Ground vibration signal adaptive decomposition and time-frequency characteristic parameter extraction method

Abstract

The invention discloses a ground vibration signal adaptive decomposition and time-frequency characteristic parameter extraction method, and belongs to the technical field of geophysical signal processing. The method comprises the steps of preprocessing seismic oscillation signals, self-adaptive modal decomposition based on entropy weight residual errors, multi-level time domain feature extraction, multi-scale convolution dictionary learning frequency domain feature coding and self-adaptive time-frequency dynamic feature joint analysis. By introducing an attention mechanism, fusion and significance weight self-learning of time domain and frequency domain and time-frequency characteristics of different modals and different scales are realized, and seismic oscillation characteristic parameters with high expressive power are formed. The method is suitable for decomposition, identification and anomaly detection of seismic signals, and has good adaptivity, discrimination capability and interpretability.

Description

Technical Field

[0001] This invention belongs to the field of geophysical signal processing technology, and more specifically relates to a method for adaptive decomposition of seismic motion signals and extraction of time-frequency feature parameters. Background Technology

[0002] In recent years, with the continuous development of earthquake monitoring and early warning systems, efficient and accurate feature extraction of strong earthquake signals has become an important research direction in earthquake engineering, geophysics, and other fields. Existing earthquake signal feature extraction techniques mainly cover time-domain feature analysis, frequency-domain feature analysis, time-frequency transform, and automatic feature extraction methods based on machine learning. Time-domain and frequency-domain analysis methods are widely used due to their simple principles and high computational efficiency. However, in actual earthquake signal processing, traditional time-domain and frequency-domain methods generally exhibit limited adaptability in the face of highly nonlinear, non-stationary abrupt changes, and complex multi-scale evolution of signals, making it difficult to accurately capture and characterize the non-stationary and sudden features in earthquake signals.

[0003] While time-frequency analysis methods such as wavelet transform and Hilbert-Huang transform can reveal the local time-frequency structure of a signal to some extent, they rely on preset or selected basis functions, making it difficult to have good adaptability to actual seismic signals. Furthermore, they have limitations in decomposing and representing the diversity of seismic signal components and extreme atypical waveforms.

[0004] In recent years, intelligent methods such as deep learning have emerged, demonstrating strong self-learning and representation capabilities for features when large-scale labeled data is readily available. However, in the field of seismology, the limited number of labeled samples and the scarcity of samples from extreme and anomalous seismic events result in insufficient generalization ability of these methods for novel or rare seismic motion signals. Furthermore, their feature extraction process is highly "black box" and lacks sufficient interpretability, affecting their reliable application in actual earthquake monitoring and early warning. Therefore, how to balance the complexity, adaptability, and effectiveness of seismic signals, and construct a novel seismic motion signal analysis method that is high-precision, interpretable, and possesses adaptive feature extraction capabilities, has become one of the urgent technical challenges to be solved. Summary of the Invention

[0005] This invention aims to address the technical problems of existing seismic motion signal feature extraction methods, such as limited adaptability, insufficient feature representation, and lack of interpretability in dealing with nonlinear, non-stationary, and complex multi-scale evolution of signals. It provides a method that can efficiently and adaptively decompose seismic motion signals and jointly extract multi-level, multi-scale, time-frequency dynamic features, thereby improving the ability to accurately characterize and identify complex features of seismic signals.

[0006] To achieve the above objectives, the present invention employs the following technical solution: the method comprises: Seismic ground motion signal preprocessing involves preprocessing the acquired raw seismic ground motion signals, specifically including denoising, normalization, and trend correction, to ensure the accuracy of subsequent decomposition and feature extraction. Based on the adaptive mode decomposition of entropy weight residual, the algorithm of adaptive mode decomposition of entropy weight residual is used to decompose the complex ground motion signal into multiple uncorrelated modal components with different energy distributions layer by layer. After each round of decomposition, the parameters of the next decomposition and the termination strategy are dynamically set according to the current decomposition effect by calculating the component energy entropy and the information entropy of the residual signal, so as to achieve an adaptive decomposition that is highly matched with the complexity of the signal. Multi-level time-domain feature extraction: After completing mode decomposition, multi-level time-domain feature extraction is performed on the original signal and each decomposed mode component, including peak acceleration, peak velocity, peak displacement, effective duration, number of extrema and energy index. Frequency domain feature encoding using multi-scale convolutional dictionary learning: For each modal component, a multi-scale convolutional dictionary learning algorithm is used to encode the frequency domain features. The joint extraction of time-frequency dynamic features employs an adaptive window function-based instantaneous frequency and local energy analysis method to perform high-precision dynamic feature extraction on the time-frequency plane. By jointly extracting the time-frequency dynamic features of instantaneous frequency, local energy, and time-frequency energy moments, the complex time-frequency evolution of seismic signals is characterized, providing a fundamental support for the rapid detection of seismic anomalies and atypical waveforms. Multimodal and multiscale feature saliency modeling and fusion: After completing multidimensional feature extraction, an attention mechanism is introduced to design a saliency weight self-learning algorithm to intelligently fuse features in the three major categories of time domain, frequency domain, and time-frequency domain, as well as features between different modalities and scales.

[0007] In one scheme, the preprocessing of the ground motion signal includes: adaptive threshold wavelet denoising, empirical mode decomposition-residual autoregressive hybrid denoising, normalization and trend correction steps.

[0008] In one scheme, the adaptive mode decomposition based on entropy weight residual includes: performing layer-by-layer decomposition using an entropy weight residual adaptive decomposition algorithm. The decomposition process is based on variational mode decomposition or empirical mode decomposition. After obtaining each mode component and the corresponding residual signal, the energy entropy and information entropy are calculated respectively. The energy entropy is used to measure the energy dispersion of the mode component, and the information entropy is used to measure the complexity of the residual signal. The decomposition is automatically terminated based on preset threshold conditions, enabling adaptive adjustment and progressive decomposition of decomposition control parameters. This results in the acquisition of mutually independent modal components with actual physical meaning, thereby improving the basic accuracy and expressiveness of multi-scale feature extraction.

[0009] In one scheme, the multi-level time-domain feature extraction includes: based on the modal components obtained from adaptive decomposition and the original signal, independently extracting peak acceleration, velocity component, displacement component, number of local extrema, effective duration and total energy time-domain feature indicators for each component; and using an energy entropy weighted fusion strategy, designing the proportion of energy of each modal component in the total energy and energy entropy as fusion weights, and weighting and integrating the time-domain feature quantities of each modality to form a global time-domain feature set.

[0010] In one scheme, the frequency domain feature encoding of the multi-scale convolution dictionary learning includes: designing a multi-scale convolution kernel dictionary to cover different time scales and frequency ranges for each modal component obtained by multi-level time domain decomposition; performing one-dimensional convolution of the modal signal with multiple convolution kernels to obtain a sparse activation sequence; and jointly learning the convolution kernel and sparse activation through an alternating optimization strategy. The frequency response of the convolution kernel is obtained by Fourier transform, and feature parameters such as dominant frequency, center frequency, and bandwidth are extracted accordingly. The activation sequence of each mode under all convolution kernels and the above features are concatenated and encoded to form a compact frequency domain feature vector, which enriches the spectral expression of seismic signals and provides a foundation for subsequent earthquake type identification and focal mechanism analysis.

[0011] In one scheme, the joint extraction of time-frequency dynamic features includes: using an adaptive window function for instantaneous frequency and local energy analysis, and dynamically adjusting the length and shape parameters of the analysis window function to perform adaptive time-frequency analysis on the width of the signal's stationary and transient segments; By using short-time Fourier transform or continuous wavelet transform, instantaneous frequency, local energy, and time-frequency energy moment are calculated to obtain the instantaneous dominant frequency, local energy peak, and energy center trajectory of each modal component. The time-frequency dynamic features are then jointly encoded to form a high-dimensional dynamic feature set describing the time-frequency structure evolution of seismic signals, thereby improving the sensitivity and accuracy of seismic event detection and differentiation.

[0012] In one scheme, the multimodal multiscale feature saliency modeling and fusion includes: uniformly normalizing and high-dimensionally concatenating the obtained multi-level, cross-modal feature set, including time-domain features, frequency-domain features and time-frequency dynamic features, into the original feature tensor; A saliency weight self-learning framework based on attention mechanism is introduced. The weights of features of different modalities, categories and scales are mapped by trainable parameters, and an end-to-end training method is adopted to minimize the loss function related to earthquake event discrimination, so as to realize automatic modeling of feature saliency.

[0013] In one approach, adaptive threshold wavelet denoising is used to suppress high-frequency noise components, empirical mode decomposition-residual autoregressive hybrid denoising is used to remove non-stationary noise and residual artifacts, normalization is used to standardize the signal amplitude to zero mean and unit variance, and trend correction uses the least squares method to correct signal drift or baseline shift.

[0014] In one approach, group attention, hierarchical attention, or self-attention strategies are employed during the fusion process to achieve multi-level feature weighting for intermodal fusion, intra-scale aggregation, and global redistribution, ultimately yielding an overall salient feature vector. This provides a fusion feature foundation that combines discriminative power and interpretability for seismic signal identification and classification.

[0015] Beneficial effects of this invention: This invention achieves effective separation and accurate representation of nonlinear and non-stationary complex components in seismic motion signals by combining an adaptive signal decomposition algorithm with multi-scale time-frequency feature analysis. Compared with traditional time-domain, frequency-domain, and conventional time-frequency analysis methods, this invention does not rely on fixed basis functions and can dynamically adjust the decomposition process according to the characteristics of the seismic signal itself, thus improving the adaptability of feature extraction and signal compatibility.

[0016] Simultaneously, by combining a joint time-frequency parameter extraction strategy, the energy distribution, frequency variations, and sudden events of seismic motion signals can be characterized from multiple levels and perspectives, effectively improving the sensitivity of identifying weak or extreme seismic events. By enhancing the interpretability of feature extraction, a more reliable and traceable basis is provided for subsequent automatic classification, event identification, and early warning decision-making of seismic signals, which helps to improve the scientific rigor and practicality of earthquake monitoring and emergency response systems.

[0017] This invention significantly improves the sensitivity and discrimination ability of feature extraction from strong ground motion signals, accurately capturing the nonlinear, non-stationary, and sudden change characteristics of seismic signals, outperforming traditional time-frequency analysis or single deep learning methods. This method possesses adaptive capabilities, enabling it to adapt to signal distribution changes under different magnitudes, geological structures, and noise environments. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart of the adaptive mode decomposition based on entropy weight residual of the present invention; Figure 3 This is the frequency domain feature encoding process for multi-scale convolutional dictionary learning in this invention. Detailed Implementation

[0019] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0020] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. To facilitate understanding, the invention will now be described more fully with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to make the disclosure of the invention more thorough and complete.

[0021] Figure 1 As shown, a method for adaptive decomposition of seismic ground motion signals and extraction of time-frequency feature parameters includes the following specific steps: Step 1: Seismic ground motion signal preprocessing First, the acquired raw ground motion signals are preprocessed, including denoising, normalization, and trend correction, to ensure the accuracy of subsequent decomposition and feature extraction. To cope with complex noise environments, this step employs adaptive threshold wavelet denoising or a hybrid denoising strategy combining empirical mode decomposition and residual autoregression. This effectively suppresses noise components in the signal while preserving the main structural features of the seismic signal, laying a solid foundation for subsequent fine decomposition and feature extraction.

[0022] In the preprocessing stage of seismic motion signals, the first step is to process the acquired raw seismic signal sequence. Noise suppression, normalization, and trend correction are performed to eliminate interference from environmental noise and instrument drift on subsequent decomposition and feature extraction. For noise suppression, this method employs a hybrid denoising strategy of adaptive threshold wavelet denoising and empirical mode decomposition (EMD)-residual autoregressive (AR). Specifically, for the input signal... First, wavelet decomposition is performed to decompose the signal into approximate components and detail components at different scales. , in For the approximate component of the j-th layer, Let k be the detail component at layer k. For each detail component, calculate its adaptive threshold. (in The noise estimation standard deviation is given, where n is the signal length, and a soft thresholding function is used. Nonlinear denoising is performed to suppress high-frequency noise components.

[0023] After wavelet denoising, EMD decomposition is used to further remove non-stationary noise and residual artifacts from the signal. Decomposed into several intrinsic mode functions (IMFs) and remainder terms. ,Right now For each IMF component, correlation and energy thresholds are used to remove noise-dominant components, retaining only components related to the main signal characteristics to reconstruct the signal. For the remaining significant low-frequency residuals and trend terms, an autoregressive (AR) model is used to further model their residuals, and the minimum mean square error criterion is applied. Perform AR modeling, the model is ,in For AR coefficients, For white noise terms, minimize The optimal parameters are obtained, and the residuals are smoothed and corrected.

[0024] After denoising, the signal is normalized to a zero mean and unit variance. ,in and These represent the mean and standard deviation of the signal, respectively. Finally, to eliminate the trend effects caused by seismic instrument drift or baseline shift, a polynomial trend correction was performed on the signal. The linear or quadratic trend term of the signal was fitted using the least squares method and then removed from the original signal to obtain a stable, drift-free preprocessed signal. Through the above series of operations, noise and artifacts were significantly suppressed, and the main characteristics of the seismic signal were effectively preserved, laying a solid data foundation for subsequent adaptive mode decomposition and multi-scale feature analysis.

[0025] Step 2: Adaptive Mode Decomposition Based on Entropy Weight Residue Based on the preprocessed signal, an entropy-weighted residual adaptive decomposition algorithm is used to decompose the complex seismic motion signal layer by layer into multiple uncorrelated modal components with different energy distributions. After each round of decomposition, the parameters for the next decomposition and the termination strategy are dynamically set according to the current decomposition effect by calculating the component energy entropy and the information entropy of the residual signal, thereby achieving adaptive decomposition that is highly matched to the complexity of the signal.

[0026] like Figure 2 As shown, this method can flexibly adjust the decomposition depth and the number of components, so that the decomposition process will not result in information loss due to under-decomposition, nor will it introduce spurious components due to over-decomposition, thus providing a solid modal foundation for multi-scale feature analysis.

[0027] S201, Regarding the preprocessed signal An entropy-weighted residual adaptive decomposition algorithm is employed to decompose the signal layer by layer, maximizing the preservation of structural differences and dynamic energy characteristics. The decomposition process is based on variational mode decomposition (VMD) or empirical mode decomposition (EMD), obtaining one mode component in each round of decomposition. and corresponding residual signals .

[0028] S202. To achieve automatic decomposition judgment and component determination, the algorithm introduces energy entropy and information entropy as adaptive adjustment criteria. First, for each modal component obtained from the decomposition... Calculate its energy The energy ratio is obtained by normalizing the energy of all current modal components. Therefore, energy entropy is defined as follows: It is used to measure the energy dispersion among modal components. The larger the energy entropy, the more dispersed the component energy is and the more comprehensive the decomposition effect.

[0029] S203, Simultaneously, for the current residual signal Probability distribution estimation (probability obtained by histogram method) ), calculate its information entropy This entropy value is used to characterize the stochastic complexity of the residual signal. When the information entropy of the decomposed residual signal is lower than a preset threshold... or residual energy ratio Below the threshold Time (of which) For residual energy, If the original signal energy is used, it is assumed that the main effective modes have been fully extracted, and the decomposition process automatically terminates. If the termination condition is not met, the next round of decomposition is performed recursively, updating the modal component set and residual information.

[0030] The above-mentioned entropy weight residual adaptive decomposition method can avoid the fragmentation or overfitting problems of traditional decomposition algorithms, making the decomposed modes both independent and have practical physical meaning, which greatly improves the basic accuracy and expressiveness of multi-scale feature extraction of seismic signals.

[0031] Step 3: Multi-level temporal feature extraction After completing modal decomposition, multi-level time-domain feature extraction is performed on the original signal and each decomposed modal component, including indicators such as peak acceleration, peak velocity, peak displacement, effective duration, number of extrema, and energy. A modal energy entropy-weighted fusion strategy is introduced, which automatically adjusts the feature contribution of each mode according to its energy distribution, improving the overall time-domain features' ability to reflect complex changes in seismic signals and various waveforms such as mainshocks and aftershocks, making the feature representation both representative and robust.

[0032] Based on the aforementioned adaptive decomposition, each modal component and the original signal Representative time-domain feature indicators are extracted independently for each component to comprehensively depict the dynamic changes of ground motion signals.

[0033] Specifically, the peak acceleration (maximum absolute value) of each component is first calculated, i.e. For the original acceleration signal, it needs to be integrated once and twice to obtain the velocity components. With displacement components Thus, their respective peak speeds are obtained. and peak displacement .

[0034] Meanwhile, by traversing each component sequence, the number of local extrema is identified and counted. This reflects the oscillations and complexity of the seismic signal. Effective duration In this regard, the start and end times when the energy accumulation reaches 95% of the total energy are usually selected, specifically by calculating the energy accumulation ratio. Based on this, the corresponding time window width is obtained; in addition, the total energy It serves as a fundamental indicator for characterizing the activity intensity of each modal signal.

[0035] After the basic time-domain features are extracted, in order to fully reflect the differences in the contribution of different modes to the overall signal and the representativeness of the time-domain features, an innovative feature integration strategy is adopted using energy entropy weighted fusion.

[0036] First, we consider the proportion of each modal component's energy in the total energy. As a weighting factor, combined with energy entropy It reflects the uniformity of modal energy distribution. The specific fusion weights can be designed as follows: This allows components with higher energy levels and more concentrated energy distributions to account for a larger proportion of the overall characteristics. Ultimately, this results in a higher proportion of the time-domain characteristics for each mode. According to weight By fusion, a global temporal feature set is formed. This overcomes the inadequacy of single-component representation or outlier interference, and improves the ability of features to describe and distinguish complex temporal changes and different waveform structures of seismic signals.

[0037] This multi-level, weighted fusion mechanism for extracting and integrating temporal features can more accurately and robustly characterize the main activity characteristics of earthquakes of different magnitudes and source types, while improving the scientific data support for subsequent earthquake event identification and risk assessment.

[0038] Step 4: Frequency domain feature encoding from multi-scale convolutional dictionary learning For each modal component, a multi-scale convolutional dictionary learning algorithm is used for frequency domain feature encoding. For example... Figure 3 As shown, the specific method involves dynamically constructing a convolution kernel dictionary adapted to different time scales and frequency ranges. By adaptively adjusting the convolution kernel size and weights, deep learning and representation of the local spectral structure of seismic signals are achieved. This method can automatically capture typical frequency domain features such as dominant frequency, center frequency, bandwidth, and spectral energy distribution, greatly expanding the ability to characterize the diversity of signal spectra and complex energy distribution patterns, and providing a solid frequency domain feature foundation for subsequent identification of different earthquake sources and earthquake types.

[0039] S401. First, for each modal component obtained from the previous process (multi-level time domain decomposition)... A convolutional dictionary learning mechanism is designed to deeply mine its spectral features. The core idea of ​​this process is to construct and train a set of multi-scale convolutional kernel dictionaries. This enables the effective extraction of the local frequency domain structure of a signal across different time scales and frequency ranges. Specifically, in the initial stage, a set of convolutional kernels is initialized for each modal component with various lengths and shapes to ensure that its frequency response covers the low-frequency, dominant-frequency, and high-frequency characteristic ranges that the target signal may contain.

[0040] For each convolution kernel The corresponding activation sequence is obtained by performing a one-dimensional convolution with the modal signal. ,Right now S402. Subsequently, a sparse reconstruction model is constructed, and the dictionary and sparse activation are learned by optimizing the lower objective function: in It is a sparsity control parameter. The sparsity of dictionary activations is encouraged to achieve efficient signal decomposition with low redundancy. The dictionary and activation sequences converge iteratively through an alternating optimization strategy to capture typical dominant frequency structures, center frequency characteristics, and energy distribution patterns at different scales.

[0041] The learned multi-scale convolutional kernels not only cover a wide spectrum and multiple time-domain windows, but their frequency response can also be obtained by performing a Fourier transform on each kernel, as shown in the formula: S403. Accordingly, the dominant frequency (maximum response frequency), center frequency (energy centroid), and bandwidth (difference between high and low half-energy frequencies) of each modal component can be directly derived from the convolution kernel and activation sequence, as shown in the following formulas: in and These represent the frequency points at both ends of the half-energy range. The weighted cumulative sum of the convolution kernel response can also characterize the diversity and concentration of the spectral energy distribution.

[0042] Unlike traditional single-scale dictionary learning, this multi-scale convolutional dictionary adaptively adjusts the temporal width and center frequency of each kernel, allowing it to flexibly fit the spectral characteristics of each mode locally or globally, automatically focusing on discriminative spectral patterns in seismic signals. Finally, the activation sequence, dominant frequency, center frequency, bandwidth, and other feature parameters of each mode under all convolutional kernels are concatenated and encoded to form a high-dimensional but compact frequency domain feature vector, providing a solid and detailed spectral representation foundation for subsequent earthquake type identification and focal mechanism analysis. This frequency domain feature encoding method greatly expands the ability to model and analyze the diversity, complex structure, and energy distribution patterns of seismic motion signals, improving the scientific effectiveness of subsequent intelligent identification and classification.

[0043] Step 5: Joint extraction of time-frequency dynamic features An adaptive window function-based instantaneous frequency and local energy analysis method is employed for high-precision dynamic feature extraction in the time-frequency plane. This method dynamically adjusts the length and shape of the analysis window function based on the local stability of the seismic signal, enhancing its ability to capture sudden and non-stationary changes in the signal's time-frequency characteristics. Through the joint extraction of time-frequency dynamic features such as instantaneous frequency, local energy, and time-frequency energy moments, the complex time-frequency evolution of seismic signals can be accurately characterized, providing a fundamental support for the rapid detection of seismic anomalies and atypical waveforms.

[0044] In the joint extraction stage of time-frequency dynamic features, based on the aforementioned modal components and their frequency domain coding features, the instantaneous frequency and local energy analysis method using adaptive window functions is further utilized to provide a high-resolution description of the time-frequency evolution of the seismic signal. First, considering the generally non-stationary and sudden characteristics of seismic signals, dynamic window function optimization based on local signal stability is employed for short-time Fourier transform (STFT) or continuous wavelet transform (CWT) analysis. For the signal to be analyzed... At each time t, the analysis window function is adaptively determined based on stability indices such as signal variance and energy variability within several preceding and following windows. length By adjusting the shape parameters, the window length is widened in the phase of gradual signal change to improve frequency resolution, and the window width is shortened in the phase of abrupt changes and transients to improve time domain resolution, thus achieving "width-narrow adaptive" time-frequency analysis coverage.

[0045] In practical implementation, the selection of the adaptive window length can be made... ,in Minimum window length, To adjust the coefficient, This represents the local variance at the current time. For each analysis window center point, the signal is weighted by an adaptive window. Then perform a short-time Fourier transform: Calculate the instantaneous frequency at each time point using the phase derivative of the Fourier spectrum or through the Hilbert transform. : in This is the analytical expression of the signal (result of Hilbert transform). Local energy can be accumulated by the square of the magnitude of the time-frequency spectrum: Furthermore, the time-frequency energy moment is extracted to reflect the dynamic structure of the time-frequency energy distribution, and the first-order time-frequency moment (energy centroid) is defined as: The corresponding second moment can reflect the discreteness and concentration of energy distribution, thereby capturing atypical time-frequency phenomena such as abnormal energy clusters and frequency drift. For multimodal components, the instantaneous dominant frequency, local energy peak, energy center trajectory, etc. of each component can be obtained separately. Finally, these time-frequency dynamic feature vectors are jointly encoded and converged into a high-dimensional dynamic feature set that characterizes the time-frequency structure evolution of the overall vibration signal.

[0046] This method can naturally and compatiblely handle multi-scale non-stationary events such as mainshocks, aftershocks, and complex waveforms in seismic signals, making the analysis links such as earthquake anomaly detection, source differentiation, and waveform evolution tracking more sensitive and robust, and providing a precise time and frequency foundation for subsequent intelligent earthquake identification and rapid response.

[0047] Step 6: Multimodal and multiscale feature saliency modeling and fusion After completing multidimensional feature extraction, an attention mechanism is introduced to design a saliency weight self-learning algorithm, which intelligently fuses features from the three major categories of time domain, frequency domain, and time-frequency domain, as well as features from different modalities and scales. Through end-to-end weight modeling, the algorithm automatically learns the importance of each feature in seismic event discrimination, effectively highlighting key features, suppressing noise and redundant information, and ultimately forming a high-dimensional saliency feature set that integrates multi-level, multi-modal, and multi-scale information. This step greatly improves the effectiveness, interpretability, and practicality of seismic signal feature representation.

[0048] First, the multi-level, cross-modal feature set obtained in the previous steps (containing time-domain, frequency-domain, and time-frequency features, which are extracted from each modal component and its different scales) is uniformly normalized and high-dimensionally concatenated to construct the original feature tensor. To fully leverage the practical role of various features in earthquake event identification, a saliency weight self-learning framework based on an attention mechanism is introduced to automatically model and adjust feature importance in an end-to-end manner. Specifically, a set of trainable parameters (multi-head attention network or additive / multiplicative attention module) is first used to weight features of different modalities, categories, and scales.

[0049] Taking general additive attention as an example, for each type of feature in the feature tensor Construct a significance scoring function in, For trainable linear transformations and nonlinear activations, and For parameter matrices and vectors, This is the bias term. During end-to-end training, this mechanism automatically optimizes the attention weights for each feature class by minimizing the loss function (cross-entropy) associated with earthquake event discrimination. This allows for the highlighting of key features and the reduction of redundancy and noise.

[0050] For multi-scale features and multi-modal sources, group attention, hierarchical attention, or self-attention strategies can be further employed to refine the fusion process into a three-level process: "inter-modal fusion," "intra-scale aggregation," and "global redistribution." For example, first, weights are assigned to features at different scales within the same modality; then, global attention is applied between cross-modal features; finally, a weighted fusion is used to obtain the overall saliency feature vector. This structure can also capture the complex interactions between features of different categories in parallel through a multi-head attention mechanism, thereby enhancing the expressive power of the fused features.

[0051] During the training phase, iterative learning of attention weights enables the system to automatically identify and amplify the most discriminative and informative feature dimensions based on the seismic signal itself and the labeled event categories, while suppressing noisy bands, redundant dimensions, and uncertainties arising during multimodal fusion. Ultimately, the saliency feature set... It not only highly summarizes the multimodal, multi-scale, and time-frequency dynamic evolution patterns of seismic signals, but also has good discrimination ability and interpretability, providing a solid data foundation for subsequent intelligent tasks such as seismic event identification, classification, and source tracing.

[0052] This attention-based adaptive feature fusion mechanism closely integrates feature engineering with deep learning, greatly improving the adaptability and practicality of feature representation for complex seismic signals.

[0053] Example: I. Explanation of Experimental Data This embodiment uses three sets of raw ground motion signal data recorded by three data acquisition stations (station A, station B, and station C) of a 5.5 magnitude earthquake event from the China Earthquake Networks Center (CENC). The sampling rate is 100Hz, and the signal duration is 60 seconds. For ease of reference, the data is presented in tabular form (see Tables 1, 2, and 3).

[0054] Table 1: Original ground motion signal sampling (first 20 data points as an example) II. Implementation Process 1. Data Preprocessing The original signal undergoes the following steps: The db4 wavelet was used to decompose the signal into five layers. Rigrsure automatic thresholding was employed to remove high-frequency noise and enhance the main signal components. The signal was scaled to the [0,1] interval for easier feature comparison. The least squares method was used to fit the residuals and eliminate long-period trends. After processing, the signal waveform noise was significantly reduced, and the baseline became more uniform.

[0055] 2. Entropy-weighted residual adaptive mode decomposition For each signal, the entropy-weighted residual algorithm is used for decomposition, with the following specific parameters: Maximum number of decomposition levels: 8; Decomposition termination criteria: The single-layer modal entropy weight is less than 0.02, or the change in residual signal energy entropy is less than 1%; Taking the data from station A as an example, it is ultimately decomposed into 5 modal components (M1~M5) plus 1 residual signal (R), as shown in the table below: 3. Multi-level temporal feature extraction Extract the following features for each component and the original signal: Peak acceleration (PGA, Gal); Peak velocity (PGV, cm / s); Peak displacement (PGD, cm); Effective duration (s, cumulative time for signal amplitude ≥ 0.05 times PGA); Number of extreme values ​​(number of times); Energy index (sum of squares and integral); Taking site A as an example, the statistics are as follows (units are shown in the table below): 4. Frequency Domain Feature Encoding from Multi-Scale Convolutional Dictionary Learning For each modal component, convolution atoms at three scales (8 / 16 / 32) are used to learn the spectral basis dictionary, and their sparse coefficients are extracted as frequency domain representations. The distribution of the top three features of their energy projection is then statistically analyzed. (See below.) 5. Joint extraction of time-frequency dynamic features Using adaptive window-length short-time Fourier transform (STFT) and Hilbert instantaneous frequency analysis, combined with time-frequency energy moments, we obtain the following: Maximum instantaneous frequency (Hz); Instantaneous spectral energy peak (Gal^2 / Hz); The rate of change of the center frequency in time (Hz / s); as follows: 6. Multimodal and multiscale feature fusion and saliency modeling Employing a hierarchical attention mechanism: The first layer groups by modality (different modalities), the second layer performs convolutional aggregation (different components at the same scale), and the third layer performs global allocation.

[0056] The system automatically learns the weights of each feature and finally outputs a significant feature vector (final length 38 dimensions) for subsequent earthquake event classification.

[0057] The partial fusion weights are as follows: The test results of this embodiment show: 1. By utilizing entropy-weighted residual adaptive decomposition, the number of signal decomposition layers can be adaptively adjusted, and the decomposed residual signal has low energy and good component discrimination.

[0058] 2. The hierarchical, multi-scale, and multi-modal feature extraction and fusion highlights key time-frequency features, and has a good discriminative effect on seismic signal anomaly identification and early classification of seismic events.

[0059] 3. This method is consistent and adaptable to signals of different magnitudes and at different stations. The fused feature vectors can be directly used in machine learning recognition models to achieve intelligent interpretation and classification of seismic waveforms.

[0060] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0061] It should be understood that the above detailed description of the technical solutions of the present invention with reference to preferred embodiments is illustrative and not restrictive. Those skilled in the art can modify the technical solutions described in the embodiments or make equivalent substitutions for some of the technical features based on reading this specification; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for adaptive decomposition and time-frequency feature parameter extraction of seismic ground motion signals, characterized in that: The method includes: Seismic ground motion signal preprocessing involves preprocessing the acquired raw seismic ground motion signals, specifically including denoising, normalization, and trend correction, to ensure the accuracy of subsequent decomposition and feature extraction. Based on the adaptive mode decomposition of entropy weight residual, the algorithm of adaptive mode decomposition of entropy weight residual is used to decompose the complex ground motion signal into multiple uncorrelated modal components with different energy distributions layer by layer. After each round of decomposition, the parameters of the next decomposition and the termination strategy are dynamically set according to the current decomposition effect by calculating the component energy entropy and the information entropy of the residual signal, so as to achieve an adaptive decomposition that is highly matched with the complexity of the signal. Multi-level time-domain feature extraction: After completing mode decomposition, multi-level time-domain feature extraction is performed on the original signal and each decomposed mode component, including peak acceleration, peak velocity, peak displacement, effective duration, number of extrema and energy index. Frequency domain feature encoding using multi-scale convolutional dictionary learning: For each modal component, a multi-scale convolutional dictionary learning algorithm is used to encode the frequency domain features. The joint extraction of time-frequency dynamic features employs an adaptive window function-based instantaneous frequency and local energy analysis method to perform high-precision dynamic feature extraction on the time-frequency plane. By jointly extracting the time-frequency dynamic features of instantaneous frequency, local energy, and time-frequency energy moments, the complex time-frequency evolution of seismic signals is characterized, providing a fundamental support for the rapid detection of seismic anomalies and atypical waveforms. Multimodal and multiscale feature saliency modeling and fusion: After completing multidimensional feature extraction, an attention mechanism is introduced to design a saliency weight self-learning algorithm to intelligently fuse features in the three major categories of time domain, frequency domain, and time-frequency domain, as well as features between different modalities and scales.

2. The method for adaptive decomposition and time-frequency feature parameter extraction of seismic ground motion signals according to claim 1, characterized in that: The aforementioned seismic motion signal preprocessing includes: adaptive threshold wavelet denoising, empirical mode decomposition-residual autoregressive hybrid denoising, normalization and trend correction steps.

3. The method for adaptive decomposition and time-frequency feature parameter extraction of seismic ground motion signals according to claim 1, characterized in that: The aforementioned adaptive mode decomposition based on entropy weight residuals includes: performing layer-by-layer decomposition using an entropy weight residual adaptive decomposition algorithm. The decomposition process is based on variational mode decomposition or empirical mode decomposition. After obtaining each mode component and the corresponding residual signal, the energy entropy and information entropy are calculated respectively. The energy entropy is used to measure the energy dispersion of the mode component, and the information entropy is used to measure the complexity of the residual signal. The decomposition is automatically terminated based on preset threshold conditions, enabling adaptive adjustment and progressive decomposition of decomposition control parameters. This results in the acquisition of mutually independent modal components with actual physical meaning, thereby improving the basic accuracy and expressiveness of multi-scale feature extraction.

4. The method for adaptive decomposition and time-frequency feature parameter extraction of seismic ground motion signals according to claim 1, characterized in that: The multi-level time-domain feature extraction includes: based on the modal components obtained from adaptive decomposition and the original signal, independently extracting peak acceleration, velocity component, displacement component, number of local extrema, effective duration and total energy time-domain feature indicators for each component; and using an energy entropy weighted fusion strategy, designing the proportion of energy of each modal component in the total energy and energy entropy as fusion weights, and weighting and integrating the time-domain feature quantities of each modality to form a global time-domain feature set.

5. The method for adaptive decomposition and time-frequency feature parameter extraction of seismic ground motion signals according to claim 1, characterized in that: The frequency domain feature encoding of the multi-scale convolution dictionary learning includes: designing a multi-scale convolution kernel dictionary to cover different time scales and frequency ranges for each modal component obtained by multi-level time domain decomposition; performing one-dimensional convolution of the modal signal with multiple convolution kernels to obtain a sparse activation sequence; and jointly learning the convolution kernel and sparse activation through an alternating optimization strategy. The frequency response of the convolution kernel is obtained by Fourier transform, and the dominant frequency, center frequency, and bandwidth feature parameters are extracted accordingly. The activation sequence of each mode under all convolution kernels and the above features are concatenated and encoded to form a compact frequency domain feature vector, which enriches the spectral expression of the seismic signal and provides a basis for subsequent earthquake type identification and focal mechanism analysis.

6. The method for adaptive decomposition and time-frequency feature parameter extraction of seismic ground motion signals according to claim 1, characterized in that: The aforementioned joint extraction of time-frequency dynamic features includes: using an adaptive window function for instantaneous frequency and local energy analysis, and dynamically adjusting the length and shape parameters of the analysis window function to perform adaptive time-frequency analysis on the width of the signal's stationary and transient segments; By using short-time Fourier transform or continuous wavelet transform, instantaneous frequency, local energy, and time-frequency energy moment characteristics are calculated, and the instantaneous dominant frequency, local energy peak, and energy center trajectory of each modal component are obtained respectively. The time-frequency dynamic features are jointly encoded to form a high-dimensional dynamic feature set describing the time-frequency structure evolution of seismic signals, thereby improving the sensitivity and accuracy of seismic event detection and differentiation.

7. The method for adaptive decomposition and time-frequency feature parameter extraction of seismic ground motion signals according to claim 1, characterized in that: The multimodal and multi-scale feature saliency modeling and fusion includes: uniformly normalizing and high-dimensionally concatenating the obtained multi-level, cross-modal feature set, including time-domain features, frequency-domain features and time-frequency dynamic features, into the original feature tensor; A saliency weight self-learning framework based on attention mechanism is introduced. The weights of features of different modalities, categories and scales are mapped by trainable parameters, and an end-to-end training method is adopted to minimize the loss function related to earthquake event discrimination, so as to realize automatic modeling of feature saliency.

8. The method for adaptive decomposition of seismic motion signals and extraction of time-frequency characteristic parameters according to claim 2, characterized in that: Adaptive threshold wavelet denoising is used to suppress high-frequency noise components, empirical mode decomposition-residual autoregressive hybrid denoising is used to remove non-stationary noise and residual artifacts, normalization is used to standardize the signal amplitude to zero mean and unit variance, and trend correction uses the least squares method to correct signal drift or baseline shift.

9. The method for adaptive decomposition of seismic motion signals and extraction of time-frequency characteristic parameters according to claim 7, characterized in that: During the fusion process, group attention, hierarchical attention, or self-attention strategies are adopted to achieve multi-level feature weighting for inter-modal fusion, intra-scale aggregation, and global redistribution, ultimately obtaining an overall saliency feature vector. This provides a fusion feature foundation with both discriminative power and interpretability for seismic signal identification and classification.

Images