CN-121502240-B - Vibration signal space-time reconstruction method based on multi-mode conditional diffusion model

CN121502240BCN 121502240 BCN121502240 BCN 121502240BCN-121502240-B

Abstract

The invention provides a vibration signal space-time reconstruction method based on a multi-mode conditional diffusion model, which relates to the technical field of vibration signal reconstruction, and comprises the steps of firstly, utilizing a multi-sensor to collect structural vibration response, constructing a multi-dimensional vibration signal matrix, and automatically identifying a space continuous missing and time random missing region; and further applying a pseudo-missing mask on the complete data, constructing a training sample through a self-supervision strategy, and guiding the model to learn the space-time related features and the missing mode. In the training stage, a diffusion model is used for generating a frame, gaussian noise disturbance is applied to a missing region, four types of conditions of time, space, trend and frequency domain are introduced in the denoising inversion process, the periodicity of signals, multi-sensor space coupling, low-frequency change and physical spectrum structure are respectively described, and high-fidelity signal reconstruction under the multi-mode information joint constraint is realized.

Inventors

WANG CHENG
ZHOU ZIBO
ZHANG YIWEN
LIN HUANGXING
FAN ZONGWEN
Gou jin
DENG JIAN
SU HANBIN
ZHANG CHENYU

Assignees

华侨大学
先峰时空(厦门)信息技术研究院有限公司

Dates

Publication Date: 20260508
Application Date: 20260113

Claims (9)

1. A vibration signal space-time reconstruction method based on a multi-mode conditional diffusion model is characterized by comprising the following steps: The method comprises the steps of obtaining a missing vibration signal acquired by a sensor component arranged on an engineering structure, inputting the vibration signal into a pre-trained diffusion model, performing space-time reconstruction processing to obtain a complete multi-sensor vibration signal matrix, wherein a missing region of the vibration signal is filled with Gaussian noise and then is used as initial input, and the amplitude and time sequence characteristics of the missing region signal are gradually recovered through iterative denoising inversion under multi-mode condition embedding constraint; carrying out quantitative evaluation and physical verification on the complete multi-sensor vibration signal matrix to ensure that the reconstructed signal accords with the structural dynamics rule; Before the trained diffusion model is called, gaussian noise filling is carried out on the signals of the missing region, the signals are used as an initial state of denoising inversion, multi-mode condition features are extracted, and the multi-mode condition features comprise frequency domain feature embedding, physical spectrum index and light convolution are combined to extract a frequency domain structure.
2. The vibration signal space-time reconstruction method based on the multi-mode conditional diffusion model according to claim 1, wherein the vibration signal is input into a pre-trained diffusion model for space-time reconstruction processing to obtain a complete multi-sensor vibration signal matrix, specifically comprising the following steps: the method comprises the steps of acquiring a missing vibration signal acquired by a sensor component arranged on an engineering structure, gradually adding Gaussian noise into the missing vibration signal, and filling the Gaussian noise; in the process of reverse denoising, a pre-trained diffusion model is adopted to predict the noise after filling, and denoising treatment is carried out on the noise so as to obtain a finally reconstructed vibration signal X recon ; Performing inverse normalization processing on the reconstructed vibration signal X recon , and recovering the signal to the physical dimension of the engineering structure to obtain a complete multi-sensor vibration signal matrix , wherein, In order to represent the multiplication by element, For the standard deviation of each channel of the original signal before normalization, Is the mean value of each channel of the original signal before normalization.
3. The vibration signal space-time reconstruction method based on the multi-modal conditional diffusion model according to claim 1, wherein the complete multi-sensor vibration signal matrix is quantitatively evaluated and physically verified to ensure that the reconstructed signal conforms to the structural dynamics law, specifically: Performing quantization evaluation and physical verification on the complete multi-sensor vibration signal matrix, and calculating mean square error, average absolute error and cos similarity indexes; And quantifying the reconstruction precision and fitting effect, analyzing the frequency spectrum characteristic and the vibration mode of the signal, and ensuring that the reconstructed signal accords with the structural dynamics rule.
4. The vibration signal space-time reconstruction method based on the multi-modal conditional diffusion model as claimed in claim 1, the method is characterized by further comprising the following steps before the trained diffusion model is called: The method comprises the steps of acquiring vibration signals acquired by a sensor component arranged on an engineering structure, preprocessing the vibration signals, and extracting to obtain an original vibration signal matrix, wherein the preprocessing comprises denoising, normalization and missing marking; Performing preliminary reconstruction on a missing region in the detected original vibration signal matrix by adopting a self-adaptive multi-scale interpolation method, and synthesizing three constraint information of time, frequency and space to generate a rough reconstruction signal; Applying an artificial deletion mask to the original vibration signal to simulate space continuous deletion and time random deletion, recording the deletion position and the preliminary interpolation result, taking a pseudo-deletion sample and a corresponding complete signal as input, and separating real data for training and pseudo-deletion signals for verification; Carrying out Gaussian noise filling on the signals of the missing region as an initial state of denoising inversion, extracting multi-modal condition features, fusing four types of features to generate side information of a diffusion model, inputting the side information of the diffusion model for guiding denoising reconstruction of the diffusion model so as to enable the model to learn a space-time distribution rule of the missing signals in self-supervision training, wherein the multi-modal condition features further comprise time feature embedding and capturing time sequence dependence and periodic information, spatial feature embedding and extracting correlations among sensors, and trend feature embedding and capturing low-frequency trends; the coarse reconstruction signals and the fused multi-mode condition features are input into a diffusion model, the noise and inversion are gradually carried out under a self-supervision training frame, the training is carried out, parameters are continuously optimized through the pseudo-missing samples until the model reaches the preset requirement, the training is finished, and the trained diffusion model is obtained.
5. The vibration signal space-time reconstruction method based on the multi-modal conditional diffusion model of claim 4, wherein the vibration signals collected by the sensor assemblies arranged on the engineering structure are obtained and preprocessed, and the original vibration signal matrix is extracted, specifically: acquiring multichannel vibration signals acquired by a plurality of acceleration sensors or displacement sensors distributed on an engineering structure, splicing the acquired multichannel vibration signals, and constructing a vibration signal matrix containing all sensor channel vibration signals Wherein M is the number of sensors, and T is the time sequence length; Automatically identifying a missing region existing in the vibration signal matrix, wherein the missing region comprises continuous space missing and random time missing, marking a missing part in the vibration signal matrix by using a NaN value, and constructing a missing mask matrix M observe used for marking whether valid data exist in each sensor at each time point, wherein a non-missing position in the missing mask matrix is marked as 1, and the missing position is marked as 0; calculating the mean value of non-missing data of each sensor channel, and filling missing positions with the mean value to obtain a preliminary complete signal matrix X filled ; The signal matrix after filling is standardized, the standard deviation of each channel is calculated, the data with the standard deviation smaller than the threshold value is subjected to zero removal prevention treatment, and the formula is adopted Calculating the data of each channel to generate a normalized signal matrix And obtaining an original vibration signal matrix.
6. The vibration signal space-time reconstruction method based on the multi-mode conditional diffusion model according to claim 5, wherein the method is characterized in that a self-adaptive multi-scale interpolation method is adopted to preliminarily reconstruct a missing region in a detected original vibration signal matrix, three constraint information of time, frequency and space are synthesized, and a rough reconstruction signal is generated, specifically: for signal matrix Performing frequency domain analysis, and performing frequency domain conversion on signals of each sensor channel through fast Fourier transform to obtain a complex frequency spectrum matrix And extracting statistical features of amplitude spectrum and energy spectrum from the complex frequency spectrum matrix to construct a frequency domain feature matrix , Mapping operators for frequency domain features; for signal matrix Performing window division processing, and extracting time tensors in each window Mask tensor Frequency domain tensor K is the number of sensor channels, L is the window length, Is the frequency domain feature dimension; In the space dimension, the missing data is compensated based on weighting and smoothing of adjacent sensors, three constraint information of time, frequency and space are integrated, and a rough reconstruction signal is generated.
7. The vibration signal space-time reconstruction method based on a multi-modal conditional diffusion model according to claim 6, wherein an artificial deletion mask is applied to an original vibration signal to simulate space continuous deletion and time random deletion, and the deletion position and the preliminary interpolation result are recorded, and a pseudo-deletion sample and a corresponding complete signal are used as inputs to separate out real data for training and pseudo-deletion signals for verification, specifically: for signal matrix Performing missing detection and passing through mask matrix Representing the signal matrix at the current mask matrix when the value is 0 Is missing, when the value is 1, the signal matrix is represented in the current mask matrix Is not missing, dividing the acquired raw data into missing signal subsets And non-deletion signalers The two parts are collected together and the two parts are arranged, A mask matrix that is a non-missing signal; masking the non-missing signal subset again to obtain pseudo-missing signal Calculating a loss function for subsequent verification, Is a pseudomiss mask.
8. The vibration signal space-time reconstruction method based on the multi-modal conditional diffusion model according to claim 7, wherein gaussian noise filling is performed on the signals of the missing region, the multi-modal conditional characteristics are extracted as an initial state of denoising inversion, and four types of characteristics are fused to generate side information input of the diffusion model, specifically: for signal matrix Initializing the missing part of the model (C) by using zero-average Gaussian noise to obtain an initial input matrix of denoising inversion, and taking the initial input matrix as a step-by-step denoising of a diffusion model to be trained to provide a random starting point; Extracting the time characteristics of the signal matrix in each window For each time point t, constructing a position coding vector PE (t) to represent time or phase information and adopting sine and cosine to code, wherein the formula is as follows , , For each window or whole sequence, construct its time-embedded tensor, where dt is the time-embedded dimension Injecting a diffusion model as a time condition; extracting spatial features of the signal matrix within each window The method comprises the steps of designing a graph convolution network to capture the spatial correlation and local coupling characteristics among multiple sensor vibration signals, extracting local spatial response through depth separable convolution, wherein the depth convolution independently extracts time sequence modes in each channel, and the point-by-point convolution realizes linear mixed mapping among channels to obtain , As a function of the non-linear activation, For the purpose of batch normalization, For the signal matrix within each window, In order to be a convolution feature, For passing through a one-dimensional convolution layer once, the convolution kernel size is 1 multiplied by 1; construction of an adjacency matrix based on a sensor spatial layout The formula of the ijth adjacency matrix is , The power of e is represented by, For the spatial position of the i-th sensor, For the spatial position of the jth sensor, Is a scale parameter; the neighborhood information diffusion is realized through graph convolution smoothing, and the formula is as follows: , The final spatial embedded representation is obtained by fusing the convolution characteristic and the graph smoothing characteristic , Is a weight matrix; extracting trend characteristics of signal matrix in each window Applying a one-dimensional convolution to the time series of each sensor to perform local smoothing to obtain smoothed signals The formula is And smoothing the smoothed signal Input unidirectional GRU module for extracting global time trend T is a transposition, and the global time trend is subjected to linear mapping and nonlinear activation function processing to obtain final trend embedding , As a matrix of weights, the weight matrix, The bias vector is a variable parameter in the training process; extracting frequency domain features of signal matrix in each window For frequency domain feature matrix Extraction of key physical frequency domain features based on amplitude spectrum, including spectrum centroid Significance of main peak And intermediate frequency energy concentration , And The first two peaks of the amplitude spectrum are respectively, For the amplitude at frequency F, F represents the upper limit of the sampling frequency; splicing the spectrum centroid, main peak significance and intermediate frequency energy concentration degree to obtain the spliced characteristics Preprocessing and extracting the amplitude spectrum through a lightweight convolutional neural network, and extracting high-dimensional spectrum characteristics And obtaining the appointed dimension through linear mapping On, obtain the frequency domain embedding The characteristic F phys and the CNN characteristic are fused through gating to obtain the final frequency domain characteristic representation of each sensor in the window The frequency domain constraints are provided as a diffusion model, , For gating the fused weight, i is the number of frequency sub-bands obtained by slicing twice on the time sequence of one L, A vector is embedded for the frequency domain features in the ith frequency domain subband in the window, For the feature vector of physical information in the ith frequency domain subband in any window, Is the feature vector extracted after one-dimensional convolution in the ith frequency domain sub-band in any window, For the multiplication operator, To straighten a multi-dimensional vector into a one-dimensional vector, Is one-dimensional pooling operation; In order to be amplitude-value, Representing the number of extracted frequency domain features within a window; Time characteristics Spatial characteristics Trend characteristics Frequency domain features Splicing to obtain side information input of diffusion model 。
9. The vibration signal space-time reconstruction method based on the multi-modal conditional diffusion model according to claim 8, wherein the coarse reconstruction signal and the fused multi-modal conditional characteristics are input into the diffusion model, the noise and inversion are gradually removed under the self-supervision training frame, the training is performed, the parameters are continuously optimized through the pseudo-missing samples until the model reaches the preset requirement, the training is finished, and the trained diffusion model is obtained, specifically: in the model training process, based on the pseudo-missing signal, gaussian noise is gradually added through a forward diffusion process, and the formula is as follows , wherein, In order to accumulate the noise attenuation coefficient, For the noise intensity of time step i, Gaussian noise conforming to a standard normal distribution; In the back-diffusion phase, through a conditional network And corresponding time step information to predict noise to obtain the noise predicted by the model after the condition information fusion, wherein the formula is as follows Model predicted noise after condition information fusion Predicting missing signals by the formula An immediate estimate of the original noise-free signal by the model is obtained, wherein, For noisy vibration signal data at time step t, In order to denoise the noise predicted by the process model, Based on current noise data And predicted noise Estimated raw data; in the iterative process, the joint noise loss function and the frequency domain loss function are adopted to optimize, and the formula is that , , , And Are all the weight coefficients of the two-dimensional space model, As a function of the total loss, As a function of the loss of noise, As a function of the loss in the frequency domain, Gaussian noise for time step t and meeting standard normal distribution The mathematical expectation under the influence is that, For the purpose of mathematical expectations, To fourier transform the response signal to a frequency, Is the square of the L2 norm.

Description

Vibration signal space-time reconstruction method based on multi-mode conditional diffusion model Technical Field The invention relates to the technical field of vibration signal reconstruction, in particular to a vibration signal space-time reconstruction method based on a multi-mode conditional diffusion model. Background In the field of structural health monitoring and vibration signal analysis, obtaining complete and continuous multi-sensor vibration response data is a basis for carrying out modal parameter identification, structural state evaluation and damage diagnosis. However, in practical engineering applications, it is often difficult to deploy a sufficient number of sensors at all critical locations, or maintain stable data acquisition during long-term monitoring, due to factors such as harsh environmental conditions, complex structural deployment, or insufficient reliability of the sensors themselves. For example, in complex scenes such as deep foundation pits, large-span bridges, offshore wind power towers or high-rise buildings, partial areas may have insufficient sensor arrangement density or signal acquisition interruption due to limited space, severe temperature and humidity changes, strong electromagnetic interference or severe installation conditions, so that the loss of structural vibration signals in time and space dimensions is caused. Such a lack may not only manifest as a random interruption of the single sensor over time sequence, but also as a continuous spatial lack of multiple sensors over a certain period of time, severely affecting the accuracy and reliability of subsequent structural modal identification, frequency domain analysis, and security assessment. The conventional data interpolation or filtering method is difficult to recover global trend, local characteristics and space-time correlation of signals at the same time, and particularly has obvious defects of reconstruction accuracy and generalization capability in a multi-sensor high-dimensional signal scene. For dynamic monitoring of complex engineering structures, the integrity of vibration signals directly affects the accuracy of subsequent modal identification, damage diagnosis and structural health assessment. Currently, the reconstruction processing method for the problem of vibration signal missing or abnormal mainly includes traditional signal processing technologies such as interpolation-based time domain reconstruction, filtering smoothing, low-rank decomposition, wavelet transformation and the like, and deep learning methods gradually introduced in recent years, such as generation of a countermeasure network (GAN), compressed Sensing (CS) and the like. Although the time domain interpolation method can restore certain signal continuity in a local missing area, the global trend of the signal and the spatial correlation between multiple channels are difficult to maintain, the filtering and frequency domain reconstruction method improves the signal smoothness to a certain extent, but high-frequency details and instantaneous modal characteristics are easy to lose, the reconstruction precision is limited when the low-rank decomposition and wavelet transformation are faced with large-scale missing, space continuous missing or random time missing, and the method is difficult to adapt to a complex signal structure of high-dimensional and multiple channels. Although the depth learning method improves the authenticity and continuity of the reconstructed signal to a certain extent and improves the recovery effect of local deficiency, the modeling capability of the existing GAN model on complex space-time dependence is still insufficient, the training process is unstable and the adaptability to the deficiency mode is poor, while the compressed sensing method realizes reconstruction under the undersampling condition by utilizing the signal sparsity and has certain advantages on the underdetermination problem, the reconstruction precision and the robustness of the structural response signal with obvious dynamic change or complex multi-mode characteristics still need to be improved. In a comprehensive view, the existing method is difficult to reliably reconstruct space-time vibration signals under a complex structure on the premise of keeping global trend, local detail and multi-sensor spatial correlation of signals. In view of this, the present application has been proposed. Disclosure of Invention The invention provides a vibration signal space-time reconstruction method based on a multi-mode conditional diffusion model, which can at least partially improve the problems. In order to achieve the above purpose, the present invention adopts the following technical scheme: a vibration signal spatiotemporal reconstruction method based on a multi-modal conditional diffusion model, comprising: The method comprises the steps of obtaining a missing vibration signal acquired by a sensor component arranged on an engineering structure,