Review of methods used for outlier detection in structural health monitoring

Identification of outliers is a vital step in a large number of structural health monitoring systems. If the system is designed to detect the occurrence of damage, then outlier detection will most likely comprise a part of its methodology. The general procedure for damage detection can be described in the following way: firstly, establish a healthy baseline for the structure, based on measured data for example, and then any future monitoring data can be compared to this baseline to check if it shows normal behaviour. This process of determining whether or not the data falls within the parameters of the baseline can be categorised as outlier detection. Outlier detection is not only used at the final stage of damage detection, but also in the training of the baseline, as outliers at this stage of the process could mask the existence of damage in future data. In this paper, the effectiveness of various outlier detection methods are reviewed. Reviewed methods include the most commonly used in structural health monitoring (e.g. Minimum Covariance Determinant), as well as some that are more commonly found in econometrics. A selection of these methods are applied to real world frequency data obtained from a short span bridge over a period of 19 days. This study informs the design of structural health monitoring systems and aids in making a decision on the most appropriate outlier detection method to use for particular applications and circumstances.