A Virtual Baseline Method to Enhance Anomaly Detection in Progressive Damage Scenarios

This study presents a novel method to enhance anomaly detection in Structural Health Monitoring systems. The method is suitable for systems characterized by non-stationary data due to environmental and operational variability as well as progressive damage. The approach in fact leverages a baseline dataset with no special requirements as it can contain both damage and environmental sensitive structural features, like modal frequencies. Initially, a regression model is trained to predict modal frequencies with environmental variables as inputs. Unlike conventional methods, the proposed approach incorporates time as another input parameter, given its connection to the progression of damage. To create the virtual baseline, the model keeps the environmental variations while fixing the time variable constant at its initial value, which ideally corresponds to the system least damaged state, to predict the modal frequencies. The resulting virtual baseline simulates the structure under varying environmental conditions but with no damage progression. In this way, the method separates the influence of environmental variations from the effects of damage, offering a more accurate reference for anomaly detection. The proposed technique is validated through experimental dataset of KW51 bridge, demonstrating its ability to effectively mitigate the effects of non-stationary data, leading to improved damage detection performance.