Bridge Health Monitoring Using a Machine Learning Strategy

The goal of this project was to cast the Structural Health Monitoring (SHM) problem within a statistical pattern recognition framework. Techniques borrowed from speaker recognition, particularly speaker verification, were used as this discipline deals with problems very similar to those addressed by structural health monitoring. Speaker recognition is the task of verifying whether the speaker is the individual he/she claims to be by analyzing his/her speech signal. Based on the principle that speaker recognition can determine whether it is John or Jane who says the word “mom," it was assumed that it would be possible to find out whether it is the healthy or the damaged bridge that provides that acceleration time history. Inspired by the use of Bayesian Maximum A Posteriori (MAP) adaptation of the Universal Background Model in speaker recognition, this work proposed a new and computationally efficient method to update the probability distribution function describing the statistical properties of an ensemble of structural parameters through the adaptation of a Gaussian Mixture Model describing the pdf of observed indices parameterizing the structural behavior. Overall the results are very promising, but at its current stage the proposed work must be considered as a proof of concept for the application of the Bayesian MAP adaptation of a training model. The major drawback of the proposed work is its inability to produce estimates of marginal distribution of the structural parameters used as indices of the structure’s performance, since the joint distribution of said variables is obtained through an implicit function of the variables of interest.


  • English

Media Info

  • Media Type: Digital/other
  • Edition: Final Report
  • Features: Figures; References;
  • Pagination: 45p

Subject/Index Terms

Filing Info

  • Accession Number: 01666078
  • Record Type: Publication
  • Report/Paper Numbers: CAIT-UTC-NC3
  • Contract Numbers: DTRT13-G-UTC28
  • Created Date: Apr 16 2018 11:20AM