A Bayesian extended sequential sensor optimization scheme for damage detection in ballasted tracks

The problems of sensor placement and evaluation are well-known in structural health monitoring. In this study, an extended sequential sensor optimization scheme for sensor placement, which is more robust and effective than conventional methods, was developed. Using the information entropy as the measure of optimality in candidate configurations, the proposed scheme implements Bayes theorem and MCMC samples to evaluate optimality. Moreover, novel optimality indices were developed to compare the efficiency of sensor configurations having the same number of sensors or otherwise. The proposed optimization scheme efficiently addresses the large computational demand involved in some existing techniques. Although relatively cheap in its demand, the proposed scheme's efficiency is improved by ensuring that alternative configurations to the optimal one are available for a given number of sensors. Most sensor optimization scheme involving information entropy suffers from sensors been placed too close to each other, leading to information redundancy. This problem of information redundancy is addressed by enforcing a minimum distance between sensors, as well as the spatial correlation of prediction error at the observed degrees-of-freedom. The proposed scheme is implemented in the damage detection of a ballasted track system. The optimal configuration obtained with the proposed scheme at a various number of sensors is compared with popular conventional ones, and the superiority of the proposed method in terms of system information is established. Furthermore, for verification purposes, the optimal configurations obtained with each of the optimization schemes been compared were used to predict the location and severity of ballast damage in a ballasted track panel, using experimentally measured acceleration data. The results demonstrate the better reliability and practicality of the proposed scheme for sensor placement, especially in field applications, compared to conventional methods.