Random forest classifier for real-time chemical leak source tracking using fence-monitoring sensors

Abstract

Fast and reliable diagnosis of chemical leak and leak location(s) can save lives and reduce the damage from chemical accidents by enabling quick response. This paper presents a method that uses random forest (RF) classifier to track the location of chemical leak in real-time. A set of big data of leak accidents, which is needed to learn the RF classifier, is extracted by performing massive CFD simulations on a real chemical plant. The RF model is designed with optimal parameters and descriptors through parameter effect experiment. Feature ranking is also implemented to eliminate unnecessary attributes from the dataset. Using the pre-processed data, the optimal RF model achieved a test accuracy of 86.9% for the classification problem of predicting the leak location among 40-potential leak sources in the plant. Furthermore, when analyzing prediction errors by visualizing the classification boundary of RF model, most of the prediction errors are confirmed to be misclassification of adjacent leak locations. Considering the high prediction accuracy of the RF model, the RF-based leak source tracking model is expected to be effectively applied to industrial leak accidents.