A busy monitoring system can produce gigabytes of vibration and strain data every day. None of it is useful until it is condensed into something an engineer can act on. SHM analytics is the discipline of that condensation — a pipeline that progressively reduces volume while increasing meaning.
Step 1 — clean the signal
Real data is noisy. Before anything else, signals are filtered, de-trended and checked for sensor faults and dropouts. A surprising share of false alarms in immature systems trace back not to the structure but to a loose cable or a saturated channel — so quality control is part of the science, not an afterthought.
Step 2 — extract damage-sensitive features
Next, the cleaned signal is reduced to <strong>features</strong>: compact quantities that are sensitive to damage but robust to everything else. Common choices include:
- modal features — natural frequencies, mode shapes, damping, derived through system identification;
- statistical features — RMS, kurtosis, and the coefficients of fitted time-series models;
- transmissibility and frequency-response functions between sensor pairs.
Good feature design is the heart of SHM. A feature that moves only when real damage appears — and ignores traffic, temperature and noise — is worth more than any algorithm applied downstream.
The confounding problem: environment
Here is the discipline's hardest practical challenge. Natural frequencies shift with temperature; a bridge can be several percent stiffer on a cold morning than a hot afternoon. These environmental and operational variations are often <em>larger</em> than the change a real defect would cause. Separating genuine damage from daily and seasonal swings — through data normalisation and models that learn the normal envelope — is where many programmes succeed or fail.
If a feature changes by 3% across the seasons, a damage indicator that triggers on a 1% shift will cry wolf. Robust SHM models the normal variation first, then looks for what is left over.
Step 3 — detect change with machine learning
With clean features in hand, damage detection becomes a pattern-recognition problem. Because labelled examples of a damaged structure are rare — you cannot break a real bridge to gather training data — most SHM leans on <strong>unsupervised, data-driven novelty detection</strong>: learn what healthy looks like from abundant baseline data, then flag statistically significant departures.
Supervised learning and physics-informed models add power where some knowledge of damage exists, helping move from mere detection toward localisation and severity. The frontier today is combining data-driven flexibility with physics-based interpretability, so a flag comes with an explanation an engineer can trust.
Step 4 — decide
The pipeline ends not in a number but in an action: keep monitoring, schedule an inspection, restrict load, or intervene. Translating a damage indicator into a defensible decision — with quantified uncertainty — is the ultimate purpose of every stage that came before.
Algorithms get the attention, but in SHM the leverage is in the features and in honestly modelling normal variation. Get those right and simple detectors work; get them wrong and no model can save you.