A method for picking up seismic P waves based on SAIP-Unet network

Abstract

The invention belongs to the technical field of earthquake phase arrival time picking and image processing, and is an earthquake P wave picking method based on a SAIP-Unet network; the SAIP-Unet network is established on the basis of a codec structure; firstly, a network encoder uses a multi-scale residual module to deeply mine effective features and fuses and outputs them through a multi-scale feature extraction module to accelerate the network convergence speed; then, a spatiotemporal attention mechanism is added to the decoder to generate a high-precision arrival sequence and then perform point-by-point prediction, thereby enhancing the network's perception of arrival features and improving recognition accuracy; compared with traditional methods and deep learning methods, the invention has higher arrival time picking accuracy; and solves the problems of poor efficiency, low accuracy, weak robustness, etc. of existing earthquake arrival time picking algorithms.

Description

Technical Field

[0001] The invention belongs to the technical field of earthquake phase arrival time picking and image processing, and relates to an earthquake P wave picking method based on a SAIP-Unet network. Background Art

[0002] Seismic P waves are the first waveforms to reach the surface after an earthquake, and are the most frequently occurring waveforms in seismic phase analysis. Accurately picking up P waves at the time of arrival is a difficult and hot topic in the current research field of seismic signal processing. Rapid and accurate picking up of P waves at the time of arrival can improve the integrity and accuracy of earthquake cataloging, and is of great significance to earthquake location, seismic phase analysis, and magnitude determination. There are currently three main methods for picking up seismic P waves at the time of arrival: manual picking, traditional methods, and deep learning methods. The manual picking method is the method with the highest picking accuracy currently, in which professionals manually analyze seismic waveforms and identify P waves, but the picking efficiency of this method is extremely low; the traditional algorithm is the most widely used method for picking up arrival time, and its representative algorithms include STA / LTA, AR-AIC, etc. These algorithms are mostly based on characteristic equations based on characteristic information such as amplitude and frequency for picking up arrival time, and have high picking efficiency, but the detection effect is extremely dependent on the setting of the threshold and the algorithm robustness is poor.

[0003] Seismic phase picking algorithms based on deep learning technology have emerged in recent years and have been widely recognized, and a variety of seismic phase identification and arrival-time picking models such as ConvNetQuake and PhaseNet have been developed. This type of method mainly achieves arrival-time picking by performing nonlinear mapping modeling on seismic signals. It has high picking efficiency and strong robustness, and is the main direction of future arrival-time picking technology development. However, the existing deep learning arrival-time picking methods have weak perception capabilities for arrival-time features, and the recognition accuracy needs to be improved urgently. Summary of the invention

[0004] The present invention overcomes the shortcomings of the prior art and proposes a method for picking up seismic P waves based on the SAIP-Unet network to improve the model's parameter control ability and the ability to perceive the characteristics of seismic P waves to pick up accurate seismic phase arrival information, thereby solving the problems of poor efficiency, low precision, and weak robustness of the existing seismic arrival time picking algorithms.

[0005] In order to achieve the above object, the present invention is implemented through the following technical solutions.

[0006] A method for picking up seismic P waves based on a SAIP-Unet network comprises the following steps:

[0007] Step S1: Use the three-component data of the seismic waveform with the P-wave arrival time manually marked as the original data for preprocessing.

[0008] Step S2: Expand the annotation range of the P wave arrival time, process the labels and prediction results, and randomly divide the samples in the data set into training set and test set.

[0009] Step S3: Construct a lightweight arrival time picking network SAIP-Unet to train the seismic signal data set; the SAIP-Unet network consists of an encoder and a decoder; the encoder performs deep feature extraction on the input seismic signal data, enhances the weight of the seismic phase arrival time in the sequence, and suppresses and filters other interference information; the decoder restores the feature sequence to the same length as the input data, and finally identifies each time step to realize Seq2Seq.

[0010] Step S4: Pick up the arrival time of earthquake P waves through the SAIP-Unet network: it is to fuse the channel attention mechanism and the time step attention mechanism to construct a spatiotemporal attention mechanism, and improve the network representation ability by analyzing the interdependence between data channels and the time correlation between time steps; after receiving the fused features, the spatiotemporal attention module first calculates the channel attention weight and the time step attention weight respectively, and then calculates the Hadamard product of the two to obtain the fused attention weight, and finally reassigns the module input based on the fused weight, so that the network pays attention to important features and suppresses unimportant features.

[0011] Step S5: Train the test set under different error thresholds, and obtain the recognition accuracy and overall performance of the network model based on the evaluation indicators.

[0012] Preferably, the preprocessing described in step S1 is to cut the data into 10 s data segments including the first arrival time of the P wave, remove the linear trend of the data by differential method, eliminate noise by bandpass filtering and wavelet transform, and normalize the data.

[0013] Preferably, the method of removing the linear trend of the data is to use the difference method to remove the linear trend of the data, that is, to calculate the difference between time t and time t-1 to construct a new sequence as shown in formula (1), where T new represents a new sequence; T old represents the old sequence; t is the processed time step, making the data stable;

[0014] T new (t) = T old (t)-T old (t-1) (1)

[0015] The bandpass filtering is to use a 0.1-20 Hz bandpass filter to filter the data to reduce background noise contained in the data.

[0016] The wavelet transform uses Daubechies wavelet to map the signal to the wavelet domain, processes the wavelet coefficients of the noisy signal according to the different properties and mechanisms of noise and the wavelet coefficients of the noise at different scales, and eliminates the noise through shortwave.

[0017] The normalization is to perform Min-Max normalization on the data to speed up the network convergence.

[0018] Preferably, in step S4, the fusion channel attention mechanism and the time step attention mechanism to construct the spatiotemporal attention mechanism includes formula (2-4):

[0019]

[0020]

[0021]

[0022] Among them, FA, TA and SA represent the channel attention weight, time step attention weight and spatiotemporal attention weight respectively; * represents the Hadamard product operation.

[0023] Preferably, in step S5, the evaluation indicators include: mean absolute error MAE, mean square error MSE, and hit rate HR.

[0024] Preferably, the mean absolute error MAE is calculated as shown in formula (5):

[0025]

[0026] in Indicates when the algorithm recognizes the arrival of the P wave; L i It indicates the arrival time of the manually marked P wave; n is the total amount of test data;

[0027] The mean square error calculation method is shown in formula (6):

[0028]

[0029] The hit rate calculation method is shown in formula (7-8):

[0030]

[0031]

[0032] Where β represents the error threshold; α represents the number of data picked up with an absolute error less than the threshold;

[0033] Preferably, after step S5, an ablation experiment is performed and the results are analyzed and evaluated.

[0034] The beneficial effects of the present invention compared with the prior art are as follows:

[0035] The method of the invention uses Unet as the basic structure, integrates the spatiotemporal attention mechanism to improve the feature perception ability of the network, and adopts the Inception structure to enhance the stability of the codec. A low-error lightweight P-wave arrival picking model Spatiotemporal Attention-Inception-Pwave-Pick-Unet (SAIP-Unet for short) is designed. The model has strong earthquake phase feature analysis and fusion capabilities. Compared with the current advanced traditional algorithms and deep learning SOTA methods, the overall performance is stronger and has higher P-wave arrival picking accuracy.

[0036] The present invention introduces the idea of ​​attention and realizes the accurate picking of earthquake P wave arrival time through a sequence generation model with a codec structure, thereby enhancing the model's ability to mine seismic signal features and improving the model's arrival time picking accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flow chart of a method for picking up seismic P waves based on SAIP-Unet network according to an embodiment of the present invention.

[0038] Figure 2 It is a network structure diagram of a seismic P wave picking method based on SAIP-Unet network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0039] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention clearer, the present invention is further described in detail in conjunction with the embodiments and the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. The technical solutions of the present invention are described in detail below in conjunction with the embodiments and the accompanying drawings, but the scope of protection is not limited thereto.

[0040] Reference Figure 1 This embodiment proposes a seismic P wave picking method based on SAIP-Unet network, which is specifically performed according to the following steps:

[0041] Step S1: Use the three-component data of the seismic waveform with the P wave arrival time manually marked as the original data and take 1200 data in the original data as experimental data for preprocessing.

[0042] Step S2: Expand the annotation range of the P wave arrival time to 20 time steps centered on the manually annotated time to process the labels and prediction results, and randomly divide the samples in the data set into a training set and a test set.

[0043] Step S3: Construct a lightweight arrival time picking network SAIP-Unet with strong earthquake phase feature analysis and fusion capabilities to train the seismic signal dataset.

[0044] Step S4: In order to improve the recognition accuracy of the model and speed up the model training, the constructed network model is improved for P-wave arrival time picking.

[0045] Step S5: Train the test set under different error thresholds, and obtain the recognition accuracy and overall performance of the model based on the evaluation indicators.

[0046] Step S6: Explore the effectiveness of each module and mechanism in the constructed network, conduct a series of ablation experiments, and analyze and evaluate the experimental results.

[0047] In step S1, image preprocessing requires cutting the data into 10 s data segments including the first arrival time of the P wave, using the difference method to remove the linear trend of the data, using bandpass filtering and wavelet transform to eliminate noise, and also normalizing the data to speed up the convergence of the network.

[0048] Removing linear trends: Use the difference method to remove the linear trends of the data, that is, calculate the difference between time t and time t-1 to construct a new sequence as shown in formula (1), where T new represents a new sequence; T old represents the old sequence; t is the processed time step to make the data stable.

[0049] T new (t) = T old (t)-T old (t-1) (1)

[0050] Bandpass filtering: Use a 0.1-20 Hz bandpass filter to filter the data to reduce background noise contained in the data.

[0051] Wavelet transform: Daubechies wavelet is used to map the signal to the wavelet domain. Since noise and the wavelet coefficients of noise have different properties and mechanisms at different scales, the wavelet coefficients of the noisy signal are processed and noise elimination is achieved through shortwave.

[0052] Normalization: Min-Max normalization is performed on the data to speed up network convergence.

[0053] In step S2, since the arrival of the P wave is only a unit time step in the data sequence, it accounts for a very small proportion compared to the sequence containing 1000 moments, which is not conducive to feature learning of the network, so it is necessary to expand the annotation range of the arrival of the P wave. After the samples in the data set are randomly divided, the training set contains 10,000 data for training the deep neural network model, and the test set contains 2,000 data for testing and evaluating the recognition effect of the model. When the network completes the recognition of the input data, it outputs an arrival probability sequence of the same length as the data, and takes the time step with the highest probability in the sequence as the arrival of the P wave.

[0054] In step S3, a time-of-arrival picking model based on the SAIP-Unet network is constructed, wherein the SAIP-Unet network is composed of an encoder and a decoder. Some structures in the encoder and decoder draw on the Inception structure, and extract the earthquake phase arrival features based on its excellent parameter control and feature extraction capabilities to improve feature utilization; the encoder can perform deep feature extraction on the input seismic signal data, enhance the weight of the phase arrival in the sequence, and suppress and filter other interference information; the decoder can restore the feature sequence to the same length as the input data, and finally identify each time step to achieve Seq2Seq.

[0055] In step S4, the channel attention mechanism and the time step attention mechanism are fused to construct a spatiotemporal attention mechanism. By analyzing the interdependence between data channels and the temporal correlation between time steps, the network representation capability is improved. After receiving the fused features, the spatiotemporal attention module first calculates the channel attention weight and the time step attention weight respectively, then calculates the Hadamard product of the two to obtain the fused attention weight, and finally re-assigns the module input based on the fused weight, so that the network pays attention to important features and suppresses unimportant features. The specific fusion process is shown in formula (2-4).

[0056]

[0057]

[0058]

[0059] Among them, FA, TA and SA represent the channel attention weight, time step attention weight and spatiotemporal attention weight respectively; * represents the Hadamard product operation.

[0060] In step S5, the evaluation indicators for training the deep learning model include: mean absolute error (MAE), mean squared error (MSE) and hit rate (HR).

[0061] The mean absolute error is used to directly reflect the error of network recognition at the time of arrival. It is an intuitive reflection of the network's arrival-time picking accuracy. The closer the index value is to 0, the higher the network's arrival-time picking accuracy is. The specific calculation method is shown in formula (5).

[0062]

[0063] in Indicates when the algorithm recognizes the arrival of the P wave; L i Indicates the manually marked P wave arrival time; n is the total amount of test data.

[0064] The mean square error is used to measure the deviation between the time when the network recognizes and the time when it is labeled, and can measure the network fitting ability. The closer the index value is to 0, the stronger the network fitting ability and recognition robustness are. The specific calculation method is shown in formula (6).

[0065]

[0066] The hit rate is used to indicate the percentage of the absolute error between the network-picked arrival time and the annotated arrival time that is less than the error threshold. It is used to measure the network's learning ability for the P-wave arrival time feature. The closer this indicator is to 100%, the stronger the network's feature extraction performance is. The specific calculation method is shown in formula (7-8).

[0067]

[0068]

[0069] Wherein, β represents an error threshold (in this embodiment, four error thresholds of 0.1s, 0.2s, 0.3s, and 0.5s are used to measure the recognition results); α represents the number of data picked up whose absolute error is less than the threshold.

[0070] In step S6, for the main modules of this paper, we build multiple networks to verify the effectiveness of each module, and finally prove that the final module of this paper has stronger on-time picking performance.

[0071] Experimental results show that the proposed network has a picking hit rate of 96.9%, 98.2%, 98.7% and 99.2% under four error thresholds of 0.1s, 0.2s, 0.3s and 0.5s, respectively, and the mean absolute error and mean square error are 0.074s and 0.072, respectively. Compared with traditional methods and deep learning methods, it has higher on-time picking accuracy.

[0072] The above content is a further detailed description of the present invention in combination with a specific preferred embodiment. It cannot be determined that the specific embodiments of the present invention are limited to this. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the present invention, which should be regarded as belonging to the present invention and the scope of patent protection determined by the submitted claims.

Claims

1. A method for picking up seismic P waves based on SAIP-Unet network, characterized in that: The following steps are involved: Step S1: using the three-component data of the seismic waveform with the P wave arrival time manually marked as the original data for preprocessing; Step S2: Expand the annotation range of the P wave arrival time, process the labels and prediction results, and randomly divide the samples in the data set into a training set and a test set; Step S3: construct a lightweight arrival time picking network SAIP-Unet to train the seismic signal data set; the SAIP-Unet network consists of an encoder and a decoder; the encoder performs deep feature extraction on the input seismic signal data, enhances the weight of the seismic phase arrival time in the sequence, and suppresses and filters other interference information; the decoder restores the feature sequence to the same length as the input data, and finally identifies each time step to realize Seq2Seq; Step S4: Pick up the arrival time of earthquake P waves through the SAIP-Unet network: It is to fuse the channel attention mechanism and the time step attention mechanism to construct a spatiotemporal attention mechanism, and improve the network representation ability by analyzing the interdependence between data channels and the time correlation between time steps; after receiving the fused features, the spatiotemporal attention module first calculates the channel attention weight and the time step attention weight respectively, and then calculates the Hadamard product of the two to obtain the fused attention weight, and finally re-assigns the module input based on the fused attention weight, so that the network pays attention to important features and suppresses unimportant features; Step S5: training the test set under different error thresholds, and obtaining the recognition accuracy and overall performance of the network model according to the evaluation indicators; In step S4, the fusion channel attention mechanism and the time step attention mechanism are used to construct the spatiotemporal attention mechanism, including formulas (2)-(4): Among them, FA, TA and SA represent channel attention weight, time step attention weight and spatiotemporal attention weight respectively; * represents the Hadamard product operation.

2. A method for picking up seismic P waves based on SAIP-Unet network according to claim 1, characterized in that: The preprocessing described in step S1 is to cut the data into 10 s data segments including the first arrival time of the P wave, use the difference method to remove the linear trend of the data, use bandpass filtering and wavelet transform to eliminate noise, and normalize the data.

3. A method for picking up seismic P waves based on SAIP-Unet network according to claim 2, characterized in that: The method of removing the linear trend of the data is to use the difference method to remove the linear trend of the data, that is, to calculate the difference between time t and time t-1 to construct a new sequence as shown in formula (1), where T new represents a new sequence; T old represents the old sequence; t is the processed time step, making the data stable; T new (t)=T old (t)-T old (t-1)(1) The bandpass filtering is to use a 0.1-20 Hz bandpass filter to filter the data to reduce the background noise contained in the data; The wavelet transform uses Daubechies wavelet to map the signal to the wavelet domain, processes the wavelet coefficients of the noisy signal according to the different properties and mechanisms of noise and the wavelet coefficients of the noise at different scales, and eliminates the noise through shortwave. The normalization is to perform Min-Max normalization on the data to speed up the network convergence.

4. The method for picking up seismic P waves based on SAIP-Unet network according to claim 1, characterized in that: In step S5, the evaluation indicators include: mean absolute error MAE, mean square error MSE, and hit rate HR.

5. A method for picking up seismic P waves based on SAIP-Unet network according to claim 4, characterized in that: The mean absolute error MAE calculation method is shown in formula (5): in Indicates when the algorithm recognizes the arrival of the P wave; L i It indicates the arrival time of the manually marked P wave; n is the total amount of test data; The mean square error calculation method is shown in formula (6): The hit rate calculation method is shown in formula (7)-(8): Where β represents the error threshold; α represents the number of data picked up whose absolute error is less than the threshold.

6. The method for picking up seismic P waves based on SAIP-Unet network according to claim 1, characterized in that: After step S5 is completed, an ablation experiment is performed and the results are analyzed and evaluated.

Images