Damage or a warm day?

Tap a steel building and hear its vibration modes as a chord. Warm weather, people and damage all shift it; an autoencoder learns which shifts are normal.

Tap the building and listen. Each note in the chord is one way the building sways or twists.

Your browser can’t show the 3D view; the sound, the spectrum and the monitoring still work.

Your browser can’t play sound here; the spectrum shows the chord.

Every mode is where the healthy reference has it.

Mode shapes enlarged (the roof moves H/30), illustrative. Click a floor to tap there.

Filled: now; outlined: healthy. Each bar sounds at

K\!\cdot\! f

with

K

= 162.1 (the first healthy mode at 110 Hz); labels give the building’s own frequency.

Storeys

Lateral system

Modes, tracked against the healthy reference
Mode	f, Hz	Type	Change
1	0.68	X	+0.0 %
2	1.15	Y	+0.0 %
3	1.34	Twist	+0.0 %
4	1.79	X	+0.0 %
5	2.88	X	+0.0 %
6	3.03	Y	+0.0 %

Method

The building is a torsionally coupled shear model: rigid floors with three degrees of freedom each, columns ( $12EI/h^{3}$ , fixed between rigid floors) and X-braces as storey springs. Its modes come from the generalized eigenproblem.

Each mode becomes a tone: its frequency is multiplied by a fixed factor so the healthy building’s first mode sounds at 110 Hz, and the ratios between the modes, the chord, are the building’s own (parameter-mapping sonification). The loudness is how much a tap at a floor corner moves the mode times how much the roof corner moves in it; the decay is that of 2 % damping.

An autoencoder (6–8–2–8–6) learns the first six frequencies over 1,000 simulated healthy days with changing temperature and occupancy; a day it cannot rebuild is flagged. Autoassociative networks have been used this way since Sohn, Worden and Farrar (2002); the Mahalanobis distance is the classic baseline (Worden, Manson and Fieller 2000). Both thresholds are the 99th percentile on 1,000 separate healthy days.

The autoencoder idea comes from my team project for the EWSHM 2026 challenge; this tool uses synthetic data only.

References Code

Assumptions and limits

A simplified model with assumed loads ( $g_{\mathrm{k}}=4.0\ \mathrm{kN/m^2}$ , the roof without imposed load) and damping; the temperature effect (0.15 % stiffness per °C) is illustrative. Real buildings vary mostly through non-structural parts (Peeters and De Roeck 2001, the Z24 bridge).

Frequencies are computed, not identified from measured vibration; the 0.2 % noise stands in for identification error. The sound is a mapping, not a recording. A symmetric building already beats a little, because some of its modes are close.

The braces sit on the two center lines, so they add no twisting stiffness: the braced building twists first.

Detection rates and false alarms describe this simulation, not a real building. A threshold taken from 1,000 days is itself uncertain: over ten training seeds, the false-alarm rate on 20,000 new healthy days ranged from 0.6 to 1.7 % (mean 1.0 %). Damage that lowers every frequency alike (an edge column, for instance) can look like a warm, busy day and go unflagged.

Cite

@software{guven_building_chord,
  author = {Güven, Murat},
  title  = {Damage or a warm day?},
  year   = {2026},
  url    = {https://likemaestro.github.io/tools/building-chord}
}

References

CEN (2002). EN 1990: Eurocode – Basis of structural design. European Committee for Standardization.
Glorot, X., & Bengio, Y. (2010). Understanding the difficulty of training deep feedforward neural networks. Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, Proceedings of Machine Learning Research, 9, 249–256. proceedings.mlr.press/v9/glorot10a.html
Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic optimization. arXiv. doi:10.48550/arXiv.1412.6980
Peeters, B., & De Roeck, G. (2001). One-year monitoring of the Z24-Bridge: Environmental effects versus damage events. Earthquake Engineering & Structural Dynamics, 30(2), 149–171. doi:10.1002/1096-9845(200102)30:2<149::AID-EQE1>3.0.CO;2-Z
Sohn, H., Worden, K., & Farrar, C. R. (2002). Statistical damage classification under changing environmental and operational conditions. Journal of Intelligent Material Systems and Structures, 13(9), 561–574. doi:10.1106/104538902030904
Worden, K., Manson, G., & Fieller, N. R. J. (2000). Damage detection using outlier analysis. Journal of Sound and Vibration, 229(3), 647–667. doi:10.1006/jsvi.1999.2514