Seismic signal phase pickup method based on multi-scale deep learning and attention mechanism

Abstract

The invention belongs to the technical field of signal processing and deep learning, and particularly relates to a seismic signal phase pickup method based on multi-scale deep learning and an attention mechanism. Aiming at the problems of high S-wave omission ratio and poor fixed threshold adaptability of the existing phase pickup method in a complex noise environment, the method is based on a multi-scale one-dimensional U-Net architecture, and adopts parallel depth separable convolution with the convolution kernel size of 3 / 7 / 15 to extract multi-scale time features. Key feature expression is enhanced in a bottleneck layer and a decoder in combination with channel and space attention (CBAM), and a dynamic threshold calibration strategy based on verification set F1 score optimization is introduced, so that an end-to-end phase pickup model is constructed. Therefore, the earthquake signal phase pickup is realized, the leak detection risk is effectively reduced, and the method is suitable for an earthquake monitoring and early warning system.

Description

Technical Field

[0001] This invention belongs to the field of signal processing and deep learning technology, specifically relating to a seismic signal phase picking method based on multi-scale deep learning and attention mechanism. Background Technology

[0002] Seismic signal phase acquisition, especially the accurate identification of P-waves and S-waves, is a core component of microseismic monitoring, natural earthquake location, and disaster early warning systems. In this field, the tolerance for missed detections is extremely low; missed S-wave detections directly lead to significant deviations in the calculation of focal depth and location. Studies show that a missed detection rate exceeding 10% can increase location errors by more than 20%, severely impacting the accuracy of emergency rescue decisions. Traditional methods (such as the STA / LTA algorithm) and early deep learning models (such as PhaseNet) generally follow an "accuracy-first" design paradigm. However, in complex and noisy environments such as mines and active faults in urban areas, this paradigm faces severe challenges: models tend to make conservative judgments in pursuit of high accuracy, resulting in a large number of missed detections of S-wave signals with low signal-to-noise ratios. False detections can be corrected through post-processing (such as multi-station joint verification), but missed detections constitute irreversible information loss, becoming a bottleneck restricting the reliability of the monitoring system. The baseline PhaseNet model achieves an S-wave recall of only 0.9350 on publicly available datasets, and its false negative risk is significant in high-precision applications. Therefore, developing a novel phase picking method that can significantly improve recall, especially S-wave recall, is of urgent practical engineering importance. The primary goal of this invention is to achieve a qualitative improvement in recall, particularly S-wave recall, within an acceptable accuracy range through technological innovation, thereby fundamentally reducing the risk of false negatives. Summary of the Invention

[0003] To address the industry pain point of high S-wave false negative rates in complex noise environments using existing phase picking methods, this invention provides a seismic signal phase picking method based on multi-scale deep learning and attention mechanisms. This method can improve recall while maintaining F1 score advantage. This invention is based on a multi-scale one-dimensional convolutional neural network (CNN) architecture, combining channel and spatial attention (CBAM) and depthwise separable convolution to reduce computational costs. It also introduces a dynamic threshold calibration strategy based on validation set F1 score optimization, thereby constructing an end-to-end phase picking model.

[0004] To achieve the above objectives, the present invention employs the following technical solution:

[0005] This invention provides a seismic signal phase picking method based on multi-scale deep learning and attention mechanism, comprising the following steps:

[0006] Step 1: Obtain three-component seismic waveform data in shape [T, 3], where T is the number of time sampling points and 3 corresponds to the three components Z / N / E. Perform zero-mean normalization on each component, i.e., subtract the mean of the component and divide by the standard deviation. Prune or slide window the continuous waveform according to a fixed time window and convert the preprocessed waveform data into tensor shape [B, 3, T], where B is the batch size.

[0007] Step 2: Generate soft tags based on the arrival times of P-waves and S-waves marked by seismologists. Generate a probability curve for P-waves using a Gaussian window function centered on the arrival time of P-waves, and a probability curve for S-waves using the same parameters centered on the arrival time of S-waves. The background noise probability is defined as 1 minus the sum of the probabilities of P-waves and S-waves, ultimately forming a three-channel soft tag tensor [B, 3, T], where the sum of the three probabilities at each time point is 1.

[0008] Step 3: Input the preprocessed seismic waveform data into the pre-trained seismic signal phase picking model. The model adopts a multi-scale one-dimensional U-Net architecture. The encoder processes the data layer by layer through multiple EncoderBlocks. Each layer first extracts waveform features at different time scales using multi-scale depth-separable convolutions, and then compresses the time dimension to 1 / 4 of the original using MaxPool1D downsampling. At the same time, the number of channels is doubled layer by layer, and the feature maps of each layer are temporarily stored for subsequent skip connections. The bottleneck layer first doubles the number of channels using multi-scale depth-separable convolutions, and then strengthens key information in two steps using the CBAM1D attention module. First, channel attention is used to learn channel weights through global pooling and shared MLP, and then temporal attention is used. The P-wave and S-wave are focused into the temporal candidate region through channel-dimensional pooling and one-dimensional convolution. The decoder recovers the temporal resolution by upsampling layer by layer through multiple DecoderBlocks. Each layer first uses transposed convolution to upsample and amplify the temporal dimension to 4 times and halve the number of channels. Then, it fuses the high-resolution features of the corresponding layer of the encoder after center alignment through skip connections. Next, the P-wave and S-wave features are further enhanced by the CBAM1D attention module. Then, the fused features are refined by multi-scale depth separable convolution. The classification head maps the decoder output to a three-channel logits tensor through 1×1 convolution. The output logits are normalized by softmax in the channel dimension to obtain the temporal probability maps of P-wave, S-wave and background noise.

[0009] Step 4: Invoke the optimal thresholds determined for P-waves and S-waves through grid search during the training phase, and judge the probability map point by point. When the probability of P-wave exceeds the optimal threshold of P-wave, mark the time point as a candidate for P-wave arrival. When the probability of S-wave exceeds the optimal threshold of S-wave, mark the time point as a candidate for S-wave arrival. For fixed window data, directly select the candidate position with the highest probability as the final arrival time of P-wave and S-wave. For continuous data stream, apply the peak detection algorithm to the candidate time points, requiring that the interval between adjacent peaks is greater than the minimum interval. Select the peak position with the highest probability as the final arrival time of P-wave and S-wave. Output the picking results of P-wave and S-wave, including the arrival timestamp and the corresponding confidence level.

[0010] Preferably, the seismic signal phase picking model is as follows:

[0011] The encoder consists of multiple EncoderBlocks cascaded together. Each EncoderBlock contains a multi-scale depthwise separable convolutional module and a max pooling layer. The multi-scale depthwise separable convolutional module uses three parallel branches with convolutional kernels of different sizes. Each branch first performs depthwise convolution and then pointwise convolution to map the number of channels to the target value. The outputs of the three branches are concatenated in the channel dimension and then fused by a 1×1 convolution. The max pooling layer compresses the time dimension to 1 / 4 of the original. The number of output channels of each layer increases according to the formula C_out=base_ch×2^d, where d is the layer depth index and base_ch is the number of base channels. The feature map before downsampling is saved as skip connection features.

[0012] The bottleneck layer first doubles the number of input channels through multi-scale depthwise separable convolution, and then enhances them through the CBAM1D attention module. The channel attention submodule of the CBAM1D attention module performs global average pooling and global max pooling on the feature map in the time dimension. The two pooling results are fed into a shared MLP. The outputs of the two MLPs are added and activated by Sigmoid to obtain channel weights, which are then multiplied element-wise with the original feature map to achieve channel weighting. The temporal attention submodule of the CBAM1D attention module performs average pooling and max pooling on the channel-weighted feature map in the channel dimension. The two results are concatenated in the channel dimension and fed into a one-dimensional convolutional layer. The output is activated by Sigmoid to obtain temporal weights, which are then multiplied element-wise with the feature map to achieve temporal weighting.

[0013] The decoder consists of multiple DecoderBlocks concatenated. Each DecoderBlock contains a transposed convolutional upsampling layer, skip connection fusion, a CBAM1D attention module, and a multi-scale depth-separable convolution. The skip connection fusion centers the skip connection features of the corresponding layer of the encoder. The classification head maps the number of channels output by the last layer of the decoder to the number of classes through a 1×1 convolutional layer, and outputs an unnormalized logits tensor.

[0014] Preferably, it also includes model training, specifically:

[0015] Seismic waveform data with expert-annotated P-wave and S-wave arrival times were collected as training and validation sets. The training data were preprocessed and soft labels were generated according to steps 1 and 2.

[0016] The logits tensor output by the model is normalized by log_softmax in the channel dimension to obtain the log probability. The log probability and the cross-entropy between the soft cross-entropy loss function are calculated using the time-by-time cross-entropy L=-Σ(y_true×log(y_pred)), where the summation iterates through all time points, channels and samples, and different weight coefficients are applied to the three categories to balance the distribution of category samples.

[0017] The Adam optimizer and backpropagation algorithm are used to update all model parameters end-to-end. The ReduceLROnPlateau learning rate scheduling strategy is adopted. When the validation set loss does not decrease for three consecutive epochs, the learning rate is multiplied by the decay factor until the minimum learning rate is reached. During training, the model weights with the minimum validation set loss are saved as the optimal model.

[0018] After each epoch during training, the current model is used to infer the validation set. The prediction confidence and time error relative to the true value of all validation samples are collected. The model is traversed on a preset threshold grid. For each candidate threshold, the precision, recall, and F1 score of the P-wave and S-wave are calculated. A true positive is defined as a prediction confidence ≥ threshold and a time error ≤ tolerance. A false positive is defined as a prediction confidence ≥ threshold but a time error > tolerance or no corresponding true value for the sample. A false negative is defined as a sample with a true value but a prediction confidence < threshold. The threshold that maximizes the F1 score for the P-wave and S-wave is selected as the optimal threshold and saved to the checkpoint file along with the model weights for use in the inference phase.

[0019] Preferably, the training phase determines the optimal thresholds for the P-wave and S-wave respectively through grid search, specifically as follows:

[0020] On the validation set, for P-waves and S-waves, candidate probability thresholds are traversed independently within a predetermined range with a preset step size. For each candidate threshold, the corresponding recall and F1 score are calculated. Finally, the thresholds that make the recall and F1 score optimal are selected for P-waves and S-waves respectively as their optimal probability thresholds.

[0021] Preferably, the multi-scale depth-separable convolution includes two operations: depthwise convolution and pointwise convolution. Multi-scale feature extraction is performed by using multiple different depthwise separable convolutions in parallel.

[0022] Preferably, the depthwise convolution performs a one-dimensional convolution independently on each channel of the input feature with a kernel size of k, the groups parameter is set to the number of input channels, the number of output channels is the same as the number of input channels, and is batch normalized and ReLU activated. The pointwise convolution uses a 1×1 kernel to map the depthwise convolution output to the target number of channels, and is batch normalized, ReLU activated, and then Dropout regularized.

[0023] Preferably, the center alignment specifically refers to:

[0024] When the time length T_skip of the skip connection feature is greater than the time length T_up of the upsampled feature, the central region of length T_up is clipped from the center of the skip connection feature to both sides.

[0025] When the time length T_skip of the skip connection feature is less than the time length T_up of the upsampled feature, zero values ​​are symmetrically padded at both ends of the skip connection feature to make the total length reach T_up.

[0026] Compared with the prior art, the present invention has the following advantages:

[0027] This invention successfully achieves high-precision and robust phase picking of seismic signals, especially S-waves, which are traditionally prone to being missed, by integrating multi-scale depthwise separable convolution, a two-layer CBAM attention mechanism, and dynamic threshold adaptive calibration. The multi-scale depthwise separable convolution, through parallel connection of depthwise separable convolution branches with different kernels (3 / 7 / 15), captures multi-scale temporal features of short-term pulses, medium-term phases, and long-term trends while reducing parameter and computational load. The two-layer CBAM attention mechanism introduces CBAM1D after skip connections and fusion at each layer of the bottleneck layer and decoder. Through the cascading of channel attention and temporal attention, it adaptively enhances the feature channels and arrival time candidate regions related to P-waves and S-waves, suppressing background noise interference. The dynamic threshold adaptive calibration, through validation set grid search, determines the optimal threshold that maximizes the F1 score for P-waves and S-waves respectively, replacing the fixed threshold of 0.5, thus improving generalization ability and picking accuracy under different datasets and noise conditions.

[0028] This invention achieves robust phase picking of seismic signals, significantly improves the F1 score of P-wave and S-wave picking on datasets, and reduces system response delay, providing reliable technical support for mine safety monitoring and earthquake early warning applications. Attached Figure Description

[0029] Figure 1 This is a flowchart of the method;

[0030] Figure 2 Detailed diagram of the core modules of the seismic signal phase picking model;

[0031] Figure 3 For training loss curves;

[0032] Figure 4 This is a plot of the phase picking results for earthquake data (showing a comparison between the P / S wave probability curve and the actual picking location). Detailed Implementation

[0033] To further illustrate the technical solution of the present invention, the present invention will be further described below through embodiments.

[0034] like Figure 1 As shown in the figure, a seismic signal phase picking method based on multi-scale deep learning and attention mechanism in this embodiment includes the following steps:

[0035] Step 1, Data Preprocessing: Obtain three-component seismic waveform data with shape [T, 3], where T is the number of time sampling points and 3 corresponds to the Z / N / E components. Perform zero-mean normalization on each component, i.e., subtract the mean of the component and divide by the standard deviation. Prune or slide window the continuous waveform according to a fixed time window (default 3000 sampling points, corresponding to 30 seconds at a sampling rate of 100Hz). Convert the preprocessed waveform data into tensor shape [B, 3, T], where B is the batch size.

[0036] Step 2: Generate soft labels based on the arrival times of P-waves and S-waves (indicated by sample indexes) marked by seismologists. Generate a P-wave probability curve using a Gaussian window function (standard deviation corresponding to 0.1 seconds) centered on the P-wave arrival time, and a S-wave probability curve using the same parameters centered on the S-wave arrival time. The background noise probability is defined as 1 minus the sum of the probabilities of P-waves and S-waves, ultimately forming a three-channel soft label tensor [B, 3, T], where the sum of the three probabilities at each time point is 1.

[0037] Step 3: Input the preprocessed seismic waveform data into the trained seismic signal phase picking model, which adopts a multi-scale one-dimensional U-Net architecture. Figure 2The encoder processes waveforms layer by layer through multiple EncoderBlocks. Each layer first uses multi-scale depthwise separable convolutions (with kernel sizes of 3 / 7 / 15 in parallel) to extract waveform features at different time scales. Then, MaxPool1D downsampling is used to compress the time dimension to 1 / 4 of the original. At the same time, the number of channels is doubled layer by layer, and the feature maps of each layer are temporarily stored for subsequent skip connections. The bottleneck layer first uses multi-scale depthwise separable convolutions to double the number of channels, and then uses the CBAM1D attention module to enhance key information in two steps. First, channel attention is used to learn channel weights through global pooling and shared MLP, and then time attention is used to learn channel weights. Inter-channel attention focuses the P-wave and S-wave in the time dimension to the time candidate region through channel-dimensional pooling and one-dimensional convolution. The decoder recovers the time resolution by upsampling layer by layer through multiple DecoderBlocks. Each layer first uses transposed convolution to upsample and amplify the time dimension to 4 times and halve the number of channels. Then, it fuses the high-resolution features of the corresponding layer of the encoder after center alignment through skip connections. Next, the CBAM1D attention module is used to further enhance the P-wave and S-wave features. Then, multi-scale depth separable convolution is used to refine the fused features. The classification head maps the decoder output to a three-channel logits tensor [B, 3, T_out] through 1×1 convolution. The output logits are normalized by softmax in the channel dimension to obtain the temporal probability maps of P-wave, S-wave and background noise.

[0038] Step 4: Invoke the optimal thresholds determined for P-waves and S-waves through grid search during the training phase, and judge the probability map point by point. When the probability of P-wave exceeds the optimal threshold of P-wave, mark the time point as a candidate for P-wave arrival. When the probability of S-wave exceeds the optimal threshold of S-wave, mark the time point as a candidate for S-wave arrival. For fixed window data, directly select the candidate position with the highest probability as the final arrival time of P-wave and S-wave. For continuous data stream, apply the peak detection algorithm to the candidate time points, requiring that the interval between adjacent peaks is greater than the minimum interval. Select the peak position with the highest probability as the final arrival time of P-wave and S-wave. Output the picking results of P-wave and S-wave, including the arrival timestamp and the corresponding confidence level.

[0039] The seismic signal phase picking model in this embodiment is as follows: The encoder includes multiple EncoderBlocks connected in series (default depth is 4 layers). Each EncoderBlock contains a multi-scale depth separable convolutional module and a max pooling layer. The multi-scale depth separable convolutional module uses three parallel branches with different convolutional kernel sizes (3, 7, 15). Each branch first performs depth convolution (convolutional kernel sizes are 3 / 7 / 15, and the groups parameter is equal to the number of input channels) and then performs point convolution (1×1 convolutional kernel) to map the number of channels to the target value. The outputs of the three branches are concatenated in the channel dimension and then fused by 1×1 convolution. The pooling kernel size and stride of the max pooling layer are both 4, which compresses the time dimension to 1 / 4 of the original. The number of output channels of each layer increases according to the formula C_out=base_ch×2^d, where d is the layer depth index, base_ch is the number of base channels, and the feature map before downsampling is saved as skip connection features.

[0040] The bottleneck layer first doubles the number of input channels through multi-scale depthwise separable convolution, and then enhances them through the CBAM1D attention module. The channel attention submodule of the CBAM1D attention module performs global average pooling and global max pooling on the feature map in the time dimension. The two pooling results are fed into a shared MLP (structure C→C / r→C, where r is the compression ratio, default 4). The outputs of the two MLPs are added together and activated by Sigmoid to obtain channel weights, which are then multiplied element-wise with the original feature map to achieve channel weighting. The time attention submodule of the CBAM1D attention module performs average pooling and max pooling on the channel-weighted feature map in the channel dimension. The two results are concatenated in the channel dimension and fed into a one-dimensional convolutional layer (convolution kernel size 7). The output is activated by Sigmoid to obtain time weights, which are then multiplied element-wise with the feature map to achieve time weighting.

[0041] The decoder consists of multiple DecoderBlocks concatenated (depth symmetrical to the encoder). Each DecoderBlock contains a transposed convolutional upsampling layer (with a kernel size and stride of 4, expanding the temporal dimension by 4 times while reducing the number of channels to match the corresponding layer in the encoder), skip connection fusion (center-aligning and plucking or padding the skip connection features of the corresponding layer in the encoder to make their temporal dimension consistent with the upsampled features, and then concatenating them in the channel dimension), a CBAM1D attention module (applying channel attention and temporal attention enhancement to the concatenated features, with a channel compression ratio of 4 and a temporal convolution kernel size of 5), and a multi-scale depth-separable convolution (reducing the number of channels from the sum of the concatenated features to the target number of channels to achieve feature refinement). The classification head maps the number of channels output by the last layer of the decoder to the number of categories (background, P-wave, S-wave) through a 1×1 convolutional layer, and outputs an unnormalized logits tensor.

[0042] This method also includes model training, specifically: 1) Data preparation: Collect seismic waveform data with expert annotations for P-wave and S-wave arrival times as training and validation sets. Preprocess the training data and generate soft labels according to steps 1 and 2; 2) Loss function calculation: Normalize the logits tensor of the model output by performing log_softmax normalization in the channel dimension to obtain the log probability. Calculate the time-point cross-entropy L=-Σ(y_true×log(y_pred)) between the log probability and the soft label using the soft cross-entropy loss function. The summation iterates through all time points, channels, and samples. Apply different weight coefficients to the three categories (background, P-wave, S-wave) to balance the category sample distribution; 3) Optimization training: Update all model parameters end-to-end using the Adam optimizer and backpropagation algorithm. Set the initial learning rate to 5×10^-4 and the batch size to 512. Use the ReduceLROnPlateau learning rate scheduling strategy. When the validation set loss does not decrease for three consecutive epochs, multiply the learning rate by a decay factor. 0.5 until the minimum learning rate is reached. During training, the model weights with the minimum loss on the validation set are saved as the optimal model. 4) Dynamic threshold calibration: After each epoch during training, the current model is used to infer the validation set. The prediction confidence (maximum probability of each channel) and time error relative to the true value of all validation samples are collected. The model is traversed on the preset threshold grid (0.10 to 0.95 with a step size of 0.05). For each candidate threshold, the precision, recall and F1 score of the P wave and S wave are calculated. The true positive (TP) is defined as the prediction confidence ≥ the threshold and the time error ≤ the tolerance (0.1 seconds at a sampling rate of 100Hz for 10 sampling points). The false positive (FP) is defined as the prediction confidence ≥ the threshold but the time error > the tolerance or the sample has no corresponding true value. The false negative (FN) is defined as the sample has a true value but the prediction confidence < the threshold. The thresholds that maximize the F1 score for the P wave and S wave are selected as the optimal thresholds and the optimal thresholds are saved to the checkpoint file along with the model weights for use in the inference phase.

[0043] During the training phase, the optimal thresholds for P-waves and S-waves are determined through grid search. Specifically, on the validation set, candidate probability thresholds for P-waves and S-waves are traversed independently within a predetermined range (0.10~0.95) with a preset step size. For each candidate threshold, the corresponding recall and F1 score are calculated. Finally, the thresholds that maximize both recall and F1 score for P-waves and S-waves are selected as their respective optimal probability thresholds.

[0044] This embodiment of multi-scale depthwise separable convolution includes two operations: depthwise convolution and pointwise convolution. By using multiple different depthwise separable convolutions in parallel for multi-scale feature extraction, this design significantly reduces the number of model parameters and computational complexity, thereby increasing the model processing speed to a maximum of approximately 6927.1 records / minute. This provides a foundation for real-time earthquake early warning deployment on embedded devices with limited computing resources. Depthwise convolution independently performs a one-dimensional convolution on each channel of the input features with a kernel size of k (k∈{3,7,15}), and the `groups` parameter is set to the number of input channels. The number of output channels is the same as the number of input channels, and batch normalization and ReLU activation are applied. Pointwise convolution uses a 1×1 kernel to map the depthwise convolution output to the target number of channels, and batch normalization, ReLU activation, and Dropout regularization are applied. Compared to standard convolution, the number of parameters in depthwise separable convolution is reduced from C_in×C_out×k to approximately C_in×k+C_in×C_out, significantly reducing computational complexity and the number of parameters.

[0045] The center alignment mechanism works as follows: when the time length T_skip of the skip connection feature is greater than the time length T_up of the upsampled feature, a central region of length T_up is trimmed from the center of the skip connection feature outwards; when the time length T_skip of the skip connection feature is less than the time length T_up of the upsampled feature, zero values ​​are symmetrically padded at both ends of the skip connection feature to make the total length T_up. This mechanism solves the time dimension mismatch problem caused by rounding errors from multiple downsampling and upsampling, ensuring that the skip connection feature and the upsampled feature can be accurately concatenated.

[0046] The core breakthrough of this method lies in the qualitative leap in recall: compared to the baseline PhaseNet model, the recall of S-waves significantly improved from 0.9092 to 0.9971, thereby reducing the risk of missed detections to an extremely low level and fundamentally enhancing the reliability of the monitoring system in complex noise environments. This design, which trades controllable accuracy fine-tuning (S-wave accuracy adjusted from 0.9758 to 0.9589) for a significant improvement in recall, precisely meets the rigid requirement of "better to have false alarms than missed alarms" in earthquake early warning scenarios, demonstrating clear engineering rationality and advancement. Ultimately, in the key comprehensive evaluation index F1 score, S-waves improved from 0.9413 to 0.9777, and P-waves also maintained a leading position, confirming the substantial improvement in the overall performance of the model. At the same time, the introduction of depthwise separable convolution effectively reduced model complexity, making real-time deployment on embedded devices possible, thus forming a highly efficient solution with a closed loop from algorithm innovation to engineering application. Specific performance comparisons and ablation experiment results are shown in the table below:

[0047] Table 1 Ablation Experiment Results

[0048] name p_prec p_rec p_f1 s_prec s_rec s_f1 Baseline (phasenet) 0.9874 0.9465 0.9665 0.9758 0.9092 0.9413 multiscale_baseline 0.9910 0.9717 0.9813 0.9888 0.8634 0.9219 multiscale_separable+cbam 0.9955 0.9164 0.9543 0.9827 0.8744 0.9254 multiscale_separable+cbam+thresh 0.9759 0.9910 0.9834 0.9589 0.9971 0.9777

[0049] Table 2 Comparison Results

[0050] name p_prec p_rec p_f1 s_prec s_rec s_f1 BaselineTCN 0.9771 0.9217 0.9486 0.9642 0.9207 0.9419 BaselineResASPP 0.9756 0.9023 0.9376 0.9688 0.6989 0.8120 BaselineInception 0.9762 0.9061 0.9399 0.9539 0.9127 0.9328

[0051] As can be seen from Tables 1 and 2, the present invention has achieved a breakthrough in S-wave recall by using a multi-scale depth-separable convolution + CBAM + dynamic threshold scheme, and has improved its F1 score, which is significantly better than the baseline model.

[0052] The foregoing has shown and described the main features and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.

[0053] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A seismic signal phase picking method based on multi-scale deep learning and attention mechanism, characterized in that, Includes the following steps: Step 1: Obtain three-component seismic waveform data in shape [T, 3], where T is the number of time sampling points and 3 corresponds to the three components Z / N / E. Perform zero-mean normalization on each component, i.e., subtract the mean of the component and divide by the standard deviation. Prune or slide window the continuous waveform according to a fixed time window and convert the preprocessed waveform data into tensor shape [B, 3, T], where B is the batch size. Step 2: Generate soft tags based on the arrival times of P-waves and S-waves marked by seismologists. Generate a probability curve for P-waves using a Gaussian window function centered on the arrival time of P-waves, and a probability curve for S-waves using the same parameters centered on the arrival time of S-waves. The background noise probability is defined as 1 minus the sum of the probabilities of P-waves and S-waves, ultimately forming a three-channel soft tag tensor [B, 3, T], where the sum of the three probabilities at each time point is 1. Step 3: Input the preprocessed seismic waveform data into the pre-trained seismic signal phase picking model. The model adopts a multi-scale one-dimensional U-Net architecture. The encoder processes the data layer by layer through multiple EncoderBlocks. Each layer first extracts waveform features at different time scales using multi-scale depth-separable convolution, and then compresses the time dimension to 1 / 4 of the original size using MaxPool1D downsampling. At the same time, the number of channels is doubled layer by layer, and the feature map of each layer is temporarily stored for subsequent skip connections. The bottleneck layer first doubles the number of channels using multi-scale depth-separable convolution, and then strengthens key information in two steps using the CBAM1D attention module. First, channel attention is used to learn channel weights through global pooling and shared MLP, and then time attention is used to focus the P-wave and S-wave in the time dimension to the time candidate region through channel-dimensional pooling and one-dimensional convolution. The decoder recovers the temporal resolution by upsampling layer by layer through multiple DecoderBlocks. Each layer first uses transposed convolution to upsample and amplify the temporal dimension to 4 times and halve the number of channels. Then, it fuses the high-resolution features of the corresponding layer of the encoder after center alignment through skip connections. Next, the CBAM1D attention module is used to further enhance the P-wave and S-wave features. Then, multi-scale depth separable convolution is used to refine the fused features. The classification head maps the decoder output to a three-channel logits tensor through 1×1 convolution. The output logits are normalized by softmax in the channel dimension to obtain the temporal probability maps of P-wave, S-wave and background noise. Step 4: Invoke the optimal thresholds determined for P-waves and S-waves through grid search during the training phase, and judge the probability map point by point. When the probability of P-wave exceeds the optimal threshold of P-wave, mark the time point as a candidate for P-wave arrival. When the probability of S-wave exceeds the optimal threshold of S-wave, mark the time point as a candidate for S-wave arrival. For fixed window data, directly select the candidate position with the highest probability as the final arrival time of P-wave and S-wave. For continuous data stream, apply the peak detection algorithm to the candidate time points, requiring that the interval between adjacent peaks is greater than the minimum interval. Select the peak position with the highest probability as the final arrival time of P-wave and S-wave. Output the picking results of P-wave and S-wave, including the arrival timestamp and the corresponding confidence level.

2. The seismic signal phase picking method based on multi-scale deep learning and attention mechanism according to claim 1, characterized in that, The specific seismic signal phase picking model is as follows: The encoder consists of multiple EncoderBlocks cascaded together. Each EncoderBlock contains a multi-scale depthwise separable convolutional module and a max pooling layer. The multi-scale depthwise separable convolutional module uses three parallel branches with convolutional kernels of different sizes. Each branch first performs depthwise convolution and then pointwise convolution to map the number of channels to the target value. The outputs of the three branches are concatenated in the channel dimension and then fused by a 1×1 convolution. The max pooling layer compresses the time dimension to 1 / 4 of the original. The number of output channels of each layer increases according to the formula C_out=base_ch×2^d, where d is the layer depth index and base_ch is the number of base channels. The feature map before downsampling is saved as skip connection features. The bottleneck layer first doubles the number of input channels through multi-scale depthwise separable convolution, and then enhances them through the CBAM1D attention module. The channel attention submodule of the CBAM1D attention module performs global average pooling and global max pooling on the feature map in the time dimension. The two pooling results are fed into a shared MLP. The outputs of the two MLPs are added and activated by Sigmoid to obtain channel weights, which are then multiplied element-wise with the original feature map to achieve channel weighting. The temporal attention submodule of the CBAM1D attention module performs average pooling and max pooling on the channel-weighted feature map in the channel dimension. The two results are concatenated in the channel dimension and fed into a one-dimensional convolutional layer. The output is activated by Sigmoid to obtain temporal weights, which are then multiplied element-wise with the feature map to achieve temporal weighting. The decoder consists of multiple DecoderBlocks concatenated. Each DecoderBlock contains a transposed convolutional upsampling layer, skip connection fusion, a CBAM1D attention module, and a multi-scale depth-separable convolution. The skip connection fusion centers the skip connection features of the corresponding layer of the encoder. The classification head maps the number of channels output by the last layer of the decoder to the number of classes through a 1×1 convolutional layer, and outputs an unnormalized logits tensor.

3. The seismic signal phase picking method based on multi-scale deep learning and attention mechanism according to claim 2, characterized in that, This also includes model training, specifically: Seismic waveform data with expert-annotated P-wave and S-wave arrival times were collected as training and validation sets. The training data were preprocessed and soft labels were generated according to steps 1 and 2. The logits tensor output by the model is normalized by log_softmax in the channel dimension to obtain the log probability. The log probability and the cross-entropy between the soft cross-entropy loss function are calculated using the time-by-time cross-entropy L=-Σ(y_true×log(y_pred)), where the summation iterates through all time points, channels and samples, and different weight coefficients are applied to the three categories to balance the distribution of category samples. The Adam optimizer and backpropagation algorithm are used to update all model parameters end-to-end. The ReduceLROnPlateau learning rate scheduling strategy is adopted. When the validation set loss does not decrease for three consecutive epochs, the learning rate is multiplied by the decay factor until the minimum learning rate is reached. During training, the model weights with the minimum validation set loss are saved as the optimal model. After each epoch during training, the current model is used to infer the validation set. The prediction confidence and time error relative to the true value of all validation samples are collected. The model is traversed on a preset threshold grid. For each candidate threshold, the precision, recall, and F1 score of the P-wave and S-wave are calculated. A true positive is defined as a prediction confidence ≥ threshold and a time error ≤ tolerance. A false positive is defined as a prediction confidence ≥ threshold but a time error > tolerance or no corresponding true value for the sample. A false negative is defined as a sample with a true value but a prediction confidence < threshold. The threshold that maximizes the F1 score for the P-wave and S-wave is selected as the optimal threshold and saved to the checkpoint file along with the model weights for use in the inference phase.

4. The seismic signal phase picking method based on multi-scale deep learning and attention mechanism according to claim 3, characterized in that, The training phase determines the optimal thresholds for P-waves and S-waves respectively through grid search, specifically as follows: On the validation set, for P-waves and S-waves, candidate probability thresholds are traversed independently within a predetermined range with a preset step size. For each candidate threshold, the corresponding recall and F1 score are calculated. Finally, the thresholds that make the recall and F1 score optimal are selected for P-waves and S-waves respectively as their optimal probability thresholds.

5. The seismic signal phase picking method based on multi-scale deep learning and attention mechanism according to claim 4, characterized in that, The multi-scale depthwise separable convolution includes two operations: depthwise convolution and pointwise convolution. Multi-scale feature extraction is performed by using multiple different depthwise separable convolutions in parallel.

6. The seismic signal phase picking method based on multi-scale deep learning and attention mechanism according to claim 5, characterized in that, The depthwise convolution performs a one-dimensional convolution independently on each channel of the input feature with a kernel size of k, the groups parameter is set to the number of input channels, the number of output channels is the same as the number of input channels, and is batch normalized and ReLU activated. The pointwise convolution uses a 1×1 kernel to map the depthwise convolution output to the target number of channels, and is batch normalized, ReLU activated, and then Dropout regularized.

7. The seismic signal phase picking method based on multi-scale deep learning and attention mechanism according to claim 6, characterized in that, The center alignment specifically refers to: When the time length T_skip of the skip connection feature is greater than the time length T_up of the upsampled feature, the central region of length T_up is clipped from the center of the skip connection feature to both sides. When the time length T_skip of the skip connection feature is less than the time length T_up of the upsampled feature, zero values ​​are symmetrically padded at both ends of the skip connection feature to make the total length reach T_up.

Images