Hidden Markov models for sequential damage detection of bridges

Damage detection is an important first step in the implementation of structural health monitoring (SHM) systems. Recent advances in sensing and wireless communications have enabled civil engineers to affordably collect vast datasets from operational structures. Given the data available, data-driven damage detection frameworks are receiving significant attention and are being established as a promising tool for SHM. Unsupervised data driven SHM frameworks treat observations sampled throughout the process of data collection as independent and identically distributed (i.i.d.). This assumption is overconstraining as these observations are sequentially dependent. For example, given three observations sampled consecutively and the i.i.d. assumption, the first and last could be labeled as undamaged while the middle one as damaged. Since structural damage does not often heal on itself by definition, this prediction is in error. The focus of this work is to relax the independence assumption by taking into account the sequential nature of observations. Hidden Markov models (HMM) are used to model sequences of observed features of structural response and the probability of observing a sequences of fixed length is used to make decision on the state of the structure. The experimental dataset from the Z-24 bridge are employed to validate the functionality of the proposed HMM method. Preliminary results indicate that using the sequential nature of observations results in superior performance in damage detection compared to a baseline case when the i.i.d. assumption is made.