Model-free damage detection of a laboratory bridge using artificial neural networks

• Published in: Journal of civil structural health monitoring 2020-04, Vol.10 (2), p.183-195
• Main Authors: Ruffels, Aaron, Gonzalez, Ignacio, Karoumi, Raid
• Format: Article
• Language: English
• Subjects: Accelerometers; Algorithms; Arch bridges; Artificial neural network; Artificial neural networks; Bridge accelerations; Civil Engineering; Computational complexity; Control; Damage detection; Damage-sensitive features; Data analysis; Dynamical Systems; Engineering; Laboratories; Laboratory models; Learning algorithms; Learning systems; Machine learning; Mahalanobis distances; Mean square error; Measurement Science and Instrumentation; Model free; Neural networks; Original Paper; Outliers (statistics); Reference state; Root mean squared errors; Statistics; Steel arch bridges; Steel bridges; Steel structures; Structural damage; Structural health monitoring; Vibration
• ISSN: 2190-5452; 2190-5479; 2190-5479
• DOI: 10.1007/s13349-019-00375-2

This paper investigates a model-free damage detection method using a laboratory model of a steel arch bridge with a five-metre span. The efficiency of the algorithm was studied for various damage cases. The structure was excited with a rolling mass and seven accelerometers were used to record its response. An artificial neural network (ANN) was trained to predict the bridge accelerations based on data collected from the undamaged structure. Damage-sensitive features were defined as the root mean squared errors between the measured data and the ANN predictions. A baseline healthy state was established with which new data could be compared to. Outliers from the reference state were taken as an indication of damage. Two outlier detection methods were used: Mahalanobis distance and the Kolmogorov–Smirnov test. The method showed promising results and damage was successfully detected for four out of the five single damage cases. The gradual damage case was also detected, however, for some instances, greater damage did not result in an increase in the damage index. The Kolmogorov–Smirnov test performed best at detecting small single damage cases, while Mahalanobis distance was better at tracking gradual damage.