Natural Frequency Identification in Noisy Environments: A Topology‐Enhanced Approach Using Deep Learning and Clustering Algorithms
Operational Modal Analysis (OMA) methods are commonly used to estimate the modal characteristics of structures, but their accuracy decreases in power plants and similar facilities where operating conditions vary continuously and noise often obscures the true structural response. In dams, which are l...
Saved in:
| Published in: | Structural control and health monitoring Vol. 2025; no. 1 |
|---|---|
| Main Authors: | , |
| Format: | Journal Article |
| Language: | English |
| Published: |
Pavia
John Wiley & Sons, Inc
01.01.2025
|
| Subjects: | |
| ISSN: | 1545-2255, 1545-2263 |
| Online Access: | Get full text |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Operational Modal Analysis (OMA) methods are commonly used to estimate the modal characteristics of structures, but their accuracy decreases in power plants and similar facilities where operating conditions vary continuously and noise often obscures the true structural response. In dams, which are large mass and highly rigid structures, the recorded response vibrations have very low amplitudes and are often contaminated by external influences (e.g., turbine operation), limiting the effectiveness of classical peak picking OMA techniques. Additionally, time domain identification methods such as Stochastic Subspace Identification (SSI) may also struggle under these conditions, as noise can obscure the impulse‐like features or modal transients required for accurate estimation. These challenges are even more pronounced in Roller Compacted Concrete (RCC) dams. While thinner arch dams may exhibit more distinct dynamic responses under ambient excitations, the massive bodies of RCC dams generate extremely low vibration amplitudes, making the reliable identification of modal parameters considerably more difficult. The integration of intake structures into the dam body causes continuous turbine‐induced vibrations from hydroelectric power generation. This persistent excitation further complicates the separation of true structural modes from machine‐induced noise. Consequently, the direct applicability of conventional OMA techniques to RCC dams is limited, and alternative approaches specifically tailored to this dam type are required. Within this framework, the present study uniquely exploits the sinusoidal excitation induced by turbine operation during electricity generation as a sustained and predictable source of ambient vibration, thereby providing new insights into the dynamic characterization of RCC dams. In the context of this research, acceleration data in the time domain, obtained from sensors installed on both the Gürsöğüt‐2 dam body and adjacent bedrock, were analyzed. The bedrock data were treated as the noise source, and complex, nonlinear effects on the dam body were filtered through a Long Short‐Term Memory (LSTM)–based deep learning model. Filtered data from different dates were analyzed in the frequency domain, and mode shapes exhibiting distinctive characteristics were selected and clustered based on their similarities using the Self‐Organizing Map (SOM) method. For the comparison of mode shapes, persistent latent representations were obtained by leveraging the topological properties of their vectors and analyzed in a low‐dimensional space. This approach facilitated the rapid and effective identification of fundamental patterns and distinctive structural features among various modal responses. From the SOM clusters, characteristic frequencies such as Maximum Energy Frequency (MEF), Minimum Damping Frequency (MDF), and Most Frequent Frequency (MEF) were extracted. These were used to evaluate their interrelationships, filter out spectral features potentially associated with structural resonance, and ultimately develop an Automated OMA method. The validity of the results was subsequently tested. The study’s results revealed that the 3–7 Hz frequency band is critical for the resonance behavior of the structure and that the mode shapes within this range represent consistent and structurally characteristic vibration modes. In contrast, mode shapes exhibiting inconsistent and irregular behavior were found to be largely noise‐induced and were filtered out by the system. Furthermore, the comparative evaluation of MEF, MDF, and MFF enabled the classification of modes into reliable natural modes, rarely observed structural modes, and nonstructural vibrations induced by persistent environmental sources. This classification, together with the damping values calculated in the 5.7%–7.6% range, provides critical insights for resonance risk assessment and establishes a robust foundation for long‐term structural health monitoring. |
|---|---|
| Bibliography: | ObjectType-Article-1 SourceType-Scholarly Journals-1 ObjectType-Feature-2 content type line 14 |
| ISSN: | 1545-2255 1545-2263 |
| DOI: | 10.1155/stc/1007014 |