Novel artificial intelligence method for multi-dimensional multi-level attention earthquake first arrival pickup

Abstract

The invention relates to the technical field of seismic signal detection, in particular to a multi-dimensional multi-level attention seismic first arrival pickup novel artificial intelligence method, which comprises three steps of model building, model training and first arrival pickup. The core components of the model building model comprise an efficient discrete wavelet device, an encoder, a feature extractor, a decoder and an inverse wavelet converter. According to the method, time-frequency decomposition is performed on the seismic waveform by using efficient discrete wavelet transform, the characteristics of the signal on different frequencies and time dimensions are extracted, and local and global correlation modeling is performed on the extracted characteristics through a hierarchical attention mechanism, so that the complex characteristics of the seismic signal are more effectively captured, and the accuracy of the seismic signal is improved. The seismic wave first arrival pickup method has the remarkable advantages of first arrival pickup precision and stability in a noise complex environment, and the accuracy of seismic wave first arrival pickup can be greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of earthquake signal detection technology, and in particular to a novel artificial intelligence method for picking earthquake first arrivals with multi-dimensional and multi-level attention. Background Art

[0002] Seismic wave first arrival detection plays a vital role in seismology, directly affecting key links such as earthquake location, magnitude assessment, and underground velocity structure inversion. However, in practical applications, since signals are often interfered with by complex background noise, traditional technologies have obvious shortcomings in low signal-to-noise ratio environments, thus restricting the accuracy and reliability of earthquake monitoring.

[0003] Traditional methods such as the long-short time window ratio and the Akaike Information Criterion rely primarily on the time-domain statistical characteristics of the signal, but face challenges in practical applications. When background noise is strong, the short-time window energy of STA / LTA is easily interfered with, resulting in frequent false positives or missed positives for P-wave and S-wave first arrivals. AIC, based on the autoregressive properties of time series, can easily misjudge sudden changes in the noise as first arrivals when encountering high-intensity and non-stationary noise, thereby weakening its accuracy. Furthermore, phase identification methods based on the skewness-to-kurtosis ratio often mistakenly identify noise as S-waves when the signal-to-noise ratio is extremely low, as the randomness of the noise may be similar to the statistical characteristics of the real seismic signal.

[0004] In recent years, deep learning has made rapid progress in earthquake first arrival detection, but the advantages and disadvantages of each model are also very obvious. Taking PhaseNet as an example, it relies on the U-Net architecture to maintain high accuracy in the positioning of P waves and S waves, but it is easy to cause large errors and increase false alarms in low signal-to-noise ratio environments. On the other hand, EQTransformer integrates Transformer with convolutional neural network to achieve integrated processing of earthquake phase picking and event detection. Although this design is suitable for processing large-scale data, its demand for computing resources is relatively strict, resulting in real-time applications being restricted. At the same time, under conditions of low signal-to-noise ratio or long distance from the epicenter, In addition, although the GPD model has a fast training speed, it is difficult to effectively distinguish P waves from S waves in a complex noise environment; PickNet performs well in P-wave recognition, but its recall rate for S waves is still insufficient; PpkNet uses convolution and recurrent networks to enhance the ability to capture temporal features, thereby improving the overall picking accuracy, but the high computational overhead limits its real-time application; FilterPicker attempts to combine filtering technology and deep learning to reduce noise interference, but it still fails to achieve the expected effect in S-wave picking. Coupled with the high debugging cost, its actual promotion faces great challenges.

[0005] To address the above problems, the present invention proposes a new artificial intelligence method for earthquake first arrival picking with multi-dimensional and multi-level attention, which is used to efficiently pick up the first arrival of seismic waves. This new method combines the high resolution capability of discrete wavelet transform with the mechanism of multi-level attention, and realizes the efficient integration of signal time-frequency characteristics and local to global characteristics. The present invention not only makes up for the deficiency of traditional methods in relying on a single feature, but also can achieve higher picking precision and accuracy under low signal-to-noise ratio conditions. Summary of the Invention

[0006] In order to overcome the problem of low accuracy of the existing deep learning network model for earthquake first arrival picking under low signal-to-noise ratio conditions, the present invention proposes a new artificial intelligence method for earthquake first arrival picking with multi-dimensional and multi-level attention.

[0007] The technical solution of the present invention is: a new artificial intelligence method for picking earthquake first arrivals with multi-dimensional and multi-level attention, including:

[0008] Model construction: The core components of the model include an efficient discrete wavelet, an encoder, a feature extractor, a decoder, and an inverse wavelet transformer. The three-component seismic waveform data sequence is input and two sets of data streams are generated for P-wave and S-wave feature extraction respectively. The input signal is decomposed into time and frequency by an efficient discrete wavelet transformer. Daubechies wavelet is selected as the basis function to calculate the low-frequency approximation coefficient X. j (k) and high-frequency detail coefficient W j (k) captures the main structure and mutation characteristics of the signal respectively. The signal after wavelet transform decomposition enters the encoder module to extract preliminary features. The model further mines deep features through the feature extractor and designs a symmetrical decoder module for signal reconstruction. The signal is restored to its original length through the inverse wavelet transformer and the Sigmoid activation function is used to generate independent probability distributions of P waves and S waves.

[0009] Model training: Use the Adam optimization algorithm to update network parameters, then use the Xavier normal distribution to initialize the weight matrix and filter parameters, set the bias parameters to zero, and use an adaptive learning rate adjustment mechanism to dynamically adjust the learning rate based on the performance of the validation set;

[0010] First arrival picking: Extract first arrival information from the P-wave probability sequence and S-wave probability sequence output by the pre-trained deep learning model.

[0011] Preferably, each channel of the three-component seismic waveform data sequence contains 6000 time series sampling points.

[0012] Preferably, the two sets of data streams are subjected to time-frequency analysis through a wavelet transformer to capture seismic phase characteristics at different time resolutions.

[0013] Preferably, the efficient discrete wavelet transformer adopts Daubechies wavelet basis function, and the calculation formula of its scaling function φ(t) and wavelet function ψ(t) is:

[0014]

[0015] Among them, X j (k) represents the low-frequency approximation coefficient of the jth layer, which can retain the main structural information of the signal; W j (k) is the high-frequency detail coefficient, which is used to capture the sudden change characteristics of the seismic phase, and n is the sample point of the time series.

[0016] Preferably, the encoder includes 7 convolution blocks, 5 residual blocks and 3 bidirectional long short-term memory network layers, the feature extractor includes a wavelet transformer, a window merging operation, a sliding window attention mechanism and an inverse wavelet transformer, and the decoder is designed to have a symmetrical structure with the encoder, including 3 bidirectional long short-term memory network layers, 5 residual blocks and 7 convolution blocks. The decoder output is restored to the original signal length through two inverse wavelet transformers, and the Sigmoid activation function is used to map the result to the [0,1] interval to generate independent probability distributions of P waves and S waves respectively.

[0017] Preferably, the Adam optimization algorithm utilizes its adaptive learning rate characteristics to accelerate the gradient descent process and improve training stability when updating the network parameters of the model.

[0018] Preferably, the learning rate adopts an adaptive adjustment mechanism, which changes dynamically according to the performance of the model on the validation set, accelerates convergence in the early stage of training, and fine-tunes in the later stage of training.

[0019] Preferably, a loss function is used in the model training step, and the loss function uses binary cross entropy to calculate the difference between the true label and the predicted output to drive the adjustment and optimization of the model parameters.

[0020] Preferably, in the model training step, to prevent overfitting, the training process is automatically terminated when the loss function value does not decrease in ten consecutive iterations.

[0021] Preferably, in the first arrival picking step, first, the P wave probability threshold is set to 0.3, and the S wave probability threshold is also set to 0.3. When the P wave probability or S wave probability in the event window exceeds the corresponding threshold, the system records the time point corresponding to the probability peak as the candidate first arrival time. In order to verify the accuracy of the picking result, the method combines signal-to-noise ratio analysis, takes the candidate first arrival point as the center, and takes the 100 sampling points before and after as the noise window and signal window respectively. The 95th percentile ratio of the two is calculated and converted into decibel form. Only when the signal-to-noise ratio reaches a certain standard, the first arrival point is confirmed as a valid pick. In addition, the SP time difference constraint mechanism is introduced, and the upper limit is set to 60 seconds. It is required that the difference between the S wave first arrival time and the P wave first arrival time shall not exceed this value to eliminate unreasonable picking results. Through the above-mentioned comprehensive strategy of probability threshold screening, signal-to-noise ratio verification and time constraint, high-precision picking results of P wave and S wave first arrival are finally obtained.

[0022] Beneficial effects of the present invention:

[0023] By utilizing efficient discrete wavelet transform to perform time-frequency decomposition of seismic waveforms, extracting the signal features in different frequency and time dimensions, and then using a hierarchical attention mechanism to perform local and global correlation modeling on the extracted features, the complex characteristics of the seismic signal can be captured more effectively. Finally, the peak of the confidence curve of the first arrival is obtained as the first arrival time. The present invention has significant advantages in the first arrival picking accuracy and stability in a noisy and complex environment, and can greatly improve the accuracy of seismic wave first arrival picking. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 What is shown is a schematic diagram of the model structure of the present invention;

[0025] Figure 2 Shown is a schematic diagram of the deep learning network model structure of the present invention;

[0026] Figure 3 Shown are schematic diagrams of the first arrivals of P waves and S waves picked up by the present invention and manually. DETAILED DESCRIPTION

[0027] 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 the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0028] See also Figure 1The present invention provides an embodiment: a novel artificial intelligence method for picking earthquake first arrivals with multi-dimensional and multi-level attention, comprising:

[0029] Model construction: The core components of the model include an efficient discrete wavelet, an encoder, a feature extractor, a decoder, and an inverse wavelet transformer. The three-component seismic waveform data sequence is input and two sets of data streams are generated for P-wave and S-wave feature extraction respectively. The input signal is decomposed into time and frequency by an efficient discrete wavelet transformer. Daubechies wavelet is selected as the basis function to calculate the low-frequency approximation coefficient X. j (k) and high-frequency detail coefficient W j (k) captures the main structure and mutation characteristics of the signal respectively. The signal after wavelet transform decomposition enters the encoder module to extract preliminary features. The model further mines deep features through the feature extractor and designs a symmetrical decoder module for signal reconstruction. The signal is restored to its original length through the inverse wavelet transformer and the Sigmoid activation function is used to generate independent probability distributions of P waves and S waves.

[0030] Model training: Use the Adam optimization algorithm to update network parameters, then use the Xavier normal distribution to initialize the weight matrix and filter parameters, set the bias parameters to zero, and use an adaptive learning rate adjustment mechanism to dynamically adjust the learning rate based on the performance of the validation set;

[0031] First arrival picking: Extract first arrival information from the P-wave probability sequence and S-wave probability sequence output by the pre-trained deep learning model.

[0032] For further information, see Figure 2 , the model building steps are explained in detail:

[0033] The core components of the new model proposed in this paper include an efficient discrete wavelet filter, an encoder, a feature extractor, a decoder, and an inverse wavelet transformer. The input of the model is a three-component seismic waveform data sequence, with each channel containing 6000 time-series sampling points. In order to process P waves and S waves separately, the input signal is first copied twice to generate two identical data streams, which are used to extract the characteristics of P waves and S waves respectively. Subsequently, these two data streams are subjected to time-frequency analysis by a wavelet transformer to capture the seismic phase characteristics at different time resolutions. The wavelet transformer uses the Daubechies wavelet as the basis function, which can efficiently decompose the local features and mutation information in the seismic waveform, laying the foundation for subsequent processing. The mathematical expression for calculating the scaling function φ(t) and the wavelet function ψ(t) is as follows:

[0034] X j (k)=∑ n X(n)φ j,k (n);

[0035] W j (k)=∑ n X(n)ψ j,k (n);

[0036] Among them, X j (k) represents the low-frequency approximation coefficient of the jth layer, which can retain the main structural information of the signal; W j (k) is the high-frequency detail coefficient, which is used to capture the sudden change characteristics of the seismic phase, and n is the sample point of the time series.

[0037] The signal after wavelet transform decomposition enters the encoder module. The encoder consists of 7 convolution blocks, 5 residual blocks and 3 bidirectional long short-term memory network layers. After the encoder extracts preliminary features, the model further mines deep features through the feature extractor. The feature extractor mainly consists of a wavelet transformer, a window merging operation, a sliding window attention mechanism and an inverse wavelet transformer. After extracting the deep features, the model reconstructs the signal through the decoder. The decoder is designed to be symmetrical with the encoder and contains 3 bidirectional long short-term memory network layers, 5 residual blocks and 7 convolution blocks. Finally, the decoder output is restored to the original signal length through two inverse wavelet transformers, and the Sigmoid activation function is used to map the result to the [0,1] interval to generate independent probability distributions of P waves and S waves respectively.

[0038] Furthermore, the model training steps are described in detail:

[0039] During the model training and optimization stage, the present invention adopts a series of technical means to ensure efficient parameter convergence and improvement of model performance. In the specific implementation, the Adam optimization algorithm is used to update the network parameters of the model, and its adaptive learning rate characteristics are used to accelerate the gradient descent process and improve training stability. The weight matrix and filter parameters of the model are set by the Xavier normal distribution initialization method. All bias parameters are uniformly set to zero to provide balanced initial conditions. The learning rate adopts an adaptive adjustment mechanism, which dynamically changes according to the performance of the model on the validation set, accelerates convergence in the early stage of training, and performs fine optimization in the later stage to avoid falling into local optimality. The loss function uses binary cross entropy to calculate the difference between the true label and the predicted output to drive the adjustment and optimization of the model parameters. To prevent overfitting, when the loss function value does not decrease in ten consecutive iterations, the training process is automatically terminated to ensure that the model completes the training in the best state.

[0040] Furthermore, the first arrival picking steps are described in detail:

[0041] First, the P-wave probability threshold is set to 0.3, and the S-wave probability threshold is also set to 0.3. When the P-wave probability or S-wave probability within the event window exceeds the corresponding threshold, the system records the time point corresponding to the probability peak as the candidate first arrival time. To verify the accuracy of the picking results, the method combines signal-to-noise ratio analysis. With the candidate first arrival point as the center, the 100 sampling points before and after are taken as the noise window and signal window, respectively. The 95th percentile ratio of the two is calculated and converted into decibel form. Only when the signal-to-noise ratio reaches a certain standard is the first arrival point confirmed as a valid picking. In addition, the SP time difference constraint mechanism is introduced, with an upper limit of 60 seconds. The difference between the S-wave first arrival time and the P-wave first arrival time must not exceed this value to eliminate unreasonable picking results. Through the comprehensive strategy of probability threshold screening, signal-to-noise ratio verification and time constraint, high-precision picking results of P-wave and S-wave first arrivals are finally obtained, providing a reliable basis for seismic waveform analysis.

[0042] See also Figure 3 The present invention provides an embodiment: taking the Mw4.0 earthquake event that occurred at 12:11:19 on April 25, 2014 (epicenter location 37.132832°E, 8.826829°S, centroid depth 33.3km) as an example, the actual application process of the present invention under severe noise conditions is elaborated in detail. This embodiment takes the three-component data recorded by the station PFVI as the research object, and each component is composed of 6000 sampling points. In order to effectively deal with the problem of high noise interference, the waveform data is first normalized. By subtracting the mean of each component and dividing by the standard deviation, the data is adjusted to a standard scale of zero mean and unit variance. Subsequently, a 1-45Hz bandpass filter is applied to the normalized waveform to filter out low-frequency and high-frequency noise and highlight the core features of the seismic signal. The preprocessed data is input into a pre-trained deep learning model to generate a P-wave probability sequence and an S-wave probability sequence. The prediction results are as follows: Figure 3 As shown, although the waveform data is significantly affected by complex noise, the present invention can still successfully achieve high-precision picking of P-wave and S-wave first arrivals, with an accuracy consistent with manual annotation. This example fully verifies that the present invention can automatically pick up earthquake first arrivals with high accuracy in high-noise conditions.

[0043] Through the above steps, the seismic waveform is decomposed into time and frequency by utilizing the efficient discrete wavelet transform, the characteristics of the signal in different frequency and time dimensions are extracted, and then the extracted characteristics are locally and globally associated with modeling through the hierarchical attention mechanism, so as to more effectively capture the complex characteristics of the seismic signal. Finally, the peak of the confidence curve of the first arrival is obtained as the first arrival time. The present invention has significant advantages in the accuracy and stability of the first arrival picking in a noisy environment, and can greatly improve the accuracy of the first arrival picking of seismic waves, so as to solve the problem of low accuracy of the existing deep learning network model for seismic first arrival picking in low signal-to-noise ratio conditions.

Claims

1. A new artificial intelligence method for earthquake first arrival detection with multi-dimensional and multi-level attention, characterized by: Includes: Model construction: The core components of the model include an efficient discrete wavelet, an encoder, a feature extractor, a decoder, and an inverse wavelet transformer. The three-component seismic waveform data sequence is input and two sets of data streams are generated for P-wave and S-wave feature extraction respectively. The input signal is decomposed into time and frequency by an efficient discrete wavelet transformer. Daubechies wavelet is selected as the basis function to calculate the low-frequency approximation coefficient X. j (k) and high-frequency detail coefficient W j (k) captures the main structure and mutation characteristics of the signal respectively. The signal after wavelet transform decomposition enters the encoder module to extract preliminary features. The model further mines deep features through the feature extractor and designs a symmetrical decoder module for signal reconstruction. The signal is restored to its original length through the inverse wavelet transformer and the Sigmoid activation function is used to generate independent probability distributions of P waves and S waves. Model training: Use the Adam optimization algorithm to update network parameters, then use the Xavier normal distribution to initialize the weight matrix and filter parameters, set the bias parameters to zero, and use an adaptive learning rate adjustment mechanism to dynamically adjust the learning rate based on the performance of the validation set; First arrival picking: Extract first arrival information from the P-wave probability sequence and S-wave probability sequence output by the pre-trained deep learning model.

2. A novel artificial intelligence method for earthquake first arrival detection with multi-dimensional and multi-level attention according to claim 1, characterized in that: The three-component seismic waveform data sequence includes 6000 time series sampling points in each channel.

3. The novel artificial intelligence method for earthquake first arrival detection with multi-dimensional and multi-level attention according to claim 1 is characterized by: The two sets of data streams will be subjected to time-frequency analysis through a wavelet transformer to capture the seismic phase characteristics at different time resolutions.

4. The novel artificial intelligence method for earthquake first arrival detection with multi-dimensional and multi-level attention according to claim 1 is characterized by: The efficient discrete wavelet transformer adopts Daubechies wavelet basis function, and the calculation formula of its scaling function φ(t) and wavelet function ψ(t) is: Among them, X j (k) represents the low-frequency approximation coefficient of the jth layer, which can retain the main structural information of the signal; W j (k) is the high-frequency detail coefficient, which is used to capture the sudden change characteristics of the seismic phase, and n is the sample point of the time series.

5. The novel artificial intelligence method for earthquake first arrival detection with multi-dimensional and multi-level attention according to claim 1 is characterized by: The encoder includes 7 convolution blocks, 5 residual blocks and 3 bidirectional long short-term memory network layers. The feature extractor includes a wavelet transformer, a window merging operation, a sliding window attention mechanism and an inverse wavelet transformer. The decoder is designed in a symmetrical structure with the encoder, including 3 bidirectional long short-term memory network layers, 5 residual blocks and 7 convolution blocks. The decoder output is restored to the original signal length through two inverse wavelet transformers, and the Sigmoid activation function is used to map the result to the [0,1] interval to generate independent probability distributions of P waves and S waves respectively.

6. The novel artificial intelligence method for earthquake first arrival detection with multi-dimensional and multi-level attention according to claim 1 is characterized by: The Adam optimization algorithm uses its adaptive learning rate characteristics to accelerate the gradient descent process and improve training stability when updating the network parameters of the model.

7. The novel artificial intelligence method for earthquake first arrival detection with multi-dimensional and multi-level attention according to claim 1 is characterized by: The learning rate adopts an adaptive adjustment mechanism, which changes dynamically according to the performance of the model on the validation set, accelerates convergence in the early stage of training, and fine-tunes in the later stage of training.

8. The novel artificial intelligence method for earthquake first arrival detection with multi-dimensional and multi-level attention according to claim 1 is characterized by: In the model training step, a loss function is used, and the loss function uses binary cross entropy to calculate the difference between the true label and the predicted output to drive the adjustment and optimization of the model parameters.

9. The novel artificial intelligence method for earthquake first arrival detection with multi-dimensional and multi-level attention according to claim 1 is characterized by: In the model training step, to prevent overfitting, the training process is automatically terminated when the loss function value does not decrease in ten consecutive iterations.

10. The novel artificial intelligence method for earthquake first arrival detection with multi-dimensional and multi-level attention according to claim 1 is characterized by: In the first arrival picking step, first, the P wave probability threshold is set to 0.3, and the S wave probability threshold is also set to 0.

3. When the P wave probability or S wave probability in the event window exceeds the corresponding threshold, the system records the time point corresponding to the probability peak as the candidate first arrival time. To verify the accuracy of the picking result, the method combines signal-to-noise ratio analysis, takes the candidate first arrival point as the center, and takes the 100 sampling points before and after as the noise window and signal window respectively. The 95th percentile ratio of the two is calculated and converted into decibel form. Only when the signal-to-noise ratio reaches a certain standard, the first arrival point is confirmed as a valid pick. In addition, the SP time difference constraint mechanism is introduced, and the upper limit is set to 60 seconds. It is required that the difference between the S wave first arrival time and the P wave first arrival time shall not exceed this value to eliminate unreasonable picking results. Through the comprehensive strategy of probability threshold screening, signal-to-noise ratio verification and time constraint, high-precision picking results of P wave and S wave first arrival are finally obtained.

Images