Feasibility Analysis of Coupling FDS Modeling with Machine Learning for Situational Awareness in Aircraft Hangars

Thumbnail Image
Publication or External Link
Davis, Alison Marie
Milke, James A
Situational awareness is a critical factor in maintaining the safety of firefighters and can be largely improved in buildings using distributed sensors that provide real-time data. A two-phase approach is used to increase situational awareness in aircraft hangars. Phase I consists of modeling a hangar with an open door in Fire Dynamics Simulator (FDS), with a high density of smoke, temperature, CO and CO2 sensors located at the ceiling. Fuels of interest including Douglas fir, polyethylene, paper, JP-8, and propane are modeled in six potential fire locations, with five locations along the centerline of the hangar and one in the corner of the hangar. Additionally, wind and beams at the ceiling are added to the simulation to determine the impact on the products of combustion that the sensors pick up. Phase II uses the data acquired from the FDS simulations to inform and build machine learning models that utilize supervised learning techniques to identify the location of the fire, the magnitude of the fire and the composition of the fuel that is burning. It is determined that temperature and smoke are the key products of combustion needed for these analyses. The location of the fire is identified within a circular area with a 5 m radius by using temperature measurements, thus reducing the amount of input data needed for the machine learning models. The magnitude of the fire is predicted using temperature as inputs to a heat release rate (HRR) model using a fully connected, three-layer, feed forward neural network. The composition of the fuel is predicted using a linear support vector machine that supports multi-class classification, using products of temperature and smoke obscuration as inputs. The location model is 80% accurate, the HRR model is 85% accurate and the fuel composition model varies between 62% and 91% accuracy depending on the classification goals. These results prove the feasibility of machine learning applications in an aircraft hangar setting.