Earthquake first arrival pickup deep learning method based on attention mechanism and hybrid expert strategy

Abstract

The invention relates to the technical field of seismic signal detection, in particular to a seismic first arrival pick-up deep learning method based on an attention mechanism and a hybrid expert strategy, which comprises three steps of model building, a training stage and first arrival prediction. The core components of the model comprise a front-end feature extraction module, a hybrid expert module and a time domain decoding module. According to the invention, a more efficient expert routing mechanism is adopted in the network to reduce calculation redundancy, a more optimized load balancing strategy is used to improve training stability, and quantification and sparse technologies are combined to further reduce calculation and storage cost. According to the method, high-precision seismic first arrival pickup can still be kept under the conditions of complex seismic phase superposition and regional noise interference, high adaptability and generalization ability to various scenes are achieved, and the computing resource occupancy rate and cost can be greatly reduced under a large number of model parameters.

Description

Technical Field

[0001] The present invention relates to the technical field of earthquake signal detection, and in particular to a deep learning method for earthquake first arrival picking based on an attention mechanism and a hybrid expert strategy. Background Art

[0002] Existing methods for automatic earthquake first-arrival picking mainly include traditional threshold triggering methods and machine learning classification models. These technologies still have obvious limitations in noise adaptability, cross-regional generalization capabilities, and multi-phase identification, such as high computational cost and low efficiency. Especially in earthquake environments with low signal-to-noise ratio, it is difficult to accurately pick first arrivals.

[0003] In seismic applications, earthquake first-arrival picking plays a vital role in research work such as earthquake early warning, seismogenic structure, and magnitude estimation. However, due to the significant attenuation of seismic signal energy during propagation, and the influence of complex geological conditions, medium inhomogeneity, and instrument noise on seismic waves during propagation, the first-arrival wave signal is usually relatively weak and often accompanied by interference from other seismic phases. In this case, the feature extraction and precise positioning of the first-arrival wave become extremely difficult, which directly affects the accuracy of subsequent earthquake positioning and parameter inversion, and thus limits and affects the reliability and response speed of the earthquake monitoring and early warning system.

[0004] Traditional methods such as the STA / LTA threshold triggering method, the AIC criterion, and signal processing techniques based on short-time Fourier transform play a certain role in real-time processing and preliminary screening due to their simple principles and low computational complexity. However, these methods have poor robustness to noise and often experience false triggering or missed detections in complex seismic phase interference situations, making them unable to adapt to earthquakes with low signal-to-noise ratios. In addition, traditional methods often rely on manual experience to adjust parameters and lack automation and adaptability. They show obvious limitations in scenarios of large-scale real-time data processing, resulting in an insufficiently timely response from the earthquake early warning system.

[0005] In recent years, deep learning methods have been gradually applied to earthquake first-arrival detection. By constructing convolutional neural networks, long short-term memory networks, and Transformer model methods, automatic learning and extraction of seismic phase features have been achieved. These models have improved the shortcomings of traditional methods in noise suppression and subtle seismic phase detection to a certain extent, but there are still obvious defects: First, deep learning models usually require massive amounts of high-quality labeled data, and due to the weak and changeable characteristics of seismic data, it is difficult to obtain sufficient labeled samples; second, the existing models have insufficient generalization capabilities across regions and source types, and often experience performance degradation when facing unknown scenarios; third, because deep learning models often have a large number of model parameters, resulting in extremely high computational complexity, the training and prediction processes will occupy a lot of computing resources, which is difficult to meet the requirements of efficient response in real-time earthquake monitoring applications. At the same time, there are limitations in dealing with seismic phase aliasing and weak signals, such as misjudgment or missed detection.

[0006] To address the above shortcomings, the present invention proposes an intelligent earthquake first arrival picking method based on attention mechanism and hybrid expert strategy to achieve efficient and accurate detection of complex seismic signals. Summary of the Invention

[0007] In order to overcome the problems of high cost and low efficiency of existing automatic earthquake first-arrival picking methods, the present invention proposes an intelligent earthquake first-arrival picking method based on attention mechanism and hybrid expert strategy.

[0008] The technical solution of the present invention is: a deep learning method for earthquake first arrival picking based on attention mechanism and hybrid expert strategy, comprising the following steps:

[0009] Model construction: Its core components include a front-end feature extraction module, a hybrid expert module, and a time-domain decoding module. First, a convolutional encoder structure consisting of alternating layers of one-dimensional convolution and maximum pooling is used to extract seismic waveform features. A bidirectional LSTM is then introduced to model temporal features, and a self-attention mechanism is used to enhance key features. The hybrid expert layer dynamically activates the Top-K expert subnetwork to process features, and a load balancing strategy is used to optimize training. Ultimately, a picking curve with the same length as the original waveform is generated.

[0010] Training phase: Using the three-component seismic waveform as input and the first arrival probability sequence of P and S waves as output, a multi-branch loss joint optimization model is adopted. A load balancing regularization term is introduced to constrain the usage frequency of the expert subnetwork, and the Adam optimizer is used to train the model.

[0011] First arrival prediction: Perform forward reasoning on the input waveform to obtain the seismic phase picking probability sequence, output the P / S wave probability sequence, extract the first arrival time through the peak detection algorithm, and filter the final result in combination with physical constraints.

[0012] Preferably, in the convolution encoder structure composed of multiple layers of one-dimensional convolution and maximum pooling alternatingly, the convolution kernel size and stride of each convolution layer are set according to established rules.

[0013] Preferably, the convolution encoder structure composed of alternating multiple layers of one-dimensional convolution and maximum pooling retains the main phase characteristics of the waveform during the feature extraction process, which is used to improve the feature expression efficiency of subsequent processing modules.

[0014] Preferably, the bidirectional LSTM can process forward and reverse time series information simultaneously and fuse the context representations of the two directions in the hidden state.

[0015] Preferably, the number of hidden units in each LSTM layer in the bidirectional LSTM network remains consistent, the number of layers can be configured as needed, and the outputs of all layers are sequentially connected to form a deep sequence representation.

[0016] Preferably, the model is built on the output of bidirectional LSTM, and introduces a self-attention mechanism and a hybrid expert structure combination module. The self-attention mechanism calculates the correlation weights between each position in the input sequence to highlight the characteristics of the key time period; the hybrid expert module distributes the enhanced features to multiple sub-networks for parallel processing, and the routing network dynamically activates the Top-K expert subset according to the input content, and uses a load balancing regularization term to constrain the distribution of router output.

[0017] Preferably, the self-attention mechanism and hybrid expert structure combination module uses residual self-attention to extract features and feeds them into the hybrid expert layer to dynamically activate the Top-K expert network to complete the feedforward calculation:

[0018]

[0019] Among them, the moment represents the feature representation at time t in layer l; SelfAttention(·) is the self-attention mechanism with causal mask; RMSNorm(·) represents the root mean square normalization operation; FFN_i(·) refers to the i-th non-shared expert feedforward subnetwork; FFN_{N+1}(·) is the shared expert subnetwork with independent parameters: g i,t and g N+1,t , respectively, the gating coefficients corresponding to the non-shared experts and the shared experts; Express The intermediate representation after RMSNorm normalization.

[0020] Preferably, the training objective of the training phase is composed of multiple loss terms, where each output corresponds to a binary cross entropy loss function, and a category weighting strategy is introduced to distinguish between positive and negative samples. The multi-branch losses are weighted and combined according to the set weights as the overall optimization goal, which is used to uniformly drive the convergence of the detection branch and the two picking branches.

[0021] Preferably, during the training phase, the model adopts a feedforward module based on self-attention and hybrid expert structure, the router dynamically allocates each time slice to several expert sub-networks, and introduces a load balancing regularization term as an auxiliary loss in the training.

[0022] Preferably, the first arrival prediction step utilizes a trained multi-branch neural network model to perform forward reasoning on the input three-component earthquake waveform to obtain a seismic phase picking probability sequence.

[0023] Beneficial effects of the present invention:

[0024] 1. By adopting a more efficient expert routing mechanism in the network to reduce computational redundancy, using a more optimized load balancing strategy to improve training stability, and combining quantization and sparsity techniques to further reduce computing and storage costs, the present invention can still maintain high-precision earthquake first arrival detection under the conditions of complex seismic phase superposition and regional noise interference, has strong adaptability and generalization capabilities for various scenarios, and can significantly reduce computing resource utilization and costs under a large number of model parameters.

[0025] 3. The present invention designs an expert module with strong versatility, which can adapt to the processing requirements of different earthquake tasks and multiple seismic phase characteristics, thereby significantly improving the accuracy and robustness of earthquake first arrival picking, and providing more efficient and reliable technical support for earthquake early warning and geophysical research. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 What is shown is a schematic diagram of the model structure of the present invention;

[0027] Figure 2 Shown is a schematic diagram of the implementation process of the present invention;

[0028] Figure 3 Shown is a schematic diagram of the first arrival picking of P waves and S waves of the present invention. DETAILED DESCRIPTION

[0029] 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.

[0030] See also Figure 1 and Figure 2 The present invention provides an embodiment: a deep learning method for earthquake first arrival picking based on attention mechanism and hybrid expert strategy, comprising the following steps:

[0031] Model construction: Its core components include a front-end feature extraction module, a hybrid expert module, and a time-domain decoding module. First, a convolutional encoder structure consisting of alternating layers of one-dimensional convolution and maximum pooling is used to extract seismic waveform features. A bidirectional LSTM is then introduced to model temporal features, and a self-attention mechanism is used to enhance key features. The hybrid expert layer dynamically activates the Top-K expert subnetwork to process features, and a load balancing strategy is used to optimize training. Ultimately, a picking curve with the same length as the original waveform is generated.

[0032] Training phase: Using the three-component seismic waveform as input and the first arrival probability sequence of P and S waves as output, a multi-branch loss joint optimization model is adopted. A load balancing regularization term is introduced to constrain the usage frequency of the expert subnetwork, and the Adam optimizer is used to train the model.

[0033] First arrival prediction: Perform forward reasoning on the input waveform to obtain the seismic phase picking probability sequence, output the P / S wave probability sequence, extract the first arrival time through the peak detection algorithm, and filter the final result in combination with physical constraints.

[0034] Furthermore, for the model building steps, specifically:

[0035] First, a convolutional encoder structure consisting of multiple layers of alternating one-dimensional convolutions and maximum pooling is used to perform step-by-step downsampling and feature extraction on the input three-component seismic waveform data. The convolution kernel size and stride of each convolution layer are set according to established rules to extract waveform features at different time scales. At the same time, compression in the time dimension is performed through maximum pooling to further enhance the local translation invariance of the feature map. The encoder retains the main seismic phase characteristics of the waveform during feature extraction and improves the feature expression efficiency of subsequent processing modules.

[0036] A residual convolution stack is added to the output of the convolutional encoder. By introducing cross-layer connections, the gradient propagation capability is enhanced while maintaining the network depth. The residual structure can alleviate the performance degradation caused by network deepening and enable the model to maintain stable expression capabilities in deeper feature spaces. The number of convolution channels and the structure of the residual stack are set according to a fixed strategy to ensure that information can be effectively accumulated and propagated in deep feature spaces.

[0037] The model introduces a bidirectional long short-term memory network structure and performs bidirectional modeling of time series features. The bidirectional structure can process forward and reverse time series information simultaneously and fuse the contextual representations of the two directions in the hidden state, thereby improving the model's ability to express the time series changes of seismic waveforms. The number of hidden units in each LSTM layer in the network remains consistent, and the number of layers can be configured as needed. The outputs of all layers are connected in sequence to form a deep sequence representation.

[0038] Based on the output of the bidirectional LSTM, the model introduces a module consisting of a self-attention mechanism and a hybrid expert structure. The self-attention mechanism highlights the features of key time periods by calculating the correlation weights between each position in the input sequence. Subsequently, the hybrid expert module is used to distribute the enhanced features to multiple sub-networks for parallel processing, thereby achieving more accurate feature extraction and expression. The routing network determines the activated expert subset based on the input content. The experts use independent parameter configurations and do not share weights. During the training phase, a load balancing regularization term is added to constrain the distribution of router outputs to ensure a balanced usage frequency among the experts.

[0039] This module uses residual self-attention to extract contextual features and feeds them into the hybrid expert layer to dynamically activate the Top-K expert network to complete the feedforward calculation:

[0040]

[0041] Among them, the moment represents the feature representation at time t in layer l; SelfAttention(·) is the self-attention mechanism with causal mask; RMSNorm(·) represents the root mean square normalization operation; FFN_i(·) refers to the i-th non-shared expert feedforward subnetwork; FFN_{N+1}(·) is the shared expert subnetwork with independent parameters: g i,t and g N+1,t , respectively, the gating coefficients corresponding to the non-shared experts and the shared experts; Express The intermediate representation after RMSNorm normalization.

[0042] The model output stage consists of detection and picking branches. The detection branch reconstructs the time dimension of features through upsampling and convolution operations, and outputs a detection curve with the same length as the input, which is used to represent the probability of event occurrence corresponding to each time point. The picking branch includes a unidirectional LSTM, a self-attention module and a decoder structure. The unidirectional LSTM is used to model the time features in a specific direction, and self-attention is used to enhance the information expression of key positions in the sequence. The decoder restores the time resolution through continuous upsampling and convolution operations, and finally generates a picking curve with the same length as the original waveform to represent the confidence of the first arrival of the seismic phase at each time point.

[0043] Furthermore, for the training phase steps, specifically:

[0044] During the model training phase, the present invention is based on a supervised learning framework, combined with a multi-branch structure and a hybrid expert mechanism, to jointly optimize the parameters of each model module. The model receives a three-component seismic waveform of fixed length as input, and outputs a P-wave first arrival curve and an S-wave first arrival curve, respectively. The training objective consists of multiple loss terms, where each output corresponds to a binary cross-entropy loss function, and a category weighting strategy is introduced to distinguish between positive and negative samples to alleviate the gradient offset problem caused by sparse seismic phase labels. The multi-branch losses are weighted and combined according to the set weights as the overall optimization target, which is used to uniformly drive the convergence of the detection branch and the two picking branches.

[0045] During training, the model uses a feedforward module based on self-attention and a hybrid expert structure. The router dynamically allocates each time slice to several expert sub-networks, and introduces a load balancing regularization term as an auxiliary loss during training to ensure balanced use of each expert during the optimization process. The optimizer uses the Adam algorithm, combined with an adaptive learning rate adjustment strategy, to jointly minimize the main loss and auxiliary loss. In each iteration, training data is loaded in batches, and forward calculations, loss evaluation, gradient backpropagation, and parameter updates are performed. The branch loss and corresponding F1 score are recorded as performance indicators. The verification phase uses frozen parameters for prediction and evaluation to dynamically determine the model's generalization ability. During training, the optimal model parameters are saved and training logs are continuously recorded to achieve an automated and efficient training process for the model in a high-performance environment.

[0046] Furthermore, for the first arrival prediction step, specifically:

[0047] The present invention utilizes a trained multi-branch neural network model to perform forward reasoning on the input three-component earthquake waveform to obtain a probability sequence for seismic phase picking. The model output includes two branches, corresponding to P-wave picking and S-wave picking, respectively. The length of the probability sequence output by each branch is consistent with the number of sampling points in the input waveform. For the picking branches, a local maximum search algorithm is used to identify the peak points in the P-wave and S-wave probability curves. By setting parameters such as the minimum peak height and the minimum adjacent peak spacing, non-real picking points with insufficient amplitude or too close intervals are filtered out. The model reasoning output is decoupled from the picking algorithm. The picking strategy is based on the event detection interval and combines the seismic phase probability curve to screen and select the picking candidates. If there are multiple seismic phase candidate points that meet the conditions, they are sorted according to their corresponding probability values, and only the results with the highest confidence are retained. To further constrain the physical rationality of the picking results, the system limits the P-S wave time difference and eliminates results with unreasonable phase time sequence or difference exceeding the set upper limit. The final output includes the start and end time of the event, the first arrival time of the P wave and S wave, their respective probability values, and the corresponding signal-to-noise ratio estimate. The entire picking process is completed through the coordinated output probability of the model and the rule screening strategy. It has end-to-end structural characteristics and is suitable for automated first arrival extraction of large-scale seismic waveforms.

[0048] See also Figure 3 The present invention provides an embodiment to illustrate the earthquake event with a magnitude of 2.8 that occurred at 11:31:19 on November 13, 2010. The epicenter of this earthquake was located at 13.416546°S and 135.655832°E, with a focal depth of 28.9 km. The purpose is to verify the practical applicability of the method of the present invention in a strong noise environment. The present invention uses the earthquake waveform recorded by the PFVI station as the analysis object. The data used is saved in HDF5 format and contains three directional components: E, N, and Z. Each component contains 6000 sampling points, the sampling rate is 100 Hz, and the total recording time is 60 seconds.

[0049] To address the significant noise interference in the original waveform, the data is first standardized by subtracting the mean of each component data and dividing it by the standard deviation to obtain a normalized result with zero mean. Subsequently, a bandpass filter in the range of 1-45Hz is applied to remove high-frequency interference components and low-frequency trend terms in the data, further enhancing the ability to identify key phases in the seismic signal. The data after the above preprocessing is input into a pre-trained deep learning model to output a probability prediction sequence of P waves and S waves. Although the original waveform contains complex noise, the present invention can still accurately identify the first arrival time of P waves and S waves. Its detection results are consistent with manual annotations, verifying the robustness and effectiveness of the present invention under strong noise conditions.

[0050] Through the above steps, by adopting a more efficient expert routing mechanism in the network to reduce computational redundancy, using a more optimized load balancing strategy to improve training stability, and combining quantization and sparsity techniques to further reduce computing and storage costs, this method is suitable for different types of seismic tasks to solve the problems of high computational cost and low efficiency of existing automatic earthquake first arrival picking methods.

Claims

1. A deep learning method for earthquake first arrival picking based on attention mechanism and hybrid expert strategy, characterized by: The following steps are included: Model construction: Its core components include a front-end feature extraction module, a hybrid expert module, and a time-domain decoding module. First, a convolutional encoder structure consisting of alternating layers of one-dimensional convolution and maximum pooling is used to extract seismic waveform features. A bidirectional LSTM is then introduced to model temporal features, and a self-attention mechanism is used to enhance key features. The hybrid expert layer dynamically activates the Top-K expert subnetwork to process features, and a load balancing strategy is used to optimize training. Ultimately, a picking curve with the same length as the original waveform is generated. Training phase: Using the three-component seismic waveform as input and the first arrival probability sequence of P and S waves as output, a multi-branch loss joint optimization model is adopted. A load balancing regularization term is introduced to constrain the usage frequency of the expert subnetwork, and the Adam optimizer is used to train the model. First arrival prediction: Perform forward reasoning on the input waveform to obtain the seismic phase picking probability sequence, output the P / S wave probability sequence, extract the first arrival time through the peak detection algorithm, and filter the final result in combination with physical constraints.

2. The deep learning method for earthquake first arrival picking based on attention mechanism and hybrid expert strategy according to claim 1 is characterized by: In the convolutional encoder structure composed of multiple layers of one-dimensional convolution and maximum pooling alternately, the convolution kernel size and stride of each convolution layer are set according to established rules.

3. The deep learning method for earthquake first arrival picking based on attention mechanism and hybrid expert strategy according to claim 2 is characterized by: The convolutional encoder structure composed of alternating multiple layers of one-dimensional convolution and maximum pooling retains the main phase characteristics of the waveform during the feature extraction process, which is used to improve the feature expression efficiency of subsequent processing modules.

4. The deep learning method for earthquake first arrival picking based on attention mechanism and hybrid expert strategy according to claim 3 is characterized by: The bidirectional LSTM can process forward and reverse time series information simultaneously and fuse the context representations of the two directions in the hidden state.

5. The deep learning method for earthquake first arrival picking based on attention mechanism and hybrid expert strategy according to claim 1 is characterized in that: The number of hidden units in each LSTM layer of the bidirectional LSTM network remains consistent, the number of layers can be configured as needed, and the outputs of all layers are connected in sequence to form a deep sequence representation.

6. The deep learning method for earthquake first arrival picking based on attention mechanism and hybrid expert strategy according to claim 1 is characterized by: The model is built on the output of a bidirectional LSTM and introduces a self-attention mechanism and a hybrid expert structure combination module. The self-attention mechanism calculates the correlation weights between each position in the input sequence to highlight the features of key time periods; the hybrid expert module distributes the enhanced features to multiple subnetworks for parallel processing. The routing network dynamically activates the Top-K expert subset based on the input content and uses a load balancing regularization term to constrain the distribution of router outputs.

7. The deep learning method for earthquake first arrival picking based on attention mechanism and hybrid expert strategy according to claim 6 is characterized by: The self-attention mechanism and hybrid expert structure combination module uses residual self-attention to extract features and sends them to the hybrid expert layer to dynamically activate the Top-K expert network to complete the feedforward calculation: Among them, the moment represents the feature representation at time t in layer l; SelfAttention(·) is the self-attention mechanism with causal mask; RMSNorm(·) represents the root mean square normalization operation; FFN_i(·) refers to the i-th non-shared expert feedforward subnetwork; FFN_{N+1}(·) is the shared expert subnetwork with independent parameters: g i,t and g N+1,t , respectively, the gating coefficients corresponding to the non-shared experts and the shared experts; Express The intermediate representation after RMSNorm normalization.

8. The deep learning method for earthquake first arrival picking based on attention mechanism and hybrid expert strategy according to claim 1 is characterized by: The training objective of the training phase consists of multiple loss terms, where each output corresponds to a binary cross entropy loss function, and a category weighting strategy is introduced to distinguish between positive and negative samples. The multi-branch losses are weighted and combined according to the set weights as the overall optimization goal, which is used to uniformly drive the convergence of the detection branch and the two picking branches.

9. The deep learning method for earthquake first arrival picking based on attention mechanism and hybrid expert strategy according to claim 8 is characterized by: During the training phase, the model adopts a feedforward module based on self-attention and hybrid expert structure, the router dynamically allocates each time slice to several expert sub-networks, and introduces a load balancing regularization term as an auxiliary loss in the training.

10. The deep learning method for earthquake first arrival picking based on attention mechanism and hybrid expert strategy according to claim 1 is characterized in that: The first arrival prediction step utilizes a trained multi-branch neural network model to perform forward reasoning on the input three-component earthquake waveform to obtain a seismic phase picking probability sequence.

Images