Bridges are one of the most important structures in our infrastructural system today.The modern society depends on these structures which in majority are nearingthe end of their design life. Therefore, the need for monitoring of bridges is increasing.It is crucial to detect damages in bridges at early stages to prevent damagesfrom developing further while they are becoming more expensive to repair. Thefield in which systems and approaches for structural monitoring of civil engineeringstructures are used is called structural health monitoring.

In this thesis, a model-free unsupervised machine learning method for damagedetection is developed and presented. It utilizes autoregressive models (AR) asdamage-sensitive features and the mahalanobis squared distance (MSD) as a methodfor outlier detection. The method is tested with simulated data consisting of accelerationtime series and strains from a railway drawbridge called the tyska bron.For this, a FE-model of the railway bridge is used to simulate the bridges vibrationresponse under a passing train load. Randomly varying speed and axle loadsare used for the different train passages to create variability. Additionally, noise isadded to represent sensor noise and introduce imperfections in the signal. Damageis introduced by reducing the stiffness of elements in the FE-model. The objectiveis to find the lowest detectable damage from several selected damage locations andto investigate the effect of strain data on the damage detection capabilities of theproposed method.

The results show that damage is detectable with the developed method, with thelowest detectable damage being a 10% stiffness reduction along a 10-centimetrelongelement. The damage locations which are closest to the highest response ofone structural member have the best performance, since the added noise has lesseffect there. When adding strains to the acceleration data the performance of thealgorithm does not improve. In conclusion, the study showed that the AR-methodin combination with MSD is a good and efficient method for detecting damages forthis specific case and shows potential for usage in other cases.