Search

CN-121995465-A - Adaptive decomposition and time-frequency characteristic parameter extraction method for earthquake signals

CN121995465ACN 121995465 ACN121995465 ACN 121995465ACN-121995465-A

Abstract

The invention discloses a self-adaptive decomposition and time-frequency characteristic parameter extraction method for a seismic signal, and belongs to the technical field of geophysical signal processing. The method comprises preprocessing of earthquake signals, 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 adaptive time-frequency dynamic feature joint analysis. By introducing an attention mechanism, fusion and significance weight self-learning of time-frequency domain and time-frequency characteristics of different modes and different scales are realized, and the earthquake motion characteristic parameters with high expressive force are formed. The method is suitable for the decomposition, identification and anomaly detection of the seismic signals, and has good adaptivity, discrimination capability and interpretability.

Inventors

  • GONG MAOSHENG
  • ZHOU BAOFENG

Assignees

  • 中国地震局工程力学研究所

Dates

Publication Date
20260508
Application Date
20260120

Claims (9)

  1. 1. A method for adaptively decomposing earthquake signals and extracting time-frequency characteristic parameters is characterized by comprising the following steps: Preprocessing the earthquake motion signal, namely preprocessing the collected original earthquake motion signal, specifically comprising denoising, normalization and trend correction so as to ensure the accuracy of subsequent decomposition and feature extraction; Based on the self-adaptive modal decomposition of the entropy weight residual error, on the basis of the preprocessed signals, adopting an entropy weight residual error self-adaptive decomposition algorithm to decompose the complex earthquake motion signals into a plurality of modal components which are uncorrelated with each other and have different energy distribution, and after each round of decomposition, dynamically setting parameters and termination strategies of the next decomposition according to the current decomposition effect by calculating the component energy entropy and the information entropy of the residual error signals, thereby realizing the self-adaptive decomposition highly matched with the complexity of the signals; after modal decomposition is completed, carrying out multi-level time domain feature extraction on the original signal and modal components obtained by each decomposition respectively, wherein the multi-level time domain feature extraction comprises peak acceleration, peak speed, peak displacement, effective duration, extremum quantity and energy indexes; the method comprises the steps of performing frequency domain feature coding of multi-scale convolution dictionary learning, and performing frequency domain feature coding by adopting a multi-scale convolution dictionary learning algorithm aiming at each modal component; By the combined extraction of the instantaneous frequency, the local energy and the time-frequency energy moment time-frequency dynamic characteristics, the complex time-frequency evolution of the seismic signal is depicted, and a basic support is provided for rapidly detecting the seismic abnormal event and atypical waveforms; After multi-mode multi-scale feature saliency modeling and fusion are completed, attention mechanism is introduced to design a saliency weight self-learning algorithm, and intelligent fusion is carried out on three types of features of time domain, frequency domain and time frequency and features of different modes and different scales.
  2. 2. The method for extracting the self-adaptive decomposition and the time-frequency characteristic parameters of the earthquake motion signal according to claim 1, wherein the preprocessing of the earthquake motion signal comprises the steps of self-adaptive threshold wavelet denoising, empirical mode decomposition-residual autoregressive hybrid denoising, normalization and trend correction.
  3. 3. The adaptive decomposition and time-frequency characteristic parameter extraction method of earthquake motion signal according to claim 1, wherein the adaptive modal decomposition based on entropy weight residual error comprises performing layer-by-layer decomposition by adopting entropy weight residual error adaptive decomposition algorithm, wherein the decomposition process is based on variational modal decomposition or empirical modal decomposition, respectively calculating energy entropy and information entropy after each modal component and corresponding residual error signal are obtained, measuring energy dispersibility of the modal component by the energy entropy and measuring complexity of the residual error signal by the information entropy; and automatically judging the decomposition termination according to a preset threshold condition, realizing the self-adaptive adjustment and progressive decomposition of the decomposition control parameters, obtaining mutually independent modal components with actual physical significance, and improving the basic accuracy and the expressive force of multi-scale feature extraction.
  4. 4. The method for extracting the time-domain characteristic parameters of the adaptive decomposition of the earthquake signal according to claim 1, wherein the multi-level time-domain characteristic extraction comprises the steps of independently extracting peak acceleration, speed components, displacement components, local extremum point number, effective duration and total energy time-domain characteristic indexes from each component according to each modal component obtained by the adaptive decomposition and an original signal, adopting an energy entropy weight weighting fusion strategy, and designing the duty ratio of energy of each modal component in total energy and energy entropy jointly into fusion weights, and carrying out weighted integration on time-domain characteristic quantities of each modality to form a global time-domain characteristic set.
  5. 5. The adaptive decomposition and time-domain feature parameter extraction method of seismic signals of claim 1, wherein the multi-scale convolution dictionary learning frequency domain feature coding comprises designing a multi-scale convolution kernel dictionary to cover different time scales and frequency intervals for each modal component obtained by multi-level time domain decomposition, carrying out one-dimensional convolution on the modal signals and a plurality of convolution kernels to obtain a sparse activation sequence, and jointly learning the convolution kernels and sparse activation by an alternate optimization strategy; The frequency response of the convolution kernel is obtained by adopting Fourier transformation, so that characteristic parameters of main frequency, center frequency and frequency bandwidth are extracted, and the activation sequence of each mode under all convolution kernels and the characteristics are spliced and encoded to form a frequency domain characteristic vector with a compact structure, enrich the spectrum expression of the seismic signal and provide a basis for the subsequent seismic type differentiation and seismic source mechanism analysis.
  6. 6. The method for extracting the adaptive decomposition and time-frequency characteristic parameters of the earthquake motion signal according to claim 1, wherein the time-frequency dynamic characteristic joint extraction comprises the steps of adopting an instantaneous frequency and local energy analysis method of an adaptive window function, and carrying out time-frequency analysis on the width adaptation of a signal stationary section and a transient section by dynamically adjusting the length and shape parameters of the analysis window function; the instantaneous frequency, local energy and time-frequency energy moment characteristics are calculated by utilizing short-time Fourier transform or continuous wavelet transform, instantaneous main frequency, local energy peak value and energy center track of each modal component are respectively obtained, and the time-frequency dynamic characteristics are jointly encoded to form a high-dimensional dynamic characteristic set for describing the time-frequency structure evolution of the seismic signal, so that the sensitivity and accuracy of the detection and the distinction of the seismic event are improved.
  7. 7. The method for extracting the time-frequency characteristic parameters by adaptively decomposing the earthquake signals according to claim 1, wherein the modeling and the fusion of the multi-mode multi-scale characteristic significance are characterized in that the obtained multi-level cross-mode characteristic set comprises time domain characteristics, frequency domain characteristics and time-frequency dynamic characteristics, and the multi-mode multi-scale feature significance modeling and the fusion are unified, normalized and spliced in a high-dimensional manner to form an original characteristic tensor; the significance weight self-learning framework based on the attention mechanism is introduced, the weight mapping is carried out on the features of different modes, different categories and different scales through the trainable parameters, and the loss function related to the seismic event discrimination is minimized by adopting an end-to-end training mode, so that the feature significance automatic modeling is realized.
  8. 8. The method for extracting the time-frequency characteristic parameters of the adaptive decomposition of the earthquake motion signal according to claim 2, wherein the adaptive threshold wavelet denoising is used for suppressing high-frequency noise components, the empirical mode decomposition-residual autoregressive hybrid denoising is used for removing non-stationary noise and residual artifacts, the normalization is used for normalizing the signal amplitude to zero mean unit variance, and the trend correction is used for correcting signal drift or baseline offset by a least square method.
  9. 9. The method for extracting the time-consuming feature parameters of the adaptive decomposition of the seismic signals according to claim 7, wherein the method is characterized in that a grouping attention, layering attention or self-attention strategy is adopted in the fusion process, multi-level feature weighting is realized on inter-mode fusion, intra-scale aggregation and global redistribution, and finally an overall significance feature vector is obtained, so that a fusion feature basis with discrimination capability and interpretability is provided for seismic signal identification and classification.

Description

Adaptive decomposition and time-frequency characteristic parameter extraction method for earthquake signals Technical Field The invention belongs to the technical field of geophysical signal processing, and particularly relates to a method for adaptively decomposing a seismic signal and extracting a time-frequency characteristic parameter. Background In recent years, with the continuous development of earthquake monitoring and early warning systems, efficient and accurate feature extraction of earthquake strong vibration signals becomes an important research direction in the fields of earthquake engineering, geophysics and the like. The existing seismic signal feature extraction technology mainly comprises time domain feature analysis, frequency domain feature analysis, time-frequency transformation, an automatic feature extraction method based on machine learning and the like. The time domain and frequency domain analysis method is widely applied due to the advantages of simple principle, high calculation efficiency and the like, but in the actual earthquake signal processing process, the traditional time domain and frequency domain method generally shows limited adaptability in the face of the phenomena of high nonlinearity, non-stationary mutation, multi-scale complex evolution and the like of signals, and the non-stationary and burst characteristics in the earthquake signals are difficult to accurately capture and describe. Although wavelet transformation, hilbert-Huang transformation and the like of the time-frequency analysis method can reveal the local time-frequency structure of the signal to a certain extent, the time-frequency analysis method is difficult to have good self-adaptive capacity for actual seismic signals due to dependence on preset or selected basis functions, and has limitations on the diversity of seismic signal components and the decomposition and characterization of extremely atypical waveforms. In the field of earthquakes, marked samples are limited, extremely abnormal earthquake event samples are rare, so that the methods are insufficient in generalization capability in the presence of new types or rare earthquake signals, and the feature extraction process is highly black and difficult to provide enough interpretability, thereby influencing the reliable application of the method in actual earthquake monitoring and early warning. Therefore, how to combine the complexity, the self-adaptability and the feature expression effectiveness of the seismic signals, and to construct a novel seismic signal analysis method which has high precision, can be interpreted and has the self-adaptive feature extraction capability, is one of the technical problems to be solved currently. Disclosure of Invention The invention aims to solve the technical problems of limited self-adaption capability, insufficient feature expression, lack of interpretability in a feature extraction process and the like of the traditional seismic signal feature extraction method when dealing with nonlinear, non-stationarity and complex multi-scale evolution of signals, and provides a method capable of efficiently and adaptively decomposing seismic signals, extracting multi-level, multi-scale and time-frequency dynamic features in a combined mode and improving the precise characterization and recognition capability of complex features of the seismic signals. In order to achieve the purpose, the invention is realized by adopting the following technical scheme that the method comprises the following steps: Preprocessing the earthquake motion signal, namely preprocessing the collected original earthquake motion signal, specifically comprising denoising, normalization and trend correction so as to ensure the accuracy of subsequent decomposition and feature extraction; Based on the self-adaptive modal decomposition of the entropy weight residual error, on the basis of the preprocessed signals, adopting an entropy weight residual error self-adaptive decomposition algorithm to decompose the complex earthquake motion signals into a plurality of modal components which are uncorrelated with each other and have different energy distribution, and after each round of decomposition, dynamically setting parameters and termination strategies of the next decomposition according to the current decomposition effect by calculating the component energy entropy and the information entropy of the residual error signals, thereby realizing the self-adaptive decomposition highly matched with the complexity of the signals; after modal decomposition is completed, carrying out multi-level time domain feature extraction on the original signal and modal components obtained by each decomposition respectively, wherein the multi-level time domain feature extraction comprises peak acceleration, peak speed, peak displacement, effective duration, extremum quantity and energy indexes; the method comprises the steps of performing frequency domain feature codin