Search

CN-121997759-A - Diffusion model-based multi-condition earthquake motion generation method

CN121997759ACN 121997759 ACN121997759 ACN 121997759ACN-121997759-A

Abstract

The invention relates to the technical field of earthquake engineering, in particular to a multi-condition earthquake motion generation method based on a diffusion model, which comprises the steps of firstly extracting multidimensional condition parameters such as a reaction spectrum, peak acceleration, arias intensity, energy arrival time, obvious duration, husid curve and the like from an earthquake motion acceleration time course, and carrying out normalization treatment; the method comprises the steps of converting acceleration time course into normalized logarithmic magnitude spectrum through short-time Fourier transform to serve as a learning target, constructing a transducer condition encoder to capture coupling relation of multi-source conditions, designing a U-Net-based denoising network and integrating a cross attention mechanism to achieve accurate condition injection, optimizing a noise prediction network by adopting a denoising diffusion probability model training strategy, adopting a denoising diffusion implicit model sampling strategy during reasoning, and carrying out iterative reconstruction of phase information from amplitude spectrum through a Griffin-Lim algorithm and carrying out inverse transformation to obtain the earthquake acceleration time course. The invention realizes the cooperative and accurate control of the response spectrum and the energy non-stationary characteristic, has stable training and high reasoning efficiency, and supports the generation of diversity.

Inventors

  • DING YI
  • HU JINJUN
  • XIE LILI
  • CHEN SU
  • LI XIAOJUN

Assignees

  • 江汉大学

Dates

Publication Date
20260508
Application Date
20260129

Claims (9)

  1. 1. The multi-condition earthquake motion generating method based on the diffusion model is characterized by comprising the following steps of: The method comprises the steps of preprocessing data, namely acquiring a seismic acceleration time course data set, and preprocessing the seismic acceleration time course data set by unifying sampling rate and time length, extracting multidimensional condition parameters, normalizing the condition parameters, converting the acceleration time course into a time spectrum chart through short-time Fourier transformation, and carrying out logarithmic compression and normalization on an amplitude spectrum; Training a condition encoder, namely constructing the condition encoder and constructing three layers of converterlers, wherein each layer of converters comprises a multi-head self-attention module and a position feedforward network; the method comprises the steps of training a denoising diffusion model, constructing a denoising network based on U-Net, integrating a cross attention mechanism in an attention block of the U-Net to realize condition information injection, performing time step coding by adopting sine position coding, and optimizing a noise prediction network by adopting a denoising diffusion probability model DDPM training strategy; and step four, reasoning and generating earthquake motion acceleration time course.
  2. 2. The diffusion model-based multi-conditional seismic motion generation method of claim 1, wherein the multi-dimensional conditional parameters in step one comprise scalar intensity indicators comprising peak acceleration PGA, arizons intensity I A , 5% energy arrival time T 5 and significant hold time D 5-95 and vector parameters comprising a reaction spectrum and normalized Husid energy accumulation curve.
  3. 3. The method for generating multi-condition earthquake motion based on diffusion model as recited in claim 2, wherein the peak acceleration PGA is the maximum absolute value of acceleration time course, aldrich strength The 5% energy arrival time T 5 is the time when the accumulated energy reaches 5%, the significant duration D 5−95 is the time interval from 5% accumulated energy to 95% accumulated energy, the reaction spectrum is the acceleration reaction spectrum of a plurality of periodic points, and the energy accumulation curve of Husid is normalized 。
  4. 4. The diffusion model-based multi-condition earthquake motion generation method according to claim 1, wherein the normalization processing of the condition parameters in the first step is specifically performed by mapping the peak acceleration PGA and the arizon intensity I A to the [ -1, 1] interval by using an empirical cumulative distribution function transformation after taking logarithms, linearly normalizing the energy arrival time T 5 and the duration time D 5-95 to the [ -1, 1] interval, normalizing the reaction spectrum by using the peak acceleration per sample, and normalizing the Husid energy accumulation curve definition to be [0, 1] directly.
  5. 5. The diffusion model-based multi-conditional earthquake motion generation method of claim 1, wherein the condition encoder comprises a scalar condition projection module, a vector condition blocking module and a position encoding module, wherein the scalar condition projection module maps each scalar condition into a single condition token through a multi-layer perceptron, the vector condition blocking module divides a reaction spectrum and Husid curves into a plurality of patches respectively, each patch is mapped into the condition token through the multi-layer perceptron, and the position encoding module adds sinusoidal position encoding to a token sequence of vector conditions.
  6. 6. The diffusion model-based multi-conditional earthquake motion generation method of claim 1, wherein the step three specifically comprises the steps of: S3.1, constructing a U-Net-based denoising network, wherein the denoising network comprises an encoder, a bottleneck layer and a decoder, and realizing multi-scale feature fusion through jump connection; S3.2, integrating a cross attention mechanism in an attention block of the U-Net, namely flattening the time-frequency spectrogram characteristic into a sequence form, and linearly projecting the sequence to obtain a query matrix Q, linearly projecting a token sequence output by a condition encoder to obtain a key matrix K and a value matrix V, calculating a scaling dot product attention to obtain attention weight for the query matrix Q and the key matrix K, and weighting and summing the value matrix by using the attention weight to obtain attention output; S3.3, time step coding, namely converting a diffusion time step t into a high-dimensional embedded vector by adopting sine position coding, and fusing the high-dimensional embedded vector with U-Net characteristics after MLP projection; S3.4, training strategies of a denoising diffusion probability model DDPM are adopted; And S3.5, smoothing model parameters by adopting an exponential moving average technology, wherein the attenuation rate is set to be 0.999.
  7. 7. The diffusion model-based multi-conditional earthquake motion generation method of claim 6, wherein the operation of S3.4 is as follows: Setting the diffusion step number t=1000, and linearly increasing the noise schedule β t from 10 −4 to 0.02; for each training sample, randomly sampling a time step t and a standard gaussian noise epsilon; According to the formula Calculating a time-frequency spectrogram after noise addition, x t is noise-containing data of the t step, alpha t =1−β t defines the noise addition rate, Representing the cumulative product of the noise schedule up to step t; Training a denoising network epsilon θ (x t , t, c) predicting noise epsilon, c being a conditional code; The optimization objective is to minimize the mean square error of the predicted noise and the real noise: 。
  8. 8. The diffusion model-based multi-conditional earthquake motion generation method of claim 1, wherein the step four specifically comprises: S4.1, encoding given target condition parameters into a condition token sequence through a condition encoder; S4.2, fast reasoning is carried out by adopting a denoising diffusion implicit model DDIM sampling strategy, namely, starting from standard Gaussian noise x T -N (0,I), wherein I is an identity matrix to represent isotropic Gaussian noise, setting the sampling step number to be 50 steps, uniformly sampling a time step sequence, and calculating x t−1 according to predicted noise epsilon θ (x t , t and c) for each time step: σ t controls the randomness of each step, z-N (0,I) is standard Gaussian noise independent of x t ; s4.3, carrying out inverse normalization processing on the generated normalized logarithmic magnitude spectrum to restore the magnitude spectrum; S4.4, iteratively reconstructing phase information from the amplitude spectrum by adopting Griffin-Lim algorithm, wherein the method comprises initializing a random phase spectrum Then, performing iteration, and obtaining a final phase spectrum by iterating 128 times; S4.5, combining the amplitude spectrum and the reconstructed phase spectrum, and obtaining the generated earthquake motion acceleration time course through inverse STFT transformation.
  9. 9. The diffusion model-based multi-conditional earthquake motion generation method of claim 8, wherein the iterative execution in S4.4 is sequentially performed by constructing a complex spectrum Inverse STFT to obtain time domain signal x and re-STFT to obtain new phase 。

Description

Diffusion model-based multi-condition earthquake motion generation method Technical Field The invention relates to the technical field of seismic engineering, in particular to a multi-condition earthquake motion generation method based on a diffusion model. Background Performance-based seismic design and evaluation typically requires nonlinear time-course analysis with the aid of a large number of seismic records to accurately evaluate the dynamic response and risk of damage of the structure under different seismic actions. In engineering practice, for major hydropower engineering, a site-related designed seismic response spectrum is required, namely, a set seismic method is adopted to determine the vibration parameters of an engineering site. Although the high-intensity seismic databases at home and abroad have already recorded a large number of seismic records, their distribution in the parameter space is often very unbalanced. For a specific scene of interest in engineering practice, such as a large earthquake magnitude, a near-field condition, a rare earthquake under a specific field category, or when multiple intensity parameter control requirements need to be met at the same time, it is often difficult for the existing observation data to meet the direct use requirement. In recent years, the development of deep learning technology provides a brand new path for earthquake motion generation. The diffusion model is taken as a novel generation modeling framework which is paid attention to in recent years, has shown excellent performance in the fields of image and voice synthesis and the like, is more stable in training process compared with GAN, is easy to introduce multi-mode condition information, and the generated sample has diversity and high authenticity. However, the conventional earthquake motion generation method based on the diffusion model mainly depends on scalar parameters such as magnitude and distance as control conditions, and is difficult to accurately guide non-stationary characteristics such as energy time distribution. For nonlinear structural response analysis, earthquake motions with the same response spectrum will cause significant differences in the cumulative effects of structural damage if the energy evolution process is different. Therefore, there is a need for a seismic energy generation method capable of precisely controlling both the reaction spectrum and the energy time distribution characteristics. Disclosure of Invention The application aims to provide a multi-condition earthquake motion generation method based on a diffusion model, and aims to solve the problems in the prior art. The embodiment of the application provides a multi-condition earthquake motion generation method based on a diffusion model, which comprises the following steps of: The method comprises the steps of preprocessing data, namely acquiring a seismic acceleration time course data set, and preprocessing the seismic acceleration time course data set by unifying sampling rate and time length, extracting multidimensional condition parameters, normalizing the condition parameters, converting the acceleration time course into a time spectrum chart through short-time Fourier transformation, and carrying out logarithmic compression and normalization on an amplitude spectrum; Training a condition encoder, namely constructing the condition encoder and constructing three layers of converterlers, wherein each layer of converters comprises a multi-head self-attention module and a position feedforward network; the method comprises the steps of training a denoising diffusion model, constructing a denoising network based on U-Net, integrating a cross attention mechanism in an attention block of the U-Net to realize condition information injection, performing time step coding by adopting sine position coding, and optimizing a noise prediction network by adopting a denoising diffusion probability model DDPM training strategy; and step four, reasoning and generating earthquake motion acceleration time course. Preferably, the multi-dimensional condition parameters in the first step include a scalar intensity index including a peak acceleration PGA, an apremilast intensity I A, a 5% energy arrival time T 5, and a significant hold time D 5-95, and a vector parameter including a reaction spectrum and a normalized Husid energy accumulation curve. Preferably, the peak acceleration PGA is the maximum absolute value of the acceleration time course, the Alyas intensityThe 5% energy arrival time T 5 is the time when the accumulated energy reaches 5%, the significant duration D 5−95 is the time interval from 5% accumulated energy to 95% accumulated energy, the reaction spectrum is the acceleration reaction spectrum of a plurality of periodic points, and the energy accumulation curve of Husid is normalized。 Preferably, the normalization processing of the condition parameters in the first step specifically includes that the peak acc