aware attention module generates channel weights through global average pooling in the channel dimension and a fully connected layer. , calculated by the formula: , for the input feature map , global average pooling takes the mean of each channel in the time dimension: , and finally the channel weight With LSTM output Add and multiply the original features channel by channel (Hadamard product) to generate the final attention features: .

[0013] Preferably, the multi-scale dilated convolutional layer module includes a multi-scale dilated convolutional layer, and the calculation formula is: , the loss function is defined as: .

[0014] The beneficial effect of the present invention is that the model integrates multi-scale dilated convolution, bidirectional time-frequency attention mechanism and LSTM-enhanced sequence modeling module to improve the feature extraction and time series modeling capabilities of low signal-to-noise ratio seismic signals and achieve robust picking of seismic phases. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Schematic diagram of the deep learning model structure for picking up seismic phase from low signal-to-noise ratio seismic signals of the present invention. DETAILED DESCRIPTION

[0016] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only 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 making creative efforts are within the scope of protection of the present invention.

[0017] See also Figure 1 , a deep learning model for picking up seismic phases from low signal-to-noise ratio seismic signals, including an architecture of a deep learning model for picking up seismic phases from low signal-to-noise ratio seismic signals.

[0018] The deep learning model for extracting seismic phases from low signal-to-noise ratio seismic signals is divided into four core components: Feature Extraction Backbone, Bidirectional Temporal-Channel Attention (BTCA), Multi-Scale Dilated Convolution, and Output Head (such as Figure 1 shown).

[0019] Given a three-channel (NEZ) seismic signal of length T, in order to provide rich feature representation for the subsequent deep network (backbone), the input seismic signal is first subjected to preliminary feature extraction and transformation through a one-dimensional convolutional layer. The network model then maps the original waveform to a sequence of P-wave and S-wave first-arrival time probabilities through the following steps:

[0020] Feature extraction backbone: Use cascaded “LiteMobileBlock” modules to gradually extract shallow high-resolution features from the original waveform to generate the initial feature map.

[0021] Context Enhancement: The "BTCA" module integrates the temporal modeling capabilities of bidirectional LSTM and the spectrum perception capabilities of the channel-aware attention mechanism (focusing on the importance of different sensors / directions) to output fused features.

[0022] Multi-scale perception: Parallel dilated convolution modules with different dilation coefficients are used to synchronously capture local mutations and global oscillation patterns of the waveform, generating multi-scale enhanced features.

[0023] Time series modeling: LSTM is used to further enhance the model's ability to capture long-range time dependencies in seismic signals and output time series enhancement features.

[0024] Probability output: Finally, a lightweight convolutional output head is used to map high-level features into first-arrival time probability distributions corresponding to noise, P waves, and S waves respectively.

[0025] For the lightweight feature extraction backbone, a lightweight module “LiteMobileBlock” based on depthwise separable convolution is adopted, such as Figure 1 As shown in (a), the structure of the depth-wise separable convolution module is as follows Figure 1 (b) (including depthwise convolution and pointwise convolution

[26] ). The backbone network consists of three levels of "LiteMobileBlock", with the number of output channels at each level being [64, 128, 256], always maintaining the input temporal resolution. Moreover, features at each level are passed to the decoder through skip connections to prevent the deep network from ignoring local details due to the large receptive field. Assuming a given input, the operation process is:

[0026]

[0027]

[0028]

[0029] The depthwise separable convolution uses a step size of s=1 to ensure feature size is not compressed; batch normalization (BN) improves network stability; and shortcuts address the vanishing gradient problem in deep networks and enhance information flow. Compared to traditional residual blocks, the "LiteMobileBlock" reduces the number of parameters by 72%, while effectively preserving high-frequency waveform details through skip connections.

[0030] For the bidirectional temporal-channel joint attention module (BTCA), the "BTCA" module can achieve robust feature extraction in high-noise environments by combining temporal modeling and channel perception. The structure is as follows Figure 1 (c) shown.

[0031] Time Series Modeling: Features (C is the number of channels, T is the time step) are fed into a bidirectional LSTM network. The network then uses forward and backward propagation to capture waveform changes before and after the first arrival phase, thereby obtaining global temporal context features. The hidden state of the forward LSTM at time t is , encoding the cumulative waveform features from the signal start to time t, capturing background noise before the first arrival phase. The hidden state of the backward LSTM at time t is , encoding future seismic phase information from the signal end to time t, assisting in identifying P-wave / S-wave oscillation characteristics after the first arrival phase. Finally, the bidirectional features are concatenated to obtain the final result.

[0032] Channel-aware attention module: Generates channel weights through global average pooling in the channel dimension and a fully connected layer, calculated using the following formula:

[0033]

[0034] in is the Sigmoid activation function, is the ReLU activation function, and GAP represents global average pooling. This design reduces the P-wave false detection rate by 12.4% and the S-wave false detection rate by 5.4% at a threshold of 0.5s.

[0035] For the input feature map, global average pooling takes the mean of each channel in the time dimension:

[0036]

[0037] Feature fusion: Finally, the channel weights are added to the LSTM output and multiplied channel-by-channel (Hadamard product) with the original features to generate the final attention features: , where represents channel-by-channel multiplication.

[0038] For the multi-scale dilated convolutional layer module, during the propagation of seismic waves, the first-arrival pulse (high frequency) and the subsequent seismic phase (low frequency) have significant scale differences. Therefore, in order to synchronously capture the local mutation of the first-arrival pulse and the global oscillation waveform characteristics of the seismic wave, a multi-scale dilated convolutional layer is designed. The calculation formula is as follows:

[0039]

[0040] in, Indicates channel dimension splicing, expansion coefficient Used to obtain different receptive field ranges (local to global).

[0041] E. Loss Function Design

[0042] In view of the sparsity and time-dependent characteristics of the P-wave / S-wave phase arrival times in seismic signals, this paper uses a hybrid loss function based on weighted negative log-likelihood. Its core goal is to enhance the model's sensitivity to weak phase signals through probability distribution alignment while suppressing noise interference. The loss function is defined as:

[0043]

[0044] in, is the true label of the phase category at time t (C is the number of categories, such as background, P wave, S wave), is the normalized probability distribution of the model output, is a numerical stability constant. Seismic phase events have a very low frequency on the time axis (usually <1%). The traditional mean square error (MSE) is easily dominated by background noise, while the negative log-likelihood enhances the gradient response to sparse positive samples (such as P / S wave jumps) through logarithmic compression of probability. And by forcing the model output to satisfy The loss function implicitly constrains the mutual exclusivity of phase predictions (after softmax processing). Compared with unnormalized regression methods, this design reduces the sensitivity to threshold settings.

[0045] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0046] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A deep learning model for extracting seismic phases from low signal-to-noise ratio seismic signals, comprising a feature extraction backbone, a bidirectional temporal-channel attention module, and a multi-scale dilated convolutional layer module; characterized by: The feature extraction backbone uses cascaded "LiteMobileBlock" modules to gradually extract shallow high-resolution features from the original waveform to generate the initial feature map ; The bidirectional time series-channel attention module integrates the time series modeling capability of the bidirectional LSTM and the spectrum perception capability of the channel attention mechanism, and outputs fusion features. ; The multi-scale dilated convolution layer module uses parallel dilated convolution modules with different expansion coefficients to synchronously capture the local mutation and global oscillation mode of the waveform to generate multi-scale enhanced features. LSTM is used to further enhance the model's ability to capture long-term temporal dependencies in seismic signals and output time series enhancement features. .

2. A deep learning model for extracting seismic phases from low signal-to-noise ratio seismic signals according to claim 1, characterized in that: The feature extraction backbone adopts a lightweight module "LiteMobileBlock" based on depthwise separable convolution. The backbone network consists of three levels of "LiteMobileBlock", which transmit features at each level to the decoder through skip connections.

3. The deep learning model for extracting seismic phases from low signal-to-noise ratio seismic signals according to claim 1, characterized in that: The bidirectional temporal-channel attention module includes a temporal modeling module and a channel-aware attention module.

4. The deep learning model for extracting seismic phases from low signal-to-noise ratio seismic signals according to claim 3, characterized in that: The time series modeling module is: Input the bidirectional LSTM network, capture the waveform changes before and after the initial arrival phase through forward and backward propagation, and then obtain the global temporal context features ; The hidden state of the forward LSTM at time t is , the hidden state of the backward LSTM at time t is Finally, the bidirectional features are concatenated to obtain ; The channel-aware attention module generates channel weights through global average pooling in the channel dimension and the fully connected layer , calculated by the formula: , for the input feature map , global average pooling takes the mean of each channel in the time dimension: , and finally the channel weight With LSTM output Add and multiply the original features channel by channel to generate the final attention feature .

5. The deep learning model for extracting seismic phase from low signal-to-noise ratio seismic signals according to claim 1, characterized in that: The multi-scale dilated convolutional layer module includes a multi-scale dilated convolutional layer, and the calculation formula is: , the loss function is defined as: .

Images