Long Term Sensor Malfunction Detection and Data Regeneration using Autoregressive Time Series Models

Abstract

Structural Health Monitoring (SHM) seeks to assess the integrity of a structure using installed sensors and a variety of structural and data analyses. SHM stands poised as a partial solution to maintain the failing infrastructure in America, but has not yet seen widespread implementation. One factor restraining SHM from universal implementation and prominence is a lack of reliability in sensing and monitoring as a whole. In many cases sensor readings can drift further and further away from the true measurement sought, or even fail altogether. This paper presents one method for the detection of drifting or failing sensors, based on a series of autoregressive (AR) time series models. These numerical models are trained on healthy sensor data, usually from the infancy of the structure or sensor use, and then the distributions of the residuals, or model error, is compared for healthy and unknown time periods on the structure. In cases of failed sensors, a multivariate AR model relating nearby, healthy sensors, is trained on early data before failure and then used to regenerate later time periods of sensor failure. This data regeneration will never perfectly replace the failed sensor, but still provides with replacement values that are sufficiently accurate so they can be meaningfully used in data analysis. These methods are validated on data collected from the Streicker Bridge, on campus at Princeton University.