Speaker
Description
Feature extraction is an important part of data-based structural health moni-toring (SHM) under the theory of statistical pattern recognition. Among most of the signal processing techniques, time series analysis by time-invariant linear representations provides reliable and sensitive features to damage. Accracy and adequacy of a time series model to vibration time-domain responses depend strongly on the model order and uncorrelatedness of residual sequences. An insufficient order may lead to the extraction of insensitive features to damage and an erroneous damage detection. This pa-per improves a well-known conventional feature extraction method based on AutoRegressive (AR) modelling. The improved method is intended to improve the order of AR model at each sensor and extract the model coeffi-cients and residuals as the main damage-sensitive features. Mahalanobis distance technique is applied to investigate and com-pare the performance of the improved and conventional feature extraction approaches. Experi-mental datasets of a well-known benchmark problem in the context of SHM are utilized to implement the mentioned investigation and comparison. Rsults show that the improved feature extraction method not only satisfies statistical criteria but also leads to achieving more sensitive features to damage in comparison with the conventional technique.
