Earthquake disaster-caused analysis method and system driven by near-field P-wave fragment

Abstract

The invention discloses a near-field P-wave fragment-driven earthquake disaster-caused analysis method and system, and the method comprises the steps: extracting the multi-dimensional fusion features of preprocessed seismic waveform data through the connection of a multi-window convolution parallel architecture and a residual error, and enabling the multi-window convolution parallel architecture to employ convolution kernels of different sizes to respectively capture the features of P waves in a specified time period; the multi-dimensional fusion features are input into a parallel integrated encoder, waveform internal time sequence association is obtained through a self-attention mechanism, different weights are adaptively distributed to features in a specified time period, and waveform features are output through global average pooling; and inputting the global features into a multi-channel decoder branch, and outputting earthquake disaster-causing analysis parameters including initial motion polarity, earthquake type, earthquake magnitude, waveform quality and risk level probability. According to the method, the problems of insufficient real-time performance, waveform quality evaluation distortion, poor disaster-causing grading accuracy and poor multi-task collaboration in the prior art are solved.

Description

Technical Field

[0001] This invention relates to the field of earthquake monitoring technology, specifically to a near-field P-wave segment-driven earthquake disaster analysis method and system. Background Technology

[0002] In fields such as earthquake monitoring, explosion monitoring, and mine safety monitoring, the monitoring and early warning of natural earthquakes and non-natural earthquake events such as collapses and industrial blasting are essential for ensuring the safety of people and property and improving disaster prevention and mitigation capabilities. Rapidly acquiring earthquake parameters, accurately identifying event types, and scientifically determining the severity of disasters directly affect the efficiency of emergency response and the effectiveness of disaster prevention decisions, and have significant application value in the fields of earthquake monitoring, geological exploration, and mineral production safety.

[0003] In terms of earthquake parameter acquisition and event identification, current mainstream earthquake monitoring technologies have significant shortcomings. Existing technologies generally rely on complete seismic wave sequences, requiring the completion of full-time recordings of P-waves and S-waves before analysis can begin, resulting in monitoring delays of up to minutes, which is insufficient to meet the real-time requirements of short-term early warning. Some short P-wave analysis techniques suffer from simple architectures, single-parameter outputs, and a lack of pre-assessment of waveform quality, making them prone to parameter estimation errors due to low-quality waveforms, and leading to a high error rate in identifying micro- and intermediate-scale earthquake types.

[0004] Waveform quality assessment is crucial for ensuring parameter accuracy, but existing methods often rely on absolute energy indices to construct judgment systems, which are strongly coupled with energy parameters such as magnitude and epicentral distance, easily leading to quality assessment distortion. Furthermore, fixed threshold rules have poor generalization ability, making it difficult to adapt to differences in instruments at different stations and regional geological media, failing to provide stable quality screening criteria. The influx of low signal-to-noise ratio signals further amplifies parameter estimation errors. Regarding disaster classification and disaster prevention decision-making, existing schemes have significant limitations. Most rely solely on magnitude for classification, ignoring the moderating effects of epicentral distance and geological conditions on disaster impact; the few schemes that incorporate epicentral distance also suffer from coarse classification and lack precise combined judgment logic, resulting in a disconnect from actual disaster prevention needs. Moreover, existing technologies do not link waveform quality with disaster classification, easily triggering invalid warnings, and the output results are mostly abstract numerical values ​​requiring manual secondary calculations to be integrated with disaster prevention decisions, thus lacking practicality.

[0005] In terms of model architecture design, mainstream single-task independent architectures suffer from problems such as high deployment costs, low inference efficiency, and poor consistency of multi-task results. A few multi-task models have not designed dedicated structures for the temporal characteristics of seismic waves, making it difficult to capture key temporal correlation features of P waves. Furthermore, they have not formed a complete technical chain from basic parameters to quality assessment to risk classification. Some related technologies also have limitations such as reliance on long-term waveforms and single-parameter output, which cannot meet the practical needs of early warning of non-natural earthquakes. Summary of the Invention

[0006] To address these issues, this invention provides a near-field P-wave segment-driven earthquake disaster analysis method and system, which solves the problems of insufficient real-time performance, distorted waveform quality assessment, lack of accuracy in disaster classification, and poor multi-task coordination in existing technologies.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a near-field P-wave segment-driven earthquake disaster analysis method, comprising the following steps:

[0008] Receive near-field P-wave seismic waveform data recorded by multiple stations and multiple channels, and perform detrending, bandpass filtering and normalization preprocessing on the waveform data;

[0009] Multi-scale fusion features of the preprocessed waveform data are extracted using a multi-window convolutional parallel architecture and residual connections. The multi-window convolutional parallel architecture uses convolutional kernels of different sizes to capture the features of the P-wave at specified time periods.

[0010] The multi-scale fused features are input into a parallel integrated encoder, which obtains the temporal correlation within the waveform through a self-attention mechanism, and adaptively assigns different weights to the features of a specified time period. The waveform global features are then output through global average pooling.

[0011] The global features are input into the multi-channel decoder branch, and the output includes earthquake disaster analysis results including initial motion polarity, earthquake type, magnitude, waveform quality, and risk level probability.

[0012] As a preferred scheme for earthquake disaster analysis method driven by near-field P-wave segments, the duration of the near-field P-wave seismic waveform data is 10 seconds, the frequency band of the bandpass filter is 1-20Hz, the normalization preprocessing removes the baseline drift trend of the waveform through linear fitting, and then linear scaling normalization is used.

[0013] As a preferred scheme for earthquake disaster analysis method driven by near-field P-wave segments, the multi-window convolution parallel architecture includes three types of convolution kernel sizes: 3, 5, and 7. The convolution kernel with a size of 3 captures the abrupt change characteristics of the P-wave initial motion segment, the convolution kernel with a size of 5 captures the dominant frequency characteristics of the vibration steady segment, and the convolution kernel with a size of 7 captures the energy change characteristics of the attenuation segment.

[0014] The multi-scale fusion features are input into the parallel ensemble encoder after 1×1 convolution dimensionality reduction. The feature dimension after dimensionality reduction is 320.

[0015] As a preferred scheme for earthquake disaster analysis method driven by near-field P-wave segments, the parallel integrated encoder has 4 layers, and each layer of the parallel integrated encoder includes a multi-head self-attention mechanism and a feedforward network, with the multi-head self-attention mechanism having 8 heads.

[0016] The global features are 320-dimensional and include the amplitude direction of the initial phase of the seismic wave, the amplitude mutation law at the sampling point level, the waveform signal-to-noise ratio, the baseline stability, the amplitude variation trend, the peak amplitude, the overall attenuation rate, and the energy distribution law.

[0017] As a preferred scheme for earthquake disaster analysis method driven by near-field P-wave segments, the multi-channel decoder branches include an initial motion polarity decoder, an earthquake type decoder, a magnitude decoder, a waveform quality decoder, and a risk level decoder.

[0018] The initial polarity decoder adopts a fully connected structure and outputs a positive or negative polarity probability distribution through the Softmax activation function;

[0019] The earthquake type decoder adopts a fully connected structure and outputs the probability of natural earthquake, artificial blasting, or collapse category through the Softmax activation function.

[0020] The magnitude decoder adopts a fully connected structure and outputs continuous magnitude values ​​through a Linear activation function; the waveform quality decoder adopts a fully connected structure and outputs high, medium, and low quality level probability distributions through a Softmax activation function.

[0021] The risk level decoder adopts a fully connected structure and outputs low, medium, high, and extremely high risk probability distributions through the Softmax activation function.

[0022] As a preferred scheme for earthquake disaster analysis methods driven by near-field P-wave segments, it also includes: optimizing the global index output at the event level by aggregating multi-station data through waveform quality adaptive weighting and combining a two-way calibration mechanism of station-level and event-level results;

[0023] The quality assessment of the waveform is achieved through triple decoupling constraints, including input layer standardization to eliminate energy differences, feature layer isolation of energy parameter influence through attention mask, and loss layer addition of correlation penalty term to avoid quality assessment from depending on energy features;

[0024] The quality of the waveform is calculated based on a weighted total quality score using four indicators: signal-to-noise ratio, baseline drift, amplitude abrupt change rate, and P-wave integrity. The specific calculation method was later determined as follows:

[0025] ,

[0026] In the formula, The signal-to-noise ratio after standardization. The standardized baseline drift. The standardized amplitude abrupt change rate, For the integrity of the standardized P-wave.

[0027] As the preferred method for earthquake hazard analysis driven by near-field P-wave segments, the standardization of the four indicators is as follows:

[0028] ,

[0029] ,

[0030] ,

[0031] ,

[0032] The signal-to-noise ratio (SNR) is the root mean square ratio of the amplitudes of the signal segment and the noise segment in a 10-second P-wave waveform. The signal segment is from the beginning to the last 8 seconds of the P-wave in the 10-second waveform, and the noise segment is from the beginning to the first 2 seconds of the P-wave.

[0033] The baseline drift BD is the product of the absolute value of the slope of the linear fitting line of the preprocessed waveform and the number of sampling points;

[0034] The amplitude mutation rate (AMR) is the proportion of the number of times the amplitude difference between adjacent sampling points in a waveform exceeds three times the overall amplitude standard deviation out of the total number of sampling points.

[0035] The P-wave integrity PI is identified by energy detection method when the P-wave arrives. If the P-wave arrives within the first 3 seconds of the 10-second waveform, then PI=1; otherwise, PI=0.

[0036] As a preferred method for earthquake hazard analysis driven by near-field P-wave segments, the risk level probability is generated through a risk index. Based on the magnitude estimate, epicentral distance, and site type, the specific calculation formula is as follows:

[0037] ,

[0038] In the formula, M is the calculated magnitude; This represents the straight-line distance between the station and the earthquake epicenter. The site type is identified by shear wave velocity and is divided into hard rock and soft soil. The disaster risk score is as follows:

[0039] When 0 When the value is less than 1, the risk level is low and the label value is 0.

[0040] When 1 When the value is less than 3, the risk level is medium risk and the label value is 1.

[0041] When 3 When the value is less than 5, the risk level is high risk, and the label value is 2.

[0042] When 5 At that time, the risk level was extremely high, and the label value was 3;

[0043] When the event type is artificial blasting, the risk level of all stations is downgraded to low risk; when the event type is collapse, the risk level of adjacent stations is calibrated to high risk; when ≥80% of stations output high risk but the event magnitude is <2.5, magnitude reassessment is triggered; when ≥90% of stations are judged to have low reliability in waveform quality, the event level result is marked as pending review.

[0044] As a preferred scheme for earthquake disaster analysis methods driven by near-field P-wave segments, it also includes a two-level task collaborative training mechanism. In the early stage of training, the station-level loss weight is 0.7 and the event-level loss weight is 0.3. In the later stage of training, the station-level loss weight is 0.3 and the event-level loss weight is 0.7.

[0045] Cross-entropy loss was used for classification tasks, Huber loss was used for regression tasks, and semi-supervised consistency loss was added. The adaptation part had missing labels. During training, the network weights were updated in reverse through the AdamW optimizer. The total loss was obtained by weighting and summing the polarity loss, earthquake type loss, magnitude loss, waveform quality loss and risk level loss in a weight ratio of 0.15:0.2:0.25:0.2:0.2.

[0046] This invention also provides a near-field P-wave segment-driven earthquake disaster analysis system, employing the aforementioned near-field P-wave segment-driven earthquake disaster analysis method, comprising:

[0047] The data preprocessing module is used to receive near-field P-wave seismic waveform data recorded by multiple stations and multiple channels, and to perform detrending, bandpass filtering and normalization preprocessing on the waveform data.

[0048] The feature extraction module is used to extract multi-scale fusion features of the preprocessed waveform data through a multi-window convolutional parallel architecture and residual connections. The multi-window convolutional parallel architecture uses convolutional kernels of multiple sizes to capture features of the P-wave at specified time periods.

[0049] The shared encoding module is used to input the multi-scale fused features into the parallel integrated encoder, obtain the temporal correlation within the waveform through a self-attention mechanism, adaptively assign different weights to the features of a specified time period, and output the global features of the waveform through global average pooling.

[0050] The multi-task decoding module is used to input the global features into the multi-channel decoder branch and output earthquake disaster analysis results including initial motion polarity, earthquake type, magnitude, waveform quality and risk level probability.

[0051] The aggregation calibration module is used to aggregate data from multiple stations through adaptive weighting of waveform quality, and to optimize the output of global indicators at the event level by combining a two-way calibration mechanism of station-level and event-level results.

[0052] The present invention has the following advantages:

[0053] First, the present invention adopts a single encoder and multi-branch decoder architecture, takes short P-waves as input, and eliminates multi-model connection delay through end-to-end parallel inference, which greatly shortens the time of the entire process from waveform acquisition to risk output, and meets the second-level early warning requirements of near-field earthquakes.

[0054] Second, this invention fully captures key features of P waves at different time periods through multi-window feature extraction and residual connection technology, avoiding the problems of short sequence feature loss and gradient vanishing, and improving the accuracy of core parameter estimation such as magnitude and event type.

[0055] Third, this invention achieves complete decoupling of waveform quality and energy parameters based on triple decoupling constraints, and relies solely on waveform morphology characteristics for quality determination, avoiding interference from factors such as magnitude and epicentral distance, and providing a stable and reliable basis for quality screening in parameter analysis.

[0056] Fourth, this invention integrates the dual attributes of stations and seismic sources to construct a risk mapping mechanism. Through cross-feature attention interaction, it achieves dynamic fusion of physical features and spatial attributes, enabling risk classification to adapt to the geological vulnerability of different regions and highly align with actual disaster prevention needs.

[0057] Fifth, this invention simultaneously outputs five core parameters: initial motion polarity, earthquake type, magnitude, waveform quality, and risk level probability, enabling feature reuse and complementarity, avoiding error accumulation caused by multi-model splicing, and improving the consistency of multi-task results. Attached Figure Description

[0058] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0059] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0060] Figure 1 This is a schematic diagram of the near-field P-wave segment-driven earthquake disaster analysis method provided in an embodiment of the present invention.

[0061] Figure 2 This is a schematic diagram of the seismic waveform sample used in the near-field P-wave segment-driven earthquake disaster analysis method provided in this embodiment of the invention.

[0062] Figure 3 This is a schematic diagram of the near-field P-wave segment-driven earthquake disaster analysis system architecture provided in an embodiment of the present invention. Detailed Implementation

[0063] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] Example 1

[0065] See Figure 1 and Figure 2 Embodiment 1 of the present invention provides a method for earthquake disaster analysis driven by near-field P-wave segments, comprising the following steps:

[0066] S1. Receive near-field P-wave seismic waveform data recorded by multiple stations and multiple channels, and perform detrending, bandpass filtering and normalization preprocessing on the waveform data.

[0067] S2. Extract multi-scale fusion features of the preprocessed waveform data through a multi-window convolutional parallel architecture and residual connections. The multi-window convolutional parallel architecture uses convolutional kernels of multiple sizes to capture the features of the P-wave at specified time periods.

[0068] S3. Input the multi-scale fusion features into the parallel integrated encoder, obtain the temporal correlation within the waveform through the self-attention mechanism, adaptively assign different weights to the features of the specified time period, and output the global features of the waveform through global average pooling.

[0069] S4. Input the global features into the multi-channel decoder branch and output the earthquake disaster analysis results, including initial motion polarity, earthquake type, magnitude, waveform quality, and risk level probability.

[0070] In one possible embodiment, the duration of the near-field P-wave seismic waveform data in step S1 is 10 seconds, the frequency band of the bandpass filter is 1-20Hz, the normalization preprocessing removes the baseline drift trend of the waveform through linear fitting, and then linear scaling normalization is applied.

[0071] Specifically, multiple stations and multiple channels can simultaneously acquire seismic wave vibration signals in different spatial dimensions, ensuring data comprehensiveness. Detrending processing eliminates drift errors caused by long-term instrument operation. By fitting a baseline and removing trend components, the waveform baseline is brought back to a stable state, avoiding baseline shift interference in subsequent feature extraction. Bandpass filtering focuses on the 1-20Hz effective frequency band, which covers the main energy signals of near-field P-waves, while filtering out high-frequency noise and low-frequency interference, achieving signal purification. Preprocessing uses linear fitting to remove the baseline drift trend of the waveform, followed by linear scaling and normalization to map the waveform data to the [0,1] interval, eliminating energy differences caused by different magnitudes and epicentral distances.

[0072] The 10-second duration was chosen to balance the timeliness requirements of near-field earthquake early warning with the integrity of P-wave characteristics. This ensures that the main energy segment of the P-wave is captured while meeting the real-time requirements of second-level early warning. The 1-20Hz bandpass filter band is the optimal range verified by a large amount of historical data, which can retain effective signals to the greatest extent and eliminate interference. Linear scaling and normalization eliminates energy differences in different scenarios through standardization, so that model training and inference are not affected by energy parameters such as magnitude and epicentral distance, thereby improving the generalization ability of the model.

[0073] In one possible embodiment, in the multi-window convolutional parallel architecture described in step S2, the convolutional kernel size includes three types: 3, 5, and 7. The convolutional kernel with a size of 3 captures the abrupt change characteristics of the P-wave initial phase, the convolutional kernel with a size of 5 captures the main frequency characteristics of the stable vibration phase, and the convolutional kernel with a size of 7 captures the energy change characteristics of the decay phase. The multi-scale fusion features are input into the parallel integrated encoder after being processed by 1×1 convolutional dimensionality reduction, and the feature dimension after dimensionality reduction is 320 dimensions.

[0074] Specifically, the characteristics of near-field P-waves differ significantly across different time periods. The abrupt changes in the initial phase, the dominant frequency characteristics in the stationary phase, and the energy changes in the decay phase are all crucial for parameter estimation. A multi-window convolutional parallel architecture uses convolutional kernels of 3, 5, and 7 sizes to specifically capture these features. Small-sized kernels have a small receptive field, capturing subtle amplitude changes in the initial phase; medium-sized kernels balance the receptive field with detail, suitable for extracting dominant frequency information in the stationary phase; and large-sized kernels have a large receptive field, covering the energy change trend in the decay phase. Residual connections solve the gradient vanishing problem in deep network training by superimposing the original input features with the convolutional output features, ensuring that key shallow features are not lost. 1×1 convolutional dimensionality reduction compresses data dimensions while preserving feature expressiveness, improving computational efficiency in the encoding process.

[0075] The choice of convolution kernel size is highly compatible with the physical characteristics of P-waves, and the parallel operation of convolution kernels of different sizes achieves comprehensive feature coverage. 1×1 convolution reduces the dimensionality to 320 dimensions, representing an optimal trade-off between feature integrity and computational efficiency. This avoids computational redundancy caused by excessive dimensionality while ensuring that key features are not lost, providing a suitable data format for efficient encoder processing.

[0076] In one possible embodiment, in step S3, the parallel integrated encoder has 4 layers, and each layer of the parallel integrated encoder includes a multi-head self-attention mechanism and a feedforward network, wherein the multi-head self-attention mechanism has 8 heads.

[0077] The global features have 320 dimensions and include the amplitude direction of the initial phase of the seismic wave, the amplitude mutation law at the sampling point level, the waveform signal-to-noise ratio, the baseline stability, the amplitude variation trend, the peak amplitude, the overall attenuation rate, and the energy distribution law.

[0078] Specifically, the self-attention mechanism can automatically uncover the temporal dependencies within the waveform, such as identifying the correlation pattern between the initial motion and decay phases. Simultaneously, it assigns high weights to key time periods strongly correlated with parameter estimation (such as the period containing the peak amplitude), thereby enhancing useful features and suppressing noise. The feedforward network enhances the expressive power of features through nonlinear transformations. After four layers of iterative learning, global average pooling transforms the temporal features into fixed-dimensional global features. These features integrate core information such as the initial motion direction, amplitude mutation patterns, signal-to-noise ratio, and stability of the seismic wave, providing a unified feature foundation for multi-task decoding.

[0079] The design of combining a 4-layer encoder with an 8-head self-attention mechanism ensures both the depth of feature extraction and the parallel capture of multi-dimensional features through multi-head attention. The 320-dimensional global feature dimension is the optimal configuration verified by extensive experiments. It contains multiple features that cover the core information required for seismic parameter estimation and risk assessment, ensuring that the decoder can accurately output various results.

[0080] In one possible embodiment, step S4, the multi-channel decoder branch includes an initial motion polarity decoder, an earthquake type decoder, a magnitude decoder, a waveform quality decoder, and a risk level decoder;

[0081] The initial polarity decoder adopts a fully connected structure and outputs a positive or negative polarity probability distribution through the Softmax activation function;

[0082] The earthquake type decoder adopts a fully connected structure and outputs the category of natural earthquake, artificial blasting, or collapse through the Softmax activation function.

[0083] The magnitude decoder adopts a fully connected structure and outputs continuous magnitude values ​​through a Linear activation function; the waveform quality decoder adopts a fully connected structure and outputs high, medium, and low quality level probability distributions through a Softmax activation function.

[0084] The risk level decoder adopts a fully connected structure and outputs low, medium, high, and extremely high risk probability distributions through the Softmax activation function.

[0085] Specifically, the multi-branch decoder branches adopt a customized design to adapt to the characteristics of different output parameters. The initial motion polarity decoder uses a binary classification structure and a Softmax activation function to output the probability distribution of positive / negative polarity, providing a foundation for source mechanism analysis. The earthquake type decoder uses a multi-classification structure to accurately identify the characteristic differences of three common events: natural earthquakes, artificial blasting, and collapse. The magnitude decoder uses a linear activation function to output continuous values, adapting to the needs of magnitude regression tasks. The waveform quality and risk level decoders both output levels and probability distributions through multi-classification structures, providing a basis for data reliability judgment and disaster prevention decision-making, respectively. This multi-branch design enables parallel inference for multiple tasks, avoiding the latency and error accumulation caused by splicing multiple models.

[0086] In one possible embodiment, it further includes:

[0087] By adaptively weighting and aggregating data from multiple stations based on waveform quality, and combining this with a two-way calibration mechanism between station-level and event-level results, the output of global event-level indicators is optimized. While multi-station data aggregation improves the reliability of event-level results, waveform quality varies between stations, and low-quality data can interfere with the overall results. Adaptive weighting aggregation assigns weights based on the quality level of each station; for example, high-quality, medium-quality, and low-quality stations are assigned weight coefficients of 0.6, 0.3, and 0.1 respectively, automatically reducing the impact of low-quality data. The two-way calibration mechanism corrects station-level risk probabilities through event-level results and verifies event-level parameters through station-level results, forming a closed-loop calibration that eliminates error propagation at a single level and ensures the accuracy of global indicators.

[0088] The waveform quality assessment is achieved through a triple decoupling constraint: input layer standardization eliminates energy differences; the feature layer isolates the influence of energy parameters through attention masking; and the loss layer adds a correlation penalty term to prevent quality assessment from depending on energy features. Traditional quality assessments are easily interfered with by energy parameters such as magnitude and epicentral distance, leading to assessment distortion. The triple decoupling constraint severs the correlation between quality assessment and energy parameters at three levels: input, feature, and loss. Input layer standardization eliminates energy differences; the feature layer attention mask isolates energy features; and the loss layer correlation penalty term forces the model to not rely on energy parameters to determine quality. Ultimately, the quality assessment is based solely on waveform morphology features, ensuring the objectivity and reliability of the assessment results.

[0089] The waveform quality is calculated based on a weighted average of four indicators: signal-to-noise ratio, baseline drift, amplitude abrupt change rate, and P-wave integrity. The specific calculation method was later determined as follows:

[0090]

[0091] In the formula, The signal-to-noise ratio after standardization. The standardized baseline drift. The standardized amplitude abrupt change rate, The standardized P-wave integrity is assessed. Signal-to-noise ratio (SNR) is a core indicator of waveform quality, reflecting the degree of separation between signal and noise; therefore, it is assigned the highest weight of 0.4. Baseline drift, amplitude abrupt change rate, and P-wave integrity are assessed from the perspectives of baseline stability, amplitude stationarity, and signal integrity, respectively, each assigned a weight of 0.2, forming a comprehensive quality assessment system. After weighted summation, a threshold is used to classify the quality into high, medium, and low levels, providing a quantitative basis for data selection and result reliability judgment.

[0092] In one possible implementation, the four metrics are standardized as follows:

[0093] ,

[0094] ,

[0095] ,

[0096] ,

[0097] The signal-to-noise ratio (SNR) is the root mean square ratio of the amplitudes of the signal segment and the noise segment in a 10-second P-wave waveform. The signal segment is the 8 seconds following the arrival of the P-wave in the 10-second waveform, and the noise segment is the 2 seconds before the arrival of the P-wave. The baseline drift (BD) is the product of the absolute slope of the linearly fitted line of the preprocessed waveform and the number of sampling points. The amplitude mutation rate (AMR) is the proportion of the number of times the amplitude difference between adjacent sampling points in the waveform exceeds three times the overall amplitude standard deviation out of the total number of sampling points. The P-wave integrity (PI) is determined by energy detection to identify the arrival time of the P-wave; if the arrival time of the P-wave occurs within the first 3 seconds of the 10-second waveform, PI = 1; otherwise, PI = 0. The purpose of standardizing the indicators is to eliminate dimensional differences and allow for direct weighted calculation of each indicator. SNR uses a Sigmoid mapping to unify signal-to-noise ratios of different ranges into the [0,1] interval, with SNR=5 as the dividing point to distinguish between high and low SNR. BD and AMR use an inverse mapping, so that a larger index value represents better waveform quality, which conforms to intuitive judgment logic. PI is a binary index that directly reflects whether the P-wave is completely captured. The definitions of each index are based on the physical characteristics of the seismic waveform to ensure the scientific and reasonable nature of the quality assessment.

[0098] In one possible embodiment, the risk level probability is generated by calculating a risk index, which is constructed based on the magnitude estimate, epicentral distance, and site type. The specific calculation formula is as follows:

[0099] ,

[0100] In the formula, M is the calculated magnitude; This represents the straight-line distance between the station and the earthquake epicenter. The site type is identified by shear wave velocity and is divided into hard rock and soft soil. The disaster risk scores are shown in Table 1 below:

[0101] Table 1 Disaster Risk Score

[0102] interval Risk level Tag value 0 <1 Low risk 0 1 <3 Medium risk 1 3 <5 High risk 2 5 Extremely high risk 3

[0103] When the event type is artificial blasting, the risk level of all stations is downgraded to low risk; when the event type is collapse, the risk level of adjacent stations is calibrated to high risk; when ≥80% of stations output high risk but the event magnitude is <2.5, magnitude reassessment is triggered; when ≥90% of stations are judged to have low reliability in waveform quality, the event level result is marked as pending review.

[0104] Among them, the risk index The system integrates three major disaster factors: earthquake energy (magnitude), propagation attenuation (epicentral distance), and site effects (site type), conforming to the physical laws of earthquake-induced disasters: magnitude ≥ 3 begins to contribute to the risk value, the risk increases rapidly when the epicentral distance is < 10 km, and the risk is amplified in soft soil sites due to the amplification effect. The empirical formula is fitted based on regional historical earthquake data to ensure applicability to near-field short P-wave scenarios. The risk index is mapped to discrete risk level labels, enabling the classification and output of earthquake-induced disaster risks, providing a basis for disaster prevention decision-making.

[0105] In one possible embodiment, the bidirectional calibration mechanism includes:

[0106] When the event type is artificial blasting, the risk level of all stations is downgraded to low risk; when the event type is collapse, the risk level of adjacent stations is calibrated to high risk; when ≥80% of stations output high risk but the event magnitude is <2.5, magnitude reassessment is triggered; when ≥90% of stations are judged to have low reliability in waveform quality, the event level result is marked as pending review.

[0107] Specifically, the two-way calibration mechanism is designed based on the actual patterns of earthquake-induced disasters. High-magnitude events have a greater impact on adjacent areas, hence the risk weight is increased; artificial blasting has a limited disaster range, while collapse causes severe damage to adjacent areas, corresponding to risk adjustments of decreasing and increasing respectively; when the risk distribution at a station contradicts the event magnitude, a magnitude reassessment is triggered to avoid misjudgment; when the proportion of low-quality data is too high, the results are marked for review to ensure output reliability. This mechanism eliminates errors in a single link and improves the accuracy of the overall analysis results through mutual verification and correction of station-level and event-level results.

[0108] In one possible embodiment, a two-level task collaborative training mechanism is also included, in which the station-level loss weight is 0.7 and the event-level loss weight is 0.3 in the early stage of training, and in the later stage of training, the station-level loss weight is 0.3 and the event-level loss weight is 0.7.

[0109] Cross-entropy loss is used for classification tasks, Huber loss is used for regression tasks, and semi-supervised consistency loss is added to address partial missing labels. During training, the network weights are updated in reverse through the AdamW optimizer. The total loss is obtained by weighting and summing the polarity loss, earthquake type loss, magnitude loss, waveform quality loss, and risk level loss in a weight ratio of 0.15:0.2:0.25:0.2:0.2.

[0110] Specifically, the dual-level task collaborative training mechanism is designed to address the correlation between station-level and event-level tasks. Initially, it focuses on station-level task training to solidify basic feature extraction capabilities; later, it focuses on event-level task optimization to improve global metric accuracy and avoid model skewness. The choice of loss function adapts to different task types: cross-entropy loss is suitable for classification tasks, Huber loss is robust to outliers and suitable for magnitude regression tasks, and semi-supervised consistency loss utilizes some unlabeled data to improve the accuracy of initial motion polarity determination. The AdamW optimizer suppresses overfitting through weight decay, and the weight allocation of the total loss is adjusted according to the importance and training difficulty of each task to ensure simultaneous optimization across multiple tasks and improve overall model performance.

[0111] In one application of this invention, the first step is to input seismic waveform data with 1000 sampling points in 3 components, corresponding to 10-second near-field station P-wave signals. This data undergoes preprocessing steps of detrending, filtering, and normalization to eliminate instrument drift, purify effective frequency band signals, and unify data distribution. Next, a multi-scale feature extraction stage is entered. Multi-window convolutional layers with kernel sizes of 3 / 5 / 7 and a kernel count of 64 are used to capture features of the P-wave initial motion, stationary phase, and attenuation phase, respectively. After multiple rounds of convolution operations (with dimensions successively adjusted to 64 and 128 dimensions) and combined with residual connections, a 128-dimensional multi-scale fused feature is obtained.

[0112] These fused features are then fed into a shared encoder consisting of four layers, each containing an eight-head self-attention mechanism and a feedforward network (FFN). The self-attention mechanism is used to mine the temporal correlations within the waveform and assign high weights to key features, while the feedforward network enhances the nonlinear expressive power of the features. Finally, the encoder output dimension is compressed to 320-dimensional encoded features.

[0113] The encoded features are then distributed to five independent decoder branches: the initial motion polarity decoder outputs positive / negative polarity probabilities in dimension [C,3] (corresponding to the initial motion polarity of the three-component waveform for each station); the earthquake type decoder outputs natural earthquake / blasting / collapse classification results in dimension [C,3] (corresponding to the earthquake event type probability for each channel); the magnitude decoder outputs continuous magnitude values ​​in dimension [C,1] (corresponding to the earthquake event magnitude for each channel); the waveform quality decoder outputs high / medium / low quality results in dimension [C,3] (corresponding to the waveform data quality for each channel); and the risk level decoder outputs low / medium / high / extremely high risk results in dimension [C,4] (corresponding to the earthquake event risk level for each channel).

[0114] During training, each decoder branch is matched with its corresponding label (such as initial motion polarity label and earthquake type label) to calculate the loss. Classification tasks (initial motion polarity, earthquake type, waveform quality, and risk level) use cross-entropy loss, while magnitude regression tasks use Huber loss (δ=1.0). The total loss is obtained by summing the losses of each branch according to their weights, and then backpropagated through an AdamW optimizer with a learning rate of 1e-4 to update the parameters of the entire network. Finally, the outputs of each decoder are aggregated to provide the multi-task prediction results, directly providing the parameters required for earthquake disaster analysis. This architecture, through multi-scale feature reuse, shared encoding, and multi-task parallel decoding, achieves efficient multi-parameter synchronous output under short P-wave input, while hierarchical loss optimization ensures the prediction accuracy of each task.

[0115] The application scenarios of this invention are as follows:

[0116] Near-field short-term earthquake early warning: Stations deployed in earthquake-prone areas such as cities and mountains use 10-second short P-waves to quickly output magnitude and risk level, providing second-level evacuation warnings for densely populated places (schools, hospitals), and are compatible with existing earthquake early warning systems such as JOPENS and EEW.

[0117] Mine / Industrial Area Safety Monitoring: Deployment of stations in coal mines, metal mines, and other areas to identify non-natural earthquake events such as collapses and blasting, distinguish between artificial blasting and natural earthquakes, determine the collapse risk level in real time, and assist in decision-making regarding equipment shutdown and personnel evacuation in the mining area.

[0118] Monitoring in sparsely populated areas: It supports independent operation of a single station and can output multi-parameter results without networking. It can cover areas with low station density, such as remote mountainous areas and rural areas, and fill monitoring blind spots.

[0119] Transportation infrastructure protection: Deployed at stations near key transportation facilities such as high-speed railways, bridges, and tunnels to quickly assess the disaster risk of earthquakes to the facilities and trigger emergency measures such as track speed reduction and tunnel closure to reduce facility damage and operational risks.

[0120] Monitoring of potential geological hazards: For areas with potential geological hazards such as karst collapses and landslides, the risk level of local hazards is accurately determined by combining site type parameters, which assists in the routine monitoring and emergency response of potential hazard points.

[0121] Earthquake science research support: The output parameters such as initial motion polarity and magnitude can serve as basic data for the analysis of natural earthquake focal mechanisms, providing high-frequency and accurate observational support for regional seismic activity research.

[0122] It should be noted that the method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In a distributed scenario, one of the multiple devices may execute only one or more steps of the method in this embodiment, and the multiple devices may interact with each other to complete the near-field P-wave segment-driven earthquake disaster analysis method.

[0123] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0124] Example 2

[0125] See Figure 3 Embodiment 2 of the present invention also provides a near-field P-wave segment-driven earthquake disaster analysis system, employing the near-field P-wave segment-driven earthquake disaster analysis method of Embodiment 1 or any possible implementation thereof, including:

[0126] The data preprocessing module 100 is used to receive near-field P-wave seismic waveform data recorded by multiple stations and multiple channels, and to perform detrending, bandpass filtering and normalization preprocessing on the waveform data.

[0127] The feature extraction module 200 is used to extract multi-scale fusion features of the preprocessed waveform data through a multi-window convolutional parallel architecture and residual connections. The multi-window convolutional parallel architecture uses convolutional kernels of multiple sizes to capture features of the P-wave at specified time periods.

[0128] The shared encoding module 300 is used to input the multi-scale fused features into the parallel integrated encoder, obtain the temporal correlation within the waveform through a self-attention mechanism, adaptively assign different weights to the features of a specified time period, and output the global features of the waveform through global average pooling.

[0129] The multi-task decoding module 400 is used to input the global features into the multi-channel decoder branch and output earthquake disaster analysis results including initial motion polarity, earthquake type, magnitude, waveform quality and risk level probability.

[0130] The aggregation calibration module 500 is used to aggregate data from multiple stations through adaptive weighted aggregation of waveform quality, and to optimize the output of global indicators at the event level by combining a two-way calibration mechanism of station-level and event-level results.

[0131] In this embodiment, in the data preprocessing module 100, the duration of the near-field P-wave seismic waveform data is 10 seconds, the frequency band of the bandpass filter is 1-20Hz, and the normalization preprocessing removes the baseline drift trend of the waveform through linear fitting, followed by linear scaling normalization.

[0132] In this embodiment, in the feature extraction module 200, the multi-window convolutional parallel architecture includes three types of convolution kernel sizes: 3, 5, and 7. A convolution kernel with a size of 3 captures the abrupt change features of the P-wave initial phase, a convolution kernel with a size of 5 captures the main frequency features of the stable vibration phase, and a convolution kernel with a size of 7 captures the energy change features of the decay phase.

[0133] The multi-scale fusion features are input into the parallel ensemble encoder after 1×1 convolution dimensionality reduction. The feature dimension after dimensionality reduction is 320.

[0134] In this embodiment, in the shared encoding module 300, the parallel integrated encoder has 4 layers, and each layer of the parallel integrated encoder includes a multi-head self-attention mechanism and a feedforward network, with the multi-head self-attention mechanism having 8 heads.

[0135] The global features are 320-dimensional and include the amplitude direction of the initial phase of the seismic wave, the amplitude mutation law at the sampling point level, the waveform signal-to-noise ratio, the baseline stability, the amplitude variation trend, the peak amplitude, the overall attenuation rate, and the energy distribution law.

[0136] In this embodiment, the multi-task decoding module 400 includes a multi-channel decoder branch comprising an initial motion polarity decoder, an earthquake type decoder, a magnitude decoder, a waveform quality decoder, and a risk level decoder.

[0137] The initial polarity decoder adopts a fully connected structure and outputs a positive or negative polarity probability distribution through the Softmax activation function;

[0138] The earthquake type decoder adopts a fully connected structure and outputs the category of natural earthquake, artificial blasting, or collapse through the Softmax activation function.

[0139] The magnitude decoder adopts a fully connected structure and outputs continuous magnitude values ​​through a Linear activation function; the waveform quality decoder adopts a fully connected structure and outputs high, medium, and low quality level probability distributions through a Softmax activation function.

[0140] The risk level decoder adopts a fully connected structure and outputs low, medium, and high risk probability distributions through the Softmax activation function.

[0141] In one possible embodiment, it further includes:

[0142] The indicator output optimization module 600 is used to optimize the global indicator output at the event level by aggregating data from multiple stations through adaptive weighted aggregation of waveform quality and combining a two-way calibration mechanism of station-level and event-level results.

[0143] The quality assessment of the waveform is achieved through triple decoupling constraints, including input layer standardization to eliminate energy differences, feature layer isolation of energy parameter influence through attention masking, and loss layer addition of correlation penalty term to avoid quality assessment depending on energy features.

[0144] It should be noted that the information interaction and execution process between the modules of the above system are based on the same concept as the method embodiment in Embodiment 1 of this application, and the resulting technical effects are the same as those in the method embodiment of this application. For details, please refer to the description in the method embodiment shown above in this application, and it will not be repeated here.

[0145] Example 3

[0146] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium storing program code of a near-field P-wave segment driven earthquake disaster analysis method. The program code includes instructions for executing the near-field P-wave segment driven earthquake disaster analysis method of Embodiment 1 or any possible implementation thereof.

[0147] Computer-readable storage media can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0148] Example 4

[0149] Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor;

[0150] The processor and the memory communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor can execute the near-field P-wave segment driven earthquake disaster analysis method of Embodiment 1 or any possible implementation thereof by calling the program instructions.

[0151] Specifically, a processor can be implemented in hardware or software. When implemented in hardware, the processor can be a logic circuit, an integrated circuit, etc. When implemented in software, the processor can be a general-purpose processor that reads software code stored in memory. This memory can be integrated into the processor or located outside the processor and exist independently.

[0152] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable system. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0153] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing systems. They can be centralized on a single computing system or distributed across a network of multiple computing systems. Optionally, they can be implemented using program code executable by a computing system, thereby storing them in a storage system for execution by the computing system. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0154] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A near-field P-wave segment-driven earthquake hazard analysis method, characterized in that, Includes the following steps: Receive near-field P-wave seismic waveform data recorded by multiple stations and multiple channels, and perform detrending, bandpass filtering and normalization preprocessing on the waveform data; Multi-scale fusion features of the preprocessed waveform data are extracted using a multi-window convolutional parallel architecture and residual connections. The multi-window convolutional parallel architecture uses convolutional kernels of different sizes to capture features of the P-wave at specified time periods. The multi-scale fused features are input into a parallel integrated encoder, which obtains the temporal correlation within the waveform through a self-attention mechanism, and adaptively assigns different weights to the features of a specified time period. The waveform global features are then output through global average pooling. The global features are input into the multi-channel decoder branch, and the output includes earthquake disaster analysis results including initial motion polarity, earthquake type, magnitude, waveform quality, and risk level probability.

2. The earthquake-induced disaster analysis method driven by near-field P-wave segments according to claim 1, characterized in that, The duration of the near-field P-wave seismic waveform data is 10 seconds, the frequency band of the bandpass filter is 1-20Hz, and the normalization preprocessing removes the baseline drift trend of the waveform through linear fitting, followed by linear scaling normalization.

3. The earthquake hazard analysis method driven by near-field P-wave segments according to claim 1, characterized in that, In the multi-window convolutional parallel architecture, the convolution kernel size includes three types: 3, 5, and 7. The convolution kernel with a size of 3 captures the abrupt change characteristics of the P-wave initial phase, the convolution kernel with a size of 5 captures the main frequency characteristics of the stable phase of vibration, and the convolution kernel with a size of 7 captures the energy change characteristics of the decay phase. The multi-scale fusion features are input into the parallel ensemble encoder after 1×1 convolution dimensionality reduction. The feature dimension after dimensionality reduction is 320.

4. The earthquake-induced disaster analysis method driven by near-field P-wave segments according to claim 1, characterized in that, The parallel integrated encoder has 4 layers, and each layer of the parallel integrated encoder includes a multi-head self-attention mechanism and a feedforward network. The multi-head self-attention mechanism has 8 heads. The global features are 320-dimensional and include the amplitude direction of the initial phase of the seismic wave, the amplitude mutation law at the sampling point level, the waveform signal-to-noise ratio, the baseline stability, the amplitude variation trend, the peak amplitude, the overall attenuation rate, and the energy distribution law.

5. The earthquake-induced disaster analysis method driven by near-field P-wave segments according to claim 1, characterized in that, The multi-channel decoder branches include initial motion polarity decoder, earthquake type decoder, magnitude decoder, waveform quality decoder, and risk level decoder; The initial polarity decoder adopts a fully connected structure and outputs a positive or negative polarity probability distribution through the Softmax activation function; The earthquake type decoder adopts a fully connected structure and outputs the probability of natural earthquake, artificial blasting, or collapse category through the Softmax activation function. The magnitude decoder adopts a fully connected structure and outputs continuous magnitude values ​​through a Linear activation function; the waveform quality decoder adopts a fully connected structure and outputs high, medium, and low quality level probability distributions through a Softmax activation function. The risk level decoder adopts a fully connected structure and outputs low, medium, high, and extremely high risk probability distributions through the Softmax activation function.

6. The near-field P-wave segment-driven earthquake hazard analysis method according to claim 5, characterized in that, Also includes: By adaptively weighting and aggregating data from multiple stations based on waveform quality, and combining a two-way calibration mechanism for station-level and event-level results, the output of global indicators at the event level is optimized. The quality assessment of the waveform is achieved through triple decoupling constraints, including input layer standardization to eliminate energy differences, feature layer isolation of energy parameter influence through attention mask, and loss layer addition of correlation penalty term to avoid quality assessment from depending on energy features; The waveform quality is calculated based on a weighted total score using four indicators: signal-to-noise ratio, baseline drift, amplitude abrupt change rate, and P-wave integrity. The specific calculation method was later determined as follows: ， In the formula, The signal-to-noise ratio after standardization. The standardized baseline drift. The standardized amplitude abrupt change rate, For the integrity of the standardized P-wave.

7. The near-field P-wave segment-driven earthquake disaster analysis method according to claim 6, characterized in that, The standardization method for the four indicators is as follows: ， ， ， ， The signal-to-noise ratio (SNR) is the root mean square ratio of the amplitudes of the signal segment and the noise segment in a 10-second P-wave waveform. The signal segment is from the beginning to the last 8 seconds of the P-wave in the 10-second waveform, and the noise segment is from the beginning to the first 2 seconds of the P-wave. The baseline drift BD is the product of the absolute value of the slope of the linear fitting line of the preprocessed waveform and the number of sampling points; The amplitude mutation rate (AMR) is the proportion of the number of times the amplitude difference between adjacent sampling points in a waveform exceeds three times the overall amplitude standard deviation out of the total number of sampling points. The P-wave integrity PI is identified by energy detection method when the P-wave arrives. If the P-wave arrives within the first 3 seconds of the 10-second waveform, then PI=1; otherwise, PI=0.

8. The near-field P-wave segment-driven earthquake hazard analysis method according to claim 6, characterized in that, The risk level probability is generated by calculating a risk index. Based on the magnitude estimate, epicentral distance, and site type, the specific calculation formula is as follows: ， In the formula, M is the calculated magnitude; This represents the straight-line distance between the station and the earthquake epicenter. The site type is identified by shear wave velocity and is divided into hard rock and soft soil. The disaster risk score is as follows: When 0 When the value is less than 1, the risk level is low and the label value is 0. When 1 When the value is less than 3, the risk level is medium risk and the label value is 1. When 3 When the value is less than 5, the risk level is high risk, and the label value is 2. When 5 At that time, the risk level was extremely high, and the label value was 3; The bidirectional calibration mechanism includes: When the event type is artificial blasting, the risk level of all stations is downgraded to low risk; when the event type is collapse, the risk level of adjacent stations is calibrated to high risk; when ≥80% of stations output high risk but the event magnitude is <2.5, magnitude reassessment is triggered; when ≥90% of stations are judged to have low reliability in waveform quality, the event level result is marked as pending review.

9. The earthquake hazard analysis method driven by near-field P-wave segments according to claim 1, characterized in that, It also includes a two-level task collaborative training mechanism, with a station-level loss weight of 0.7 and an event-level loss weight of 0.3 in the early stage of training, and a station-level loss weight of 0.3 and an event-level loss weight of 0.7 in the later stage of training. Cross-entropy loss was used for classification tasks, Huber loss was used for regression tasks, and semi-supervised consistency loss was added. The adaptation part had missing labels. During training, the network weights were updated in reverse through the AdamW optimizer. The total loss was obtained by weighting and summing the polarity loss, earthquake type loss, magnitude loss, waveform quality loss and risk level loss in a weight ratio of 0.15:0.2:0.25:0.2:0.

2.

10. A near-field P-wave segment-driven earthquake hazard analysis system, employing the near-field P-wave segment-driven earthquake hazard analysis method according to any one of claims 1 to 9, characterized in that, include: The data preprocessing module is used to receive near-field P-wave seismic waveform data recorded by multiple stations and multiple channels, and to perform detrending, bandpass filtering and normalization preprocessing on the waveform data. The feature extraction module is used to extract multi-scale fusion features of the preprocessed waveform data through a multi-window convolutional parallel architecture and residual connections. The multi-window convolutional parallel architecture uses convolutional kernels of different sizes to capture features of the P-wave at specified time periods. The shared encoding module is used to input the multi-scale fused features into the parallel integrated encoder, obtain the temporal correlation within the waveform through a self-attention mechanism, adaptively assign different weights to the features of a specified time period, and output the global features of the waveform through global average pooling. The multi-task decoding module is used to input the global features into the multi-channel decoder branch and output earthquake disaster analysis results including initial motion polarity, earthquake type, magnitude, waveform quality and risk level probability. The aggregation calibration module is used to aggregate data from multiple stations through adaptive weighting of waveform quality, and to optimize the output of global indicators at the event level by combining a two-way calibration mechanism of station-level and event-level results.

Images