Frequency enhanced vector quantized variational autoencoder for structural vibration response compression
•The frequency enhanced vector quantized variational autoencoder (FEVQVAE) is proposed for structural vibration response compression.•The proposed dual-branch block in both time and frequency domains enhances feature extraction, to some extent mitigating the challenge for neural networks to extract...
Saved in:
| Published in: | Mechanical systems and signal processing Vol. 224; p. 112136 |
|---|---|
| Main Authors: | , , |
| Format: | Journal Article |
| Language: | English |
| Published: |
Elsevier Ltd
01.02.2025
|
| Subjects: | |
| ISSN: | 0888-3270 |
| Online Access: | Get full text |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | •The frequency enhanced vector quantized variational autoencoder (FEVQVAE) is proposed for structural vibration response compression.•The proposed dual-branch block in both time and frequency domains enhances feature extraction, to some extent mitigating the challenge for neural networks to extract high-frequency features.•The proposed frequency block separates the extraction of features for a single resonant band and multiple resonant bands, enabling the FEVQVAE to more effectively extract frequency-domain features of vibration responses.
The structural health monitoring system has been widely installed on large civil structures, generating a significant amount of structural vibration response data over their long-term service life, which poses challenges for data transmission and storage. The compression method for structural vibration responses based on the traditional deep autoencoder (AE) can only compress the original data into low-dimensional floating-point features and not into low-dimensional integer features, which limits its compression capability. To address this issue, this paper proposes a frequency enhanced vector quantized variational autoencoder (FEVQVAE) method for compressing structural vibration responses with higher compression ratios. The proposed method has three key innovations. Firstly, the proposed time-domain and frequency-domain dual-branch block enhances the feature extraction capability of both the encoder and decoder, to some extent mitigating the challenge for neural networks to extract high-frequency features. Secondly, the proposed frequency block separates the extraction of features for a single resonant band and multiple resonant bands, enabling the encoder and decoder to more effectively extract frequency-domain features of the vibration responses. Thirdly, by introducing sensor position encoding, compression of multiple sensor data can be achieved with only one model. The effectiveness of the proposed method is validated using acceleration responses from a Dowling Hall Footbridge under normal operating conditions and a long-span suspension bridge subjected to non-stationary excitations such as wind and vehicle loads. Experimental results demonstrate that the compression performance of the proposed FEVQVAE method is significantly improved compared to the AE method and the vector quantized variational autoencoder method. Modal parameter identification results of the original and reconstructed responses show excellent consistency at a compression ratio of 19.2, with a maximum relative frequency error of only 0.952% for the first six mode frequencies and a minimum modal confidence criterion of 0.9614 for the first six mode shapes. Overall, the proposed method exhibits high precision in the compression of structural vibration response, effectively alleviating the storage and transmission challenges of monitoring big data. |
|---|---|
| ISSN: | 0888-3270 |
| DOI: | 10.1016/j.ymssp.2024.112136 |