Structural vibration response compression frequency enhancement vector quantization variational auto-encoder method

Abstract

The invention belongs to the field of structural vibration response compression and reconstruction, and discloses a frequency-enhanced vector quantization variational auto-encoder method for structural vibration response compression, which comprises the following steps of: acquiring structural vibration response; making a segmented data set; building an FEVQVAE model with a double-branch module and a frequency module; the sum of a time domain data reconstruction loss function, a frequency domain data reconstruction loss function and an inherent loss function of a vector quantization variational auto-encoder is used as a loss function of an FEVQVAE model, and the initialized FEVQVAE model is optimized by using gradient descent; and selecting an optimal model weight and storing network weight parameters to realize the compression and reconstruction of the vibration response of the engineering structure. According to the method, through the proposed double-branch module and the frequency module, the problem that some frequency components are inaccurate when the vector quantization variational auto-encoder method is directly used for engineering structure vibration response reconstruction is solved, and accurate reconstruction of the engineering structure vibration response under the high compression ratio is achieved.

Description

Technical Field

[0001] The invention belongs to the field of structural vibration response compression and reconstruction, and relates to a frequency enhanced vector quantization variational autoencoder (FEVQVAE) method for structural vibration response compression, which is used for fast compression and reconstruction of engineering structure vibration response. Background Art

[0002] Structural health monitoring systems need to collect multiple types of data, among which structural dynamic response is particularly important, such as acceleration. Structural dynamic response can reflect the dynamic response characteristics of the structure under different working conditions, which is of great significance for evaluating the overall performance and safety of the structure. However, with the increase in structural complexity, the number of sensors required also increases greatly. A real bridge often requires dozens of sensors for all-round data collection, and the collection of structural dynamic response usually has the characteristics of long-term high sampling rate. The generation of large-scale data brings challenges to data transmission and storage.

[0003] In recent years, deep learning algorithms have developed rapidly and are increasingly used to compress various data types, such as video and audio. In the field of civil engineering, deep neural networks have also been used to compress structural vibration responses. Some scholars have proposed and studied a monitoring data anomaly detection and data compression framework based on traditional deep autoencoders (AE). This method verifies the compression performance of AE on the acceleration response of the cables of a long-span suspension bridge. In the compression stage, the encoder encodes the original data into abstract features with a dimension much lower than the original data; in the reconstruction stage, the stored features and the corresponding cable labels are input into the decoder to obtain the reconstructed data. Experimental results show that when the compression ratio is 10, the reconstruction accuracy meets the engineering requirements, that is, the average correlation coefficient between the reconstructed structural vibration response and the original vibration response can reach 0.9.

[0004] The vibration response compression method based on AE compresses the original data into low-dimensional floating-point features. Although floating-point numbers take up more storage space than integers, this method can still achieve a higher compression ratio than the method based on compressed sensing. Since compressing the original data into low-dimensional integer features can significantly improve the compression ratio, and the vector quantization variational autoencoder (VQVAE) model proposed by some scholars can map the original data into low-dimensional integer features (AvdOord, O.Vinyals, Neural discrete representation learning, 2017.), a higher data compression ratio can be achieved based on VQVAE. Therefore, VQVAE has great potential in image (Z.Duan, M.Lu, J.Ma, Y.Huang, Z.Ma, F.Zhu, QARV: Quantization-Aware ResNet VAE for Lossy Image Compression, 2024.) and audio compression (C. AvdOord, Y.Li, FSC Lim, A.Luebs, O.Vinyals, T.C. Walters, Low Bit-rate Speech Coding with VQ-VAE and a WaveNet Decoder, 2019.) has been widely studied. However, there are still some challenges in directly applying VQVAE to the compression of the vibration response of engineering structures. Direct use of this method will lead to inaccurate frequency components of some vibration responses during reconstruction, thereby affecting the accuracy of the overall state assessment of the structure. It is urgent to improve this method to be suitable for the compression and reconstruction of the vibration response of engineering structures. Summary of the invention

[0005] The present invention provides a FEVQVAE method for compressing structural vibration response, aiming to improve the accuracy of VQVAE reconstruction of structural vibration response, so that the main frequency components of the reconstructed structural vibration response are accurate.

[0006] The technical solution of the present invention:

[0007] A frequency-enhanced vector quantized variational autoencoder method for compressing structural vibration response, the steps are as follows:

[0008] S1. Collect the vibration response of the specified measuring point of the structure;

[0009] S2, segmenting the vibration response of each channel of each sensor, merging all segmented data segments together, and then segmenting all data segments into data sets to form a training set, a validation set and a test set;

[0010] Segment the vibration response of each channel of each sensor:

[0011] S2.1. Divide the collected vibration responses into three data sets in proportion, namely, training set, validation set and test set. The training set and validation set are used for model training and model selection respectively, and the test set will be used to test the compression and reconstruction performance of the model for vibration responses.

[0012] S2.1.1, standardizing the vibration response in each data set generated in step S2.1 to reduce the learning difficulty of the model by limiting the amplitude range of the vibration response;

[0013] S3. Use these three datasets to train the proposed FEVQVAE model, and finally save the best weights of FEVQVAE;

[0014] S3.1. Using these three datasets to train and test FEVQVAE means randomly initializing a FEVQVAE, then using the training set and validation set constructed in step S2.1 to train and select the model, and finally saving the selected best model and weights;

[0015] The FEVQVAE model is composed of an encoder, a vector quantizer and a decoder. The detailed architecture of the encoder and the decoder are as follows: Figures 3 to 7 ;like Figure 3 As shown, by introducing sensor encoding into the encoder and decoder, the position information of the sensor is introduced into the FEVQVAE model, and one model is implemented to compress the vibration response of all sensors;

[0016] The encoder is composed of a one-dimensional convolution layer, an encoder module and a sensor code, and the detailed architecture is: input signal->one-dimensional convolution layer+sensor code->encoder module+sensor code->encoder module+sensor code->encoder module+sensor code->encoder module+sensor code->encoder module+sensor code->one-dimensional convolution layer+sensor code->D-dimensional feature; the vector quantizer is a two-dimensional matrix composed of multiple one-dimensional vectors; the decoder part is composed of a one-dimensional convolution, a decoder module and a sensor code, and the detailed architecture is:D-dimensional feature->one-dimensional convolution layer+sensor code->encoder module+sensor code->encoder module+sensor code->encoder module+sensor code->encoder module+sensor code->one-dimensional convolution layer+sensor code->reconstructed input signal;

[0017] The encoder module is composed of three dual-branch modules and a down-sampling one-dimensional convolution layer, and the detailed architecture is: dual-branch module->dual-branch module->dual-branch module->down-sampling one-dimensional convolution layer; the decoder module is composed of three dual-branch modules and an up-sampling one-dimensional transposed convolution layer, and the detailed architecture is: up-sampling one-dimensional convolution layer->dual-branch module->dual-branch module->dual-branch module;

[0018] like Figure 6 and 7 As shown, the frequency domain feature extraction capability of the encoder and decoder is enhanced through the proposed dual-branch module and frequency module;

[0019] The dual-branch module is as follows Figure 6 As shown, it consists of a time domain branch and a frequency domain branch, and features are fused through a star convolution structure, and residual connections are used to prevent the gradient in the model from disappearing; it is used to simultaneously extract time domain features and frequency domain features, and automatically fuse the extracted time domain and frequency domain features; the dual-branch module consists of a time domain branch and a frequency domain branch, and automatically fuses the time domain and frequency domain features through a star convolution structure. The time domain branch extracts time domain features by a one-dimensional convolution layer with a convolution kernel of 3, while the frequency domain branch extracts frequency domain features by a frequency module, and then the features of the two branches are each passed through a one-dimensional convolution layer, and the star convolution structure is used to fuse the features of the two branches, and finally the fused features are passed through a one-dimensional convolution layer with a convolution kernel of 1, and residual connections are used between the input and output;

[0020] The frequency module is as follows Figure 7 As shown in the figure, the rich frequency domain features used to extract vibration response are composed of two branches: one has a small receptive field and is used to extract local frequency domain features, namely, the features of a single resonance frequency band; the other has a larger receptive field and is used to capture a wider range of frequency domain features, namely, the features of multiple resonance frequency bands; subsequently, a star convolution structure is used to enable the two branches to adaptively interact and fuse, and retain the useful information of the two branches; the frequency module is composed of a single resonance frequency band branch and a multi-resonance frequency band branch, the features are first converted to the frequency domain using the fast Fourier transform, and then the features of the single resonance frequency band are extracted using a one-dimensional convolution layer, while the features of multiple resonance frequency bands are extracted using a one-dimensional hole convolution layer, and then the features of the two branches are each passed through a one-dimensional convolution, and the features of the two branches are fused using a star convolution structure, the fused features are passed through a one-dimensional convolution layer with a convolution kernel of 1, and then the inverse fast Fourier transform is used to transform the features from the frequency domain to the time domain, and the features after the fast Fourier transform and before the inverse fast Fourier transform are connected with a residual connection;

[0021] S4. After the FEVQVAE model training is completed, the discrete features obtained by the encoder and vector quantizer of the original data will be stored, and the original data will be discarded; when the original data is needed, the original data can be reconstructed through the stored discrete features, vector quantizer and decoder.

[0022] Beneficial effects of the present invention: The FEVQVAE model is proposed, and the problem of inaccurate frequency components when directly using the VQVAE method to reconstruct the structural vibration response is solved through the proposed dual-branch module and frequency module. The vibration response of the engineering structure can be compressed at a high compression ratio, and the reconstructed vibration response is very accurate, which is expected to be applied to the compression of large-scale structural vibration response in engineering practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a flow chart of the invented FEVQVAE method for compression of structural vibration response.

[0024] Figure 2 This is a schematic diagram of the training and inference process of the FEVQVAE model.

[0025] Figure 3 This is the encoder and decoder architecture diagram of the FEVQVAE model.

[0026] Figure 4 This is the vector quantizer architecture diagram of the FEVQVAE model.

[0027] Figure 5 This is the architecture diagram of the encoder module and decoder module of the FEVQVAE model.

[0028] Figure 6 This is the dual-branch module architecture diagram of the FEVQVAE model.

[0029] Figure 7 This is the frequency module architecture diagram of the FEVQVAE model. DETAILED DESCRIPTION

[0030] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0031] like Figure 1 As shown, the present invention provides a FEVQVAE method for structural vibration response compression, which includes four steps S1-S4:

[0032] The step S2.1 divides the data into three data sets in proportion, namely, a training set, a validation set, and a test set, which means that all vibration responses are divided into a training set, a validation set, and a test set in a ratio of 8:1:1;

[0033] The normalization in step S2.1.1 refers to linearly scaling each data segment to the range of [-1, 1].

[0034] In step S3, the FEVQVAE model includes three parts: encoder, vector quantizer and decoder. The configurations of encoder and decoder are as follows: Figure 3 As shown in Figure 1, the channel dimension c is equal to 32, and the D in the D-dimensional features in the encoder and decoder is equal to 256; sensor encoding refers to the position encoding in Transformer; the vector quantizer of the FEVQVAE model is as follows Figure 4 As shown (e1, e2, ..., e k ) represents a learnable dictionary.

[0035] In step S3, the encoder module in the FEVQVAE model encoder and the decoder module in the decoder are as follows: Figure 5 As shown in the figure, the encoder module consists of three dual-branch modules and a one-dimensional convolutional layer, the decoder consists of a one-dimensional transposed convolutional layer and three dual-branch modules, and SELU represents the non-linear activation function.

[0036] In step S3, the time domain features can reflect the evolution and dynamic changes of the signal over time, which helps to understand the overall behavior of the vibration signal; on the other hand, the vibration signal in the frequency domain can intuitively display the frequency components and periodic characteristics of the signal, which makes it easier to highlight the main frequency components of the signal, thereby making the features clearer. In addition, the frequency domain features are very intuitive for identifying the resonant frequency of the structure. Therefore, a dual-branch module is proposed for simultaneously extracting time domain features and frequency domain features, such as Figure 6 As shown, it contains two branches, namely the time domain branch and the frequency domain branch, and automatically fuses the extracted time domain and frequency domain features. Tanh represents the nonlinear activation function.

[0037] In step S3, in order to help the FEVQVAE model extract fine frequency domain features of the structural vibration response, such as resonant frequency and damping information, the present invention proposes a frequency module based on the characteristics that the structural vibration response has multiple resonant frequency bands in the frequency domain, such as Figure 7 As shown, it contains two branches, namely a single resonance frequency band branch and a multi-resonance frequency band branch, and Tanh represents a nonlinear activation function.

[0038] In step S3, the specific settings of hyperparameters are: initial learning rate is 0.005, number of training rounds is 200 rounds, number of batches is 16, momentum is 0.9, weight decay is 0.0001; the loss function is back-propagated, and the weights are updated using stochastic gradient descent (SGD). The error of the loss function is passed from the last layer to the front layer, and then the parameters of each layer are updated with the goal of minimizing the loss function. The number of iterations is set to 200 rounds. When the model reaches the maximum number of iterations and meets the process of reducing the loss function to convergence, the training is stopped; if the loss of the training set and the validation set no longer decreases, and the loss of the validation set is not higher than 120% of the loss of the training set, it is considered that the training process and results meet the applicable standards.

[0039] As shown in Table 1, the real monitoring data of a large suspension bridge is used to compare various vibration response compression methods. The results show that the reconstruction accuracy of the proposed FEVQVAE model is significantly improved under several different compression ratios. The present invention can be deployed in actual structural health monitoring systems and is expected to solve the storage and transmission challenges faced by monitoring big data.

[0040] Table 1 Root mean square error (m / s) of various vibration response compression methods at different compression ratios 2 )contrast

[0041]

Claims

1. A frequency-enhanced vector quantized variational autoencoder method for compressing structural vibration response, characterized in that: Here are the steps: S1. Collect the vibration response of the specified measuring point of the structure; S2, segmenting the vibration response of each channel of each sensor, merging all segmented data segments together, and then segmenting all data segments into data sets to form a training set, a validation set and a test set; The data segmentation of the vibration response of each channel of each sensor is as follows: S2.

1. Divide the collected vibration responses into three data sets in proportion, namely, training set, validation set and test set; S2.1.1, normalizing the vibration response in each data set generated in step S2.1; S3. Use the above three datasets to train the proposed FEVQVAE model, and finally save the best weights of FEVQVAE; S3.

1. Using the above three datasets to train and test FEVQVAE means randomly initializing a FEVQVAE, then using the training set and validation set constructed in step S2.1 to train and select the model, and finally saving the selected best model and weights; The FEVQVAE model is composed of an encoder, a vector quantizer and a decoder. By introducing the sensor encoding into the encoder and the decoder, the position information of the sensor is introduced into the FEVQVAE model, so that one FEVQVAE model is implemented to compress the vibration responses of all sensors. The encoder is composed of a one-dimensional convolutional layer, an encoder module and a sensor code, and the architecture is: input signal->one-dimensional convolutional layer+sensor code->encoder module+sensor code->encoder module+sensor code->encoder module+sensor code->encoder module+sensor code->one-dimensional convolutional layer+sensor code->D-dimensional feature; The vector quantizer is a two-dimensional matrix composed of multiple one-dimensional vectors; the decoder part is composed of one-dimensional convolution, decoder module and sensor encoding, and the architecture is: D-dimensional feature -> one-dimensional convolution layer + sensor encoding -> encoder module + sensor encoding -> encoder module + sensor encoding -> encoder module + sensor encoding -> encoder module + sensor encoding -> one-dimensional convolution layer + sensor encoding -> reconstructed input signal; The encoder module is composed of three dual-branch modules and a down-sampling one-dimensional convolution layer, and the architecture is: dual-branch module->dual-branch module->dual-branch module->down-sampling one-dimensional convolution layer; The decoder module consists of three dual-branch modules and an upsampling one-dimensional transposed convolutional layer, and the architecture is: upsampling one-dimensional convolutional layer -> dual-branch module -> dual-branch module -> dual-branch module; The dual-branch module performs feature fusion through a star convolution structure and adopts a residual connection to prevent the gradient in the model from disappearing; it is used to simultaneously extract time domain features and frequency domain features, and automatically fuse the extracted time domain and frequency domain features; the dual-branch module consists of a time domain branch and a frequency domain branch, and automatically fuses the time domain and frequency domain features through a star convolution structure, the time domain branch extracts the time domain features by a one-dimensional convolution layer with a convolution kernel of 3, and the frequency domain branch extracts the frequency domain features by a frequency module, then the features of the two branches are each passed through a one-dimensional convolution layer, and the star convolution structure is used to fuse the features of the two branches, and finally the fused features are passed through a one-dimensional convolution layer with a convolution kernel of 1, and a residual connection is adopted between the input and the output; The frequency module is used to extract rich frequency domain features of vibration response, and is composed of two branches: one has a small receptive field and is used to extract local frequency domain features, i.e., features of a single resonance frequency band; the other has a larger receptive field and is used to capture frequency domain features in a wider range, i.e., features of multiple resonance frequency bands; subsequently, a star convolution structure is used to enable the two branches to adaptively interact and fuse, and to retain useful information of the two branches; the frequency module is composed of a single resonance frequency band branch and a multi-resonance frequency band branch, and the features are first converted to the frequency domain using a fast Fourier transform, and then a one-dimensional convolution layer is used to extract the features of a single resonance frequency band, and a one-dimensional hole convolution layer is used to extract the features of multiple resonance frequency bands, and then the features of the two branches are each passed through a one-dimensional convolution, and the star convolution structure is used to fuse the features of the two branches, and the fused features are passed through a one-dimensional convolution layer with a convolution kernel of 1, and then the inverse fast Fourier transform is used to transform the features from the frequency domain to the time domain, and a residual connection is used between the features after the fast Fourier transform and the features before the inverse fast Fourier transform; S4. After the FEVQVAE model training is completed, the discrete features obtained by the encoder and vector quantizer of the original data will be stored, and the original data will be discarded; when the original data is needed, the original data can be reconstructed through the stored discrete features, vector quantizer and decoder.

Images