A Data-Based Framework for Identifying a Source Location of a Contaminant Spill in a River System with Random Measurement Errors.

This study addresses the problem of identifying the source location of a contaminant spill in a river system when a sensor network returns observations containing random measurement errors. To solve this problem, we suggest a new framework comprising three main steps: (i) spill detection, (ii) data preprocessing, and (iii) source identification. Specifically, we applied a statistical process control chart to detect a contaminant spill with measurement errors while keeping the false alarm rate at less than or equal to a user-specified value. After detecting a spill, we generated a nonlinear regression model to estimate a breakthrough curve of the observations and derive a characteristic vector of the estimated curve. Using the characteristic vector as an input, a random forest model was constructed with the sensor raising the first alarm. The model provides output values between 0 and 1 to represent the possibility of each candidate location being the true spill source. These possibility values allow users to identify strong candidate locations for the spill. The accuracy of our framework was tested on part of the Altamaha River system in Georgia, USA.