Machine learning and sensor swarm for structural health monitoring of a bridge

The present work proposes an unsupervised early-stage damage detection method, which relies on the combined application of the Principal Component Analysis, for features extraction and dimensionality reduction, and Symbolic Data Analysis, for automatically cluster different patterns. The structure considered is a Warren truss bridge, which is numerically simulated by a Finite Element Model. It is excited by a thermal cycle and a static load; the damage is modelled as a sudden reduction of the area of the section. The validity of the proposed algorithm is numerically tested over one month of vibration data. The damage is properly identified by some PCAs; furthermore, Symbolic Data Analysis allows an effective clustering of damaged and undamaged PCA samples. Robustness of the algorithm is tested at different noise level, timing of damage, damage position and depth, the influence of the sensors' number is also tested.