Virtual baseline to improve anomaly detection of SHM systems with non-stationary data

An approach is proposed to improve anomaly detection of structural health monitoring systems by constructing “virtual baselines” for structures undergoing non-stationarities due to environmental and operational variability (EOV) and growing damage. The process requires a baseline dataset of structural damage-sensitive (SDS) parameters as well as environmental and operational (EO) variables. On this data, at first a regression model is trained with SDS parameters as the target dependent variables, and EO parameters as independent features. In contrast to classical models that rely solely on EO independent features, the proposed method incorporates the time information of the samples. This addition allows time to represent the progression of damage in the regression model, as time and damage growth are closely related.

The regression model is utilized to construct a virtual baseline by incorporating the corresponding EO parameters while fixing the time information to that of the initial sample. This approach preserves EO variations while setting the damage information to a constant value, specifically that of the first sample, which is assumed to represent minimum damage. The virtual baseline is then employed in the anomaly detection and EOV compensation process. Through examples on numerical and experimental datasets, with and without EOV compensation, the effectiveness of the proposed method is demonstrated, highlighting its capability to mitigate both damage-related and EOV-related non-stationarities from the baseline and improve the probability of damage detection.