Abstract
This paper evaluates two powerful denoising techniques—Discrete Wavelet Transform (DWT) and Singular Value Decomposition (SVD)—to clean up noisy structural vibration data. While both methods excel at clarifying "bursty" seismic signals, SVD proves far more robust for the constant, low-level ambient vibrations that are essential for long-term structural health monitoring. This assessment is based on synthetic signals from a ten-story building model subjected to both earthquake and ambient excitations.
Introduction
Structural Health Monitoring (SHM) is shifting toward the use of low-cost sensors to monitor massive infrastructures like bridges and skyscrapers. However, "low-cost" often means "high-noise." When sensors produce signals buried in noise, traditional filters often fail because the noise and the actual structural vibration frequencies overlap. If we cannot separate the structural response from the sensor noise, accurate damage detection becomes impossible. This study compares DWT and SVD to determine which is better suited for the specific vibrations civil engineers encounter: sudden earthquakes and steady ambient loads.
Methodology
The researchers compared two primary algorithmic approaches:
- Discrete Wavelet Transform (DWT): This method breaks a signal down into different scales using "mother wavelets." By applying mathematical thresholds to these scales, it suppresses components likely to be noise.
- Singular Value Decomposition (SVD): This approach organizes the signal into a Hankel matrix. By performing a decomposition, the algorithm separates high-energy "singular values" (the signal) from low-energy ones (the noise floor).
Experimental Setup
The study relies on synthetic signals generated from a numerical model of a one-bay, ten-story shear-type building. The model was excited using two inputs: the 1986 Kalamata earthquake (non-stationary) and a linear force distribution simulating ambient vibrations (stationary). Zero-mean Gaussian white noise was added to the clean signals at varying Noise-to-Signal (N/S) ratios, ranging from 5% to 50%.
Results and Discussion
Performance was measured using the Percentage Root Mean Square Difference (PrmsD), where lower values indicate higher accuracy.
Seismic Response
Both methods performed excellently for seismic signals. At a 25% noise level, DWT achieved a PrmsD of 0.88%, while SVD was even more precise at 0.44%. Even at a massive 50% N/S ratio, SVD kept the error for seismic signals under 1%.
Ambient Vibration
The gap widened significantly for ambient vibrations. At a 25% noise level, DWT’s error jumped to 4.58%, whereas SVD maintained a clean 0.57%.
The nature of the vibration—its stationarity—is the deciding factor. Seismic signals are non-stationary with clear peaks, making them easier for Wavelets to isolate. Ambient vibrations are stationary and constant over time, making them harder to distinguish from white noise. SVD’s matrix-based approach manages to find the underlying structural patterns where DWT struggles.
Conclusion
SVD proves to be a remarkably stable all-rounder for structural denoising. Unlike DWT, which requires careful selection of wavelets and decomposition levels, SVD’s performance remains consistent across different noise levels and signal types. Future research should focus on testing these pipelines on real-world experimental data and expanding the scope to include displacement signals and non-Gaussian noise environments.
Originally presented at the 7th International Conference on Computational Methods in Structural Dynamics and Earthquake Engineering (COMPDYN 2019), 24-26 June 2019, Crete, Greece.
Summary updated