A multi-stage feature-enhanced seismic phase picking method and system

Abstract

The application discloses a multi-stage feature enhancement seismic phase picking method and system, comprising: acquiring three-component seismic waveform data, carrying out initial feature mapping through a shallow embedding network to obtain initial embedding features; carrying out multi-stage feature extraction through an encoder network to obtain multiple stage features corresponding to different semantic levels; carrying out selective enhancement under signal-to-noise guidance according to the multiple stage features to generate signal-to-noise modulation features; carrying out feature calibration and global context modeling on the signal-to-noise modulation features to obtain enhanced features of each stage; then carrying out resolution recovery and cross-layer fusion through a decoding network to obtain fusion features; carrying out point-by-point classification through a classifier according to the fusion features to output a probability sequence representing different seismic phase categories and determine seismic phase arrival time. The application realizes high-precision and high-robustness end-to-end automatic picking of microseismic P waves and S waves through the cooperation of intelligent enhancement and differential calibration modeling guided by signal-to-noise.

Description

Technical Field

[0001] This invention belongs to the field of seismic phase recognition technology, and particularly relates to a multi-stage feature-enhanced seismic phase picking method and system. Background Technology

[0002] Traditional phase picking methods mainly include the long-short time window energy ratio method and autoregressive analysis method. These methods perform well under high signal-to-noise ratio conditions, but their accuracy decreases significantly in low signal-to-noise ratio environments.

[0003] In recent years, deep learning-based methods have made groundbreaking progress in the field of seismic signal processing. Models such as PhaseNet and EQTransformer have demonstrated excellent performance on large-scale seismic datasets. However, existing deep learning methods still have the following limitations when processing microseismic signals: P-waves and S-waves in microseismic events exhibit significant differences in both the time and frequency domains: P-waves have higher frequencies (typically 10-50 Hz), shorter durations, and sharp initial motions; S-waves have lower frequencies (typically 5-30 Hz), larger amplitudes, and longer durations. Most existing methods only extract features in the time domain, failing to fully utilize these time-frequency differences for targeted modeling. While P-waves and S-waves have fundamentally different physical characteristics, existing methods typically employ a uniform feature extraction and attention mechanism, failing to design differentiated recognition strategies for the characteristics of different seismic phases, resulting in limited recognition accuracy under complex waveform conditions.

[0004] In actual microseismic monitoring, the signal-to-noise ratio varies greatly. Existing models are usually trained under fixed signal-to-noise ratio conditions and lack the ability to adaptively adjust to different signal-to-noise ratio conditions.

[0005] Therefore, designing an intelligent acquisition method that can deeply integrate prior knowledge in the time and frequency domains, implement phase differentiation processing, and possess signal-to-noise ratio adaptive capability has become the key to improving the accuracy and reliability of microseismic monitoring. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a multi-stage feature-enhanced seismic phase picking method, comprising the following steps:

[0007] Acquire three-component seismic waveform data;

[0008] The three-component seismic waveform data are initially feature-mapped through a shallow embedding network to obtain initial embedding features;

[0009] Based on the initial embedded features, multi-stage feature extraction is performed through an encoder network containing multiple cascaded stages to obtain multiple stage features corresponding to different semantic levels.

[0010] Based on the multiple stage features, selective enhancement guided by signal-to-noise ratio is performed to generate signal-to-noise modulation features. The selective enhancement includes generating guiding weights based on the energy distribution of each stage feature in the time series, and using the guiding weights to spatially modulate the original stage features.

[0011] Feature calibration and global context modeling are performed on the signal-to-noise modulation features to obtain the enhanced features at each stage;

[0012] The enhanced features from each stage are then processed through a decoding network for resolution restoration and cross-layer fusion to obtain fused features.

[0013] Based on the fusion features, a classifier is used to classify the seismic phases point by point, outputting a probability sequence representing different seismic phase categories, and the arrival time of the seismic phase is determined based on the probability sequence.

[0014] Optionally, the three-component seismic waveform data is initially feature-mapped through a shallow embedding network to obtain initial embedding features, specifically including:

[0015] The three-component seismic waveform data are convolved using a one-dimensional convolution kernel with a predetermined size and stride, mapping the number of input channels to a predetermined embedding dimension.

[0016] The convolution results are sequentially batch normalized and activated by a linear rectified function to obtain the initial embedded features.

[0017] Optionally, based on the initial embedded features, multi-stage feature extraction is performed through an encoder network containing multiple cascaded stages to obtain multiple stage features corresponding to different semantic levels, specifically including:

[0018] The initial embedded features are input into a backbone network consisting of at least three cascaded stages;

[0019] Each stage extracts features at different scales through a cascaded operation that includes downsampling convolution and feature extraction convolution. The spatial resolution decreases progressively with each stage, while the number of output channels increases progressively with each stage.

[0020] The first stage outputs shallow local features, the second stage outputs mid-level semantic features, and the third stage outputs deep abstract features.

[0021] Optionally, based on the multiple stage features, selective enhancement guided by signal-to-noise ratio is performed to generate signal-to-noise modulation features, specifically including:

[0022] For the features of each stage, the average value is calculated by squaring the feature values ​​at each time position along its channel dimension to obtain the feature energy distribution sequence in the time dimension.

[0023] The feature energy distribution sequence is subjected to short-window convolution and long-window convolution respectively. The result of the short-window convolution is divided by the sum of the result of the long-window convolution and a very small constant value to obtain the instantaneous saliency sequence.

[0024] The instantaneous saliency sequence is convolved and normalized using the Sigmoid function to generate a guiding weight sequence with values ​​between zero and one.

[0025] The guiding weight sequence is extended along the channel dimension, multiplied by a learnable modulation intensity parameter, and then multiplied by one. Finally, it is multiplied element-wise with the original stage features to obtain the signal-to-noise modulation features.

[0026] Optionally, feature calibration and global context modeling are performed on the signal-to-noise modulation features to obtain enhanced features at each stage, specifically including:

[0027] The signal-to-noise modulation features of the first stage are processed sequentially through the signal-to-noise guidance unit, the feature compensation module, and the linear attention mechanism module, wherein the feedforward network in the linear attention mechanism module adopts the first expansion factor;

[0028] For the signal-to-noise modulation features in the second stage, they are processed sequentially through the signal-to-noise guidance unit, the feature compensation module, and the linear attention mechanism module, wherein the feedforward network in the linear attention mechanism module adopts a second expansion factor that is smaller than the first expansion factor;

[0029] The signal-to-noise modulation characteristics of the third stage are processed sequentially by the signal-to-noise guiding unit and the feature compensation module.

[0030] Optionally, the feature compensation module performs feature calibration, specifically including:

[0031] The input features are first subjected to a first-dimensional convolution operation to compress the number of channels, and then a second-dimensional convolution operation to restore the number of channels, to obtain the reference features.

[0032] The element-wise product of the reference feature and the input feature is calculated and then passed through a third one-dimensional convolution to obtain the multiplicative calibration feature;

[0033] The difference between the input feature and the reference feature is calculated and then passed through a fourth one-dimensional convolution to obtain the differential calibration feature;

[0034] The multiplicative calibration feature and the differential calibration feature are concatenated, and then after a fifth one-dimensional convolution, the residuals are added to the input feature to output the calibrated feature.

[0035] Optionally, the enhanced features from each stage are used to restore resolution and fuse across layers through a decoding network to obtain fused features, specifically including:

[0036] The deepest enhanced features are used as the starting decoding features;

[0037] From deep to shallow, convolution and upsampling operations are performed on the current decoded features in sequence, and the upsampling results are added to the residuals of the enhanced features of the corresponding stage to obtain the decoded features of the previous stage.

[0038] Repeat this process until the scale of the initial embedded features is restored, and then add the final decoded features to the initial embedded features to obtain the fused features.

[0039] Optionally, based on the fused features, a classifier performs point-by-point classification, outputting a probability sequence representing different seismic phase categories, and determines the arrival time of the seismic phase based on the probability sequence, specifically including:

[0040] The fused features are convolutionally processed to adjust the number of channels, and then activated by the Softmax function to obtain the probability that each time point belongs to the P-wave, S-wave, or noise category.

[0041] The probability sequences of P-waves and S-waves are filtered by applying probability thresholds respectively;

[0042] Find local peaks in the filtered probability sequence, take the time point corresponding to the peak as the arrival time of the phase, and output the peak probability as the confidence level.

[0043] On the other hand, the present invention also provides a multi-stage feature-enhanced seismic phase picking system for implementing the method, comprising:

[0044] The data acquisition module is used to acquire and preprocess three-component seismic waveform data;

[0045] The feature extraction module, connected to the data acquisition module, includes a shallow embedding network and a multi-stage encoder network, used to perform initial feature mapping and multi-scale feature extraction on the three-component seismic waveform data, and output features at multiple stages.

[0046] The signal-to-noise guidance module, connected to the feature extraction module, is used to estimate the signal-to-noise distribution of features at each stage and generate guidance weights, selectively spatially modulate the features at the multiple stages, and output signal-to-noise modulation features.

[0047] The feature calibration module, connected to the signal-to-noise guidance module, is used to perform channel calibration and global context modeling on the signal-to-noise modulation features according to their stage using a differentiated configuration strategy, and output the enhanced features.

[0048] The decoding and fusion module, connected to the feature calibration module, is used to restore the resolution and fuse the enhanced features at each stage, and output the fused features.

[0049] The detection output module is connected to the decoding and fusion module and is used to perform phase classification based on the fusion features and output the arrival times of P-waves and S-waves.

[0050] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0051] Compared with the prior art, the present invention has the following advantages and technical effects:

[0052] Traditional methods and some existing deep learning models are sensitive to noise, and their performance drops sharply when the signal-to-noise ratio is low. This invention introduces a "signal-to-noise guided selective enhancement" mechanism, enabling the model to automatically analyze the temporal energy distribution of input features, identify and focus on potential effective signal regions, and suppress interference from noise-dominated regions. This achieves adaptive signal enhancement and significantly improves the ability to extract effective features in complex noise environments.

[0053] P-waves and S-waves differ significantly in frequency, duration, and initial arrival characteristics. Existing methods often employ a uniform feature processing strategy, failing to fully utilize and differentiate these differences. This invention, through a strategy of "differentiated feature calibration and global context modeling," adaptively combines feature calibration and attention mechanisms for different semantic levels of features extracted at different network stages. This provides a targeted processing path for accurately capturing different patterns, such as the sharp initial arrival of P-waves and the sustained waveform of S-waves.

[0054] Complex model structures often lead to high computational costs, and poor coordination among feature extraction, enhancement, and decoding modules can result in information loss or error accumulation. This invention constructs a coherent "multi-stage feature extraction-selective enhancement-differential calibration-decoding fusion" technical chain. In the encoding stage, multi-level, targeted feature optimization is performed; in the decoding stage, resolution is restored through efficient cross-layer fusion; and finally, accurate point-by-point classification is achieved through a classifier. This complete process ensures the effective transmission and utilization of feature information, achieving stable and reliable output from the original three-component waveform data to the final P-wave, S-wave, and time information while maintaining high computational efficiency. Attached Figure Description

[0055] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0056] Figure 1 This is a block diagram of the microseismic phase acquisition system architecture according to an embodiment of the present invention;

[0057] Figure 2 This is a flowchart of the microseismic phase picking method according to an embodiment of the present invention;

[0058] Figure 3 This is a flowchart of the signal-to-noise guided selective enhancement method according to an embodiment of the present invention;

[0059] Figure 4 This is a flowchart of the feature calibration and compensation method according to an embodiment of the present invention;

[0060] Figure 5 This is a schematic diagram of the differentiated configuration strategy in an embodiment of the present invention. Detailed Implementation

[0061] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0062] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0063] Example 1

[0064] This embodiment provides a multi-stage feature enhancement seismic phase picking method. This method constructs an encoder-decoder architecture, and the overall data flow is processed sequentially according to the path of "input preprocessing → multi-stage feature extraction → signal-to-noise guided enhancement → differential feature calibration → decoding fusion → detection output," including the following steps:

[0065] Acquire three-component seismic waveform data;

[0066] The three-component seismic waveform data are initially feature-mapped through a shallow embedding network to obtain initial embedding features;

[0067] Based on the initial embedded features, multi-stage feature extraction is performed through an encoder network containing multiple cascaded stages to obtain multiple stage features corresponding to different semantic levels.

[0068] Based on the multiple stage features, selective enhancement guided by signal-to-noise ratio is performed to generate signal-to-noise modulation features. The selective enhancement includes generating guiding weights based on the energy distribution of each stage feature in the time series, and using the guiding weights to spatially modulate the original stage features.

[0069] Feature calibration and global context modeling are performed on the signal-to-noise modulation features to obtain the enhanced features at each stage;

[0070] The enhanced features from each stage are then processed through a decoding network for resolution restoration and cross-layer fusion to obtain fused features.

[0071] Based on the fusion features, a classifier is used to classify the seismic phases point by point, outputting a probability sequence representing different seismic phase categories, and the arrival time of the seismic phase is determined based on the probability sequence.

[0072] Implementably, the three-component seismic waveform data is subjected to initial feature mapping through a shallow embedding network to obtain initial embedding features, specifically including:

[0073] The three-component seismic waveform data are convolved using a one-dimensional convolution kernel with a predetermined size and stride, mapping the number of input channels to a predetermined embedding dimension; the convolution results are then subjected to batch normalization and linear rectified function activation to obtain the initial embedding features.

[0074] Furthermore, based on the initial embedded features, multi-stage feature extraction is performed through an encoder network containing multiple cascaded stages to obtain multiple stage features corresponding to different semantic levels, specifically including:

[0075] The initial embedded features are input into a backbone network consisting of at least three cascaded stages. Each stage extracts features at different scales through cascaded operations including downsampling convolution and feature extraction convolution. The spatial resolution decreases progressively with each stage, while the number of output channels increases progressively with each stage. The first stage outputs shallow local features, the second stage outputs mid-level semantic features, and the third stage outputs deep abstract features.

[0076] As a specific implementation method, for the input three-component microseismic waveform ,in Given the sequence length, initial feature extraction is first performed using a shallow embedding module:

[0077] ;

[0078] The convolution kernel size is 7 and the stride is 2, mapping the 3-channel input to the embedding dimension. The backbone network employs a three-stage cascaded structure for multi-scale feature extraction. Each stage includes a cascade of downsampling convolutions and feature extraction convolutions, with the spatial resolution halved in each stage: the first stage ( The output channel count is 64, extracting shallow local features; the second stage ( The output channel count is 128, and mid-level semantic features are extracted; the third stage ( The output channel count is 256, and deep abstract features are extracted.

[0079] Implementable, based on the aforementioned multiple stage characteristics, selective enhancement guided by signal-to-noise ratio (SNR) is performed to generate SNR modulation features, specifically including:

[0080] For each stage's features, the average value is calculated by squaring the feature values ​​at each time position along its channel dimension, resulting in a feature energy distribution sequence along the time dimension. Short-window convolution and long-window convolution are then performed on the feature energy distribution sequence, and the result of the short-window convolution is divided by the sum of the result of the long-window convolution and a very small constant value to obtain an instantaneous saliency sequence. This instantaneous saliency sequence is then convolved and normalized using the Sigmoid function to generate a guiding weight sequence with values ​​between zero and one. This guiding weight sequence is then expanded along the channel dimension, multiplied by a learnable modulation intensity parameter, incremented by one, and then multiplied element-wise with the original stage features to obtain the signal-to-noise modulation feature.

[0081] As a specific implementation method, this embodiment introduces a signal-noise guided selective enhancement strategy to address the characteristics of low signal-to-noise ratio and small effective signal ratio in microseismic signals. The core idea is to first estimate the signal-to-noise distribution of the feature sequence, generate guiding weights indicating the signal region, and then spatially modulate the feature enhancement process, concentrating the enhancement operation on the signal region while suppressing the response to the noise region.

[0082] Specifically, given the encoder's first... Characteristics of stage output The signal-to-noise guiding unit first calculates the characteristic energy distribution at each time position along the channel dimension:

[0083] ;

[0084] To capture the characteristics of local energy abrupt changes, a learnable energy ratio is used to calculate instantaneous significance:

[0085] ;

[0086] in and These represent the short-time and long-time convolution kernel sizes, respectively. This is a numerically stable term. Instantaneous significance. Guiding weights are generated by Sigmoid normalization. Its value range is [0,1], with high values ​​indicating the location of potential signals. The guiding weights are modulated with the input features:

[0087] ;

[0088] in This is a learnable modulation intensity parameter. This mechanism enhances the characteristics of the signal region while preserving the characteristics of the noise region.

[0089] Implementably, feature calibration and global context modeling are performed on the signal-to-noise modulation features to obtain enhanced features at each stage, specifically including:

[0090] For the signal-to-noise modulation features of the first stage, they are processed sequentially through the signal-to-noise guidance unit, the feature compensation module, and the linear attention mechanism module, wherein the feedforward network in the linear attention mechanism module adopts a first expansion factor; for the signal-to-noise modulation features of the second stage, they are processed sequentially through the signal-to-noise guidance unit, the feature compensation module, and the linear attention mechanism module, wherein the feedforward network in the linear attention mechanism module adopts a second expansion factor smaller than the first expansion factor; for the signal-to-noise modulation features of the third stage, they are processed sequentially through the signal-to-noise guidance unit and the feature compensation module.

[0091] Implementable, the feature compensation module performs feature calibration, specifically including:

[0092] The input features are first subjected to a first one-dimensional convolution operation to compress the number of channels, and then a second one-dimensional convolution operation to restore the number of channels, to obtain reference features. The element-wise product of the reference features and the input features is calculated and then subjected to a third one-dimensional convolution to obtain multiplicative calibration features. The difference between the input features and the reference features is calculated and then subjected to a fourth one-dimensional convolution to obtain differential calibration features. The multiplicative calibration features and the differential calibration features are concatenated, and then subjected to a fifth one-dimensional convolution. Finally, the residuals are added to the input features to output the calibrated features.

[0093] As a specific implementation, the feature compensation module receives the modulated features, generates reference features through a channel compression-recovery mechanism, and then uses multiplicative and differential dual-path methods to capture different variation patterns between features:

[0094] ;

[0095] ;

[0096] The multiplicative path captures the amplification effect of correlations between features, while the differential path captures the details filtered out by the compression process. The calibration features are obtained by concatenating and fusing the two feature paths and then connecting them via residuals. The signal-to-noise guidance unit identifies signal regions from a spatial dimension, while the feature compensation module performs calibration from a channel dimension, forming a spatial-channel collaborative enhancement mechanism.

[0097] Calibrated features Global dependency modeling is further performed using a linear attention mechanism module. This module is implemented based on a state-space model. The sequence processing with low complexity is achieved by capturing global dependencies in long sequences through the following discretized recursive relationship:

[0098] ;

[0099] in For the first Input characteristics at time step For the corresponding output features, It is in a hidden state. , These are the discretized state transition matrix and input matrix. The output projection matrix is ​​used. This embodiment employs a selective mechanism to dynamically generate parameters from the input, enabling the selective transmission or omission of information based on the input content.

[0100] Considering the differences in semantic hierarchy and receptive field of features at different stages, this embodiment adopts a differentiated configuration strategy for the combination of the above-mentioned signal-to-noise guidance unit, feature compensation module, and linear attention module:

[0101] The first stage uses a complete configuration, including all three modules, with the feedforward network expansion factor in the linear attention module set to 2.

[0102] The second stage adopts a simplified configuration, reducing the feedforward network expansion factor to 1 to reduce the number of parameters;

[0103] The third stage adopts a lightweight configuration, which only includes a signal-to-noise guidance unit and a feature compensation module, and does not use a linear attention module.

[0104] This design is based on the consideration that shallow features have high spatial resolution but low semantic level, requiring stronger global modeling capabilities; while deep features have a sufficiently large receptive field, requiring less additional global modeling.

[0105] Feasible approach: Enhanced features from each stage are processed through a decoding network for resolution restoration and cross-layer fusion to obtain fused features, specifically including:

[0106] The deepest enhanced feature is used as the starting decoding feature; from deep to shallow, the current decoding feature is sequentially convolved and upsampled, and the upsampling result is added to the residual of the enhanced feature of the corresponding stage to obtain the decoding feature of the previous stage; this process is repeated until the scale of the initial embedded feature is restored, and the final decoding feature is added to the initial embedded feature to obtain the fused feature.

[0107] Furthermore, based on the fused features, a classifier performs point-by-point classification, outputting probability sequences representing different seismic phase categories, and determines the arrival time of the seismic phase based on the probability sequences, specifically including:

[0108] The fused features are convolved to adjust the number of channels, and then activated by the Softmax function to obtain the probability that each time point belongs to the P-wave, S-wave, or noise category. The probability sequences of P-wave and S-wave are filtered by applying probability thresholds respectively. Local peaks are found in the filtered probability sequences, the time points corresponding to the peaks are taken as the arrival time of the seismic phase, and the peak probability is output as the confidence level.

[0109] As a specific implementation method, the decoding stage employs a lightweight design, recovering feature resolution through layer-by-layer upsampling. Let the calibrated features output by the feature enhancement module in the three stages be respectively... , , Decoding starts from the deepest features and initializes... The decoding module's processing flow is a convolution-upsampling-convolution structure. Decoded features are residually connected with the calibration features from the corresponding stage to achieve cross-layer feature fusion.

[0110] ;

[0111] This design reduces computational overhead through a simplified structure, while utilizing residual connections to directly pass enhanced features from the encoding stage to the decoding path. The decoder ultimately outputs fused features. The detector classifies the data point by point, and the point-by-point classification probability is obtained through Softmax activation.

[0112] ;

[0113] The three channels correspond to the probabilities of P-wave, S-wave, and noise, respectively. Post-processing outputs the phase arrival time and confidence level through peak detection. Model training uses the vector cross-entropy loss function, and the model output is set to... The real label is ,in Given the batch size, the loss function is defined as:

[0114] ;

[0115] The labels use Gaussian soft labels, which construct a Gaussian distribution centered on the actual arrival times of P-waves and S-waves, providing tolerance for picking errors.

[0116] This embodiment addresses the low signal-to-noise ratio (SNR) of microseismic signals by introducing a signal-to-noise guided selective enhancement strategy. It organically integrates the classic STA / LTA detection principle from seismology with a deep learning framework, preserving both physical interpretability and data-driven adaptive capabilities. The signal-to-noise guided unit and feature compensation module form a spatial-channel collaborative enhancement mechanism, which has stronger feature representation capabilities compared to single-dimensional attention mechanisms. By capturing different variation patterns through multiplicative and differential calibration paths, it achieves effective calibration and compensation for multi-scale features. A differentiated configuration strategy effectively controls computational complexity while maintaining model performance. The entire model is trained end-to-end, facilitating rapid deployment and application.

[0117] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0118] Example 2

[0119] like Figure 1 As shown, this embodiment provides a microseismic phase acquisition system based on multi-stage feature enhancement, comprising: a data acquisition module for acquiring and preprocessing three-component seismic waveform data; a feature extraction module connected to the data acquisition module, including a shallow embedding network and a multi-stage encoder network, for performing initial feature mapping and multi-scale feature extraction on the three-component seismic waveform data, and outputting features at multiple stages; a signal-to-noise (SNR) guidance module connected to the feature extraction module for estimating the SNR distribution of features at each stage and generating guidance weights, selectively spatially modulating the features at multiple stages, and outputting SNR modulation features; a feature calibration module connected to the SNR guidance module for performing channel calibration and global context modeling on the SNR modulation features according to their stage using a differentiated configuration strategy, and outputting enhanced features; a decoding and fusion module connected to the feature calibration module for performing resolution restoration and cross-layer fusion of the enhanced features at each stage, and outputting fused features; and a detection output module connected to the decoding and fusion module for performing phase classification based on the fused features and outputting P-wave arrival times and S-wave arrival times. This system can be deployed on edge computing devices or cloud servers and supports real-time streaming processing.

[0120] like Figure 2 As shown, this embodiment provides a microseismic phase picking method based on multi-stage feature enhancement, including the following steps:

[0121] Step 1: Obtain three-component seismic waveforms. The input is three-component waveform data acquired by microseismic monitoring stations. It includes a vertical component Z, a north-south component N, and an east-west component E. In this embodiment, the sampling rate is 100Hz, and the sequence length is... Each sampling point corresponds to a 30-second monitoring window. The input data is preprocessed, including mean removal, bandpass filtering, and peak normalization, before being fed into the model.

[0122] Step 2: Multi-stage feature extraction. First, a shallow embedding module maps the three-component input to a high-dimensional feature space. The shallow embedding module uses a one-dimensional convolution with a kernel size of 7 and a stride of 2 to map the 3-channel input to the embedding dimension. And after batch normalization and ReLU activation:

[0123] ;

[0124] Output embedded features The backbone network employs a three-stage cascaded structure, with each stage consisting of a cascade of downsampling convolutions with a stride of 2 and feature extraction convolutions. The first stage outputs 64 channels with a feature dimension of [missing information]. The second stage output has 128 channels and a feature dimension of [missing information]. The third stage output has 256 channels and a feature dimension of [missing information]. The three stages correspond to shallow local features, mid-level semantic features, and deep abstract features, respectively.

[0125] Step 3: Signal-to-noise guided selective enhancement. For example... Figure 3 As shown, to address the characteristics of low signal-to-noise ratio and small proportion of effective signal in the time series of microseismic signals, this invention introduces a signal-to-noise guided selective enhancement strategy. Given the encoder's... Characteristics of stage output The signal-to-noise guiding unit first calculates the characteristic energy distribution at each time position along the channel dimension:

[0126] ;

[0127] To capture local energy mutation characteristics, the instantaneous significance is calculated using a learnable short-time to long-time energy ratio:

[0128] ;

[0129] in and These represent the kernel sizes for the short-term and long-term windows, respectively. For stable terms, instantaneous significance is used to generate guiding weights through 1×1 convolution and Sigmoid normalization. The value range is [0,1]. The guiding weights are broadcast along the channel dimension and then modulated with the input features:

[0130] ;

[0131] in The modulation intensity parameter is learned and initialized to 0.1 to ensure stability during the initial training phase. This modulation mechanism amplifies the features of the signal region (high saliency) while keeping the features of the noise region (low saliency) unchanged, thus realizing the "localization before enhancement" processing strategy.

[0132] Step 4: Differentiated Feature Calibration. For example... Figure 4 As shown, the feature compensation module receives the modulation features output by the signal-to-noise guiding unit. Reference features are generated through a channel compression-reconstruction mechanism, and then multiplicative and differential dual-path methods are used to capture different variation patterns between features. First, channel compression and reconstruction operations are performed on the modulated features:

[0133] ;

[0134] ;

[0135] This bottleneck structure forces the features to undergo information reorganization, generating reference features containing core semantic information. Dual-path calibration calculations are performed using the reference characteristics and modulation characteristics:

[0136] ;

[0137] ;

[0138] The multiplicative path captures the amplification effect of correlations between features, while the differential path captures the details filtered out by the compression process. The calibrated features are obtained by concatenating and fusing the two feature paths and then connecting them via residuals.

[0139] ;

[0140] The signal-to-noise ratio (SNR) guidance unit identifies signal regions and performs pre-enhancement from a spatial dimension, while the feature compensation module performs feature calibration from a channel dimension. Both optimize the features from different dimensions, forming a spatial-channel collaborative enhancement mechanism to jointly improve the characterization capability of low SNR microseismic signals.

[0141] like Figure 5 As shown, this embodiment employs a differentiated configuration strategy for the three stages. The calibrated features are further modeled globally using a linear attention mechanism module, which is based on a state-space model. The sequence processing complexity is as follows: The first stage processes shallow features using a full configuration, including a signal-to-noise (SNR) guidance unit, a feature compensation module, and a linear attention module. The feedforward network expansion factor of the linear attention module is set to 2. The second stage processes mid-level features using a simplified configuration, with the feedforward network expansion factor reduced to 1. The third stage processes deep features using a lightweight configuration, employing only the SNR guidance unit and the feature compensation module, without using the linear attention module. This differentiated design is based on the following considerations: shallow features have high spatial resolution but low semantic level, requiring a stronger attention mechanism for global context modeling; deep features have high semantic level, and their receptive field already covers a longer time range, so the need for additional global modeling is relatively low. In this embodiment, the linear attention module is implemented using the Mamba architecture based on a selective state space model, where the state dimension is set to 16, the convolutional kernel size is 4, and the expansion factor is 2.

[0142] Step 5: Decoding and Feature Fusion. The decoding stage receives features processed by the feature enhancement module from each stage of the encoder, restores feature resolution through layer-by-layer upsampling, and performs cross-stage feature fusion. Let the enhanced features of the three stages be... , , The decoding module employs a convolution-upsampling-convolution structure, and the decoded features are residually connected with the enhancement features of the corresponding stage.

[0143] ;

[0144] Decoder from deep features Initially, the resolution is restored layer by layer, with the output dimensions being respectively , , The decoded output is added to the shallow embedding features to obtain the fused features. .

[0145] Step Six: Seismic Phase Detection and Output. The detection head fuses the features. For point-by-point classification, a two-stage convolutional structure is used for channel dimensionality reduction and upsampling. The point-by-point classification probability is obtained through Softmax activation.

[0146] ;

[0147] Output The classification results are dimensional, with three channels corresponding to the probabilities of P-wave, S-wave, and noise, respectively. In the post-processing stage, a threshold of 0.3 is applied to filter the P-wave and S-wave probability sequences, and the P-wave arrival time is output through peak detection. When the S wave arrives The corresponding peak probability is used as the confidence level.

[0148] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multi-stage feature-enhanced seismic phase picking method, characterized in that, Includes the following steps: Acquire three-component seismic waveform data; The three-component seismic waveform data are initially feature-mapped through a shallow embedding network to obtain initial embedding features; Based on the initial embedded features, multi-stage feature extraction is performed through an encoder network containing multiple cascaded stages to obtain multiple stage features corresponding to different semantic levels. Based on the multiple stage features, selective enhancement guided by signal-to-noise ratio is performed to generate signal-to-noise modulation features. The selective enhancement includes generating guiding weights based on the energy distribution of each stage feature in the time series, and using the guiding weights to spatially modulate the original stage features. Feature calibration and global context modeling are performed on the signal-to-noise modulation features to obtain the enhanced features at each stage; The enhanced features from each stage are then processed through a decoding network for resolution restoration and cross-layer fusion to obtain fused features. Based on the fusion features, a classifier is used to classify the seismic phases point by point, outputting a probability sequence representing different seismic phase categories, and the arrival time of the seismic phase is determined based on the probability sequence.

2. The method according to claim 1, characterized in that, The three-component seismic waveform data are initially feature-mapped using a shallow embedding network to obtain initial embedding features, specifically including: The three-component seismic waveform data are convolved using a one-dimensional convolution kernel with a predetermined size and stride, mapping the number of input channels to a predetermined embedding dimension. The convolution results are sequentially batch normalized and activated by a linear rectified function to obtain the initial embedded features.

3. The method according to claim 1, characterized in that, Based on the initial embedded features, multi-stage feature extraction is performed through an encoder network containing multiple cascaded stages to obtain multiple stage features corresponding to different semantic levels, specifically including: The initial embedded features are input into a backbone network consisting of at least three cascaded stages; Each stage extracts features at different scales through a cascaded operation that includes downsampling convolution and feature extraction convolution. The spatial resolution decreases progressively with each stage, while the number of output channels increases progressively with each stage. The first stage outputs shallow local features, the second stage outputs mid-level semantic features, and the third stage outputs deep abstract features.

4. The method according to claim 1, characterized in that, Based on the aforementioned multiple stage characteristics, selective enhancement guided by signal-to-noise ratio (SNR) is performed to generate SNR modulation features, specifically including: For the features of each stage, the average value is calculated by squaring the feature values ​​at each time position along its channel dimension to obtain the feature energy distribution sequence in the time dimension. The feature energy distribution sequence is subjected to short-window convolution and long-window convolution respectively. The result of the short-window convolution is divided by the sum of the result of the long-window convolution and a very small constant value to obtain the instantaneous saliency sequence. The instantaneous saliency sequence is convolved and normalized using the Sigmoid function to generate a guiding weight sequence with values ​​between zero and one. The guiding weight sequence is extended along the channel dimension, multiplied by a learnable modulation intensity parameter, and then multiplied by one. Finally, it is multiplied element-wise with the original stage features to obtain the signal-to-noise modulation features.

5. The method according to claim 1, characterized in that, The signal-to-noise modulation features are calibrated and global context modeled to obtain the enhanced features at each stage, specifically including: The signal-to-noise modulation features of the first stage are processed sequentially through the signal-to-noise guidance unit, the feature compensation module, and the linear attention mechanism module, wherein the feedforward network in the linear attention mechanism module adopts the first expansion factor; For the signal-to-noise modulation features in the second stage, they are processed sequentially through the signal-to-noise guidance unit, the feature compensation module, and the linear attention mechanism module, wherein the feedforward network in the linear attention mechanism module adopts a second expansion factor that is smaller than the first expansion factor; The signal-to-noise modulation characteristics of the third stage are processed sequentially by the signal-to-noise guiding unit and the feature compensation module.

6. The method according to claim 5, characterized in that, The feature compensation module performs feature calibration, specifically including: The input features are first subjected to a first-dimensional convolution operation to compress the number of channels, and then a second-dimensional convolution operation to restore the number of channels, to obtain the reference features. The element-wise product of the reference feature and the input feature is calculated and then passed through a third one-dimensional convolution to obtain the multiplicative calibration feature; The difference between the input feature and the reference feature is calculated and then passed through a fourth one-dimensional convolution to obtain the differential calibration feature; The multiplicative calibration feature and the differential calibration feature are concatenated, and then after a fifth one-dimensional convolution, the residuals are added to the input feature to output the calibrated feature.

7. The method according to claim 1, characterized in that, The enhanced features from each stage are then processed through a decoding network for resolution restoration and cross-layer fusion to obtain fused features, specifically including: The deepest enhanced features are used as the starting decoding features; From deep to shallow, convolution and upsampling operations are performed on the current decoded features in sequence, and the upsampling results are added to the residuals of the enhanced features of the corresponding stage to obtain the decoded features of the previous stage. Repeat this process until the scale of the initial embedded features is restored, and then add the final decoded features to the initial embedded features to obtain the fused features.

8. The method according to claim 1, characterized in that, Based on the fused features, a classifier performs point-by-point classification, outputting a probability sequence representing different seismic phase categories, and determining the arrival time of the seismic phase based on the probability sequence, specifically including: The fused features are convolutionally processed to adjust the number of channels, and then activated by the Softmax function to obtain the probability that each time point belongs to the P-wave, S-wave, or noise category. The probability sequences of P-waves and S-waves are filtered by applying probability thresholds respectively; Find local peaks in the filtered probability sequence, take the time point corresponding to the peak as the arrival time of the phase, and output the peak probability as the confidence level.

9. A multi-stage feature-enhanced seismic phase picking system, characterized in that, For implementing the method according to any one of claims 1-8, comprising: The data acquisition module is used to acquire and preprocess three-component seismic waveform data; The feature extraction module, connected to the data acquisition module, includes a shallow embedding network and a multi-stage encoder network, used to perform initial feature mapping and multi-scale feature extraction on the three-component seismic waveform data, and output features at multiple stages. The signal-to-noise guidance module, connected to the feature extraction module, is used to estimate the signal-to-noise distribution of features at each stage and generate guidance weights, selectively spatially modulate the features at the multiple stages, and output signal-to-noise modulation features. The feature calibration module, connected to the signal-to-noise guidance module, is used to perform channel calibration and global context modeling on the signal-to-noise modulation features according to their stage using a differentiated configuration strategy, and output the enhanced features. The decoding and fusion module, connected to the feature calibration module, is used to restore the resolution and fuse the enhanced features at each stage, and output the fused features. The detection output module is connected to the decoding and fusion module and is used to perform phase classification based on the fusion features and output the arrival times of P-waves and S-waves.

10. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method of any one of claims 1-8.

Images