<p>Well-established output-only dynamic identification methods rely on strong assumptions, such as stationarity, time-invariance, and linearity of the investigated processes. These methods also require constant broadband spectrum inputs and long acquisition durations. However, in many practical situations, the unknown input excitations consist of many short, unexpected events, which vary suddenly in energy and frequency content. These variations may trigger a non-linear response in the structural system, thus violating one or more of the aforementioned assumptions. In the present work, the recently developed EMILIA algorithm is integrated into a broader Structural Health Monitoring (SHM) strategy aimed at early warning for structural condition assessment, enabling the automated detection of damage-induced instantaneous frequency shifts within a single acquisition record. The proposed methodology is tested using data from a steel shear frame specimen subject to controlled damage in laboratory environment in order to investigate its capability to perform output-only dynamic identification with reduced amount of data and time-varying structural responses. It is shown that EMILIA’s three main components – discrete wavelet time-domain data decomposition, Hilbert transform time-frequency analysis, and Bayesian conditional probabilities – enable the robust extraction of modal properties even in the presence of noisy data and fast time-varying events. The effectiveness of the proposed strategy is validated under both nonstationary conditions and with non-linear data using experimental response signals from the aforementioned shear frame. Unlike conventional SHM approaches, the time-frequency analysis enables a straightforward and accurate estimation of the time instant at which sudden variations in the structural dynamic behaviour occur, facilitating timely intervention planning in the event of permanent alterations in the targeted dynamic features.</p>