A novelty detection method for large-scale structures under varying environmental conditions

Data-based approaches for damage detection are rather demanding processes for application in vibration-based structural health monitoring (SHM). Statistical pattern recognition by time series analysis for feature extraction, and novelty detection based on statistical decision-making can work efficiently in SHM applications; anyway, some limitations and shortcomings make a strategy based on them difficult and even unreliable, especially if adopted for large-scale structures. This is due to the fact that, in the process of feature extraction, the time series model most compatible with the vibration responses must be selected, and then appropriate orders determined to assure accuracy by extracting uncorrelated residuals without overfitting. For the statistical decision-making, the major issue is represented by the detection of damage under varying environmental conditions, which may induce false alarms. Aim of this work is to cope with the aforementioned issues when the goal is the early damage detection. We propose an automatic model identification algorithm and a novelty detection method that combines of a statistical distance measure, the partition-based Kullback-Leibler divergence, and the Mahalanobis-squared distance technique. Experimental datasets relevant to a cable-stayed bridge are exploited to validate the proposed approach and assess its accuracy. Results testify that the method turns out to be highly successful in detecting damage, even under environmental variability.