A variational inference based learning approach for decentralized building fire estimation

Abstract

Real-time estimation of the fire status is critical for reliable decision-making in building fire emergencies. Learning-based approaches for fire estimation have achieved great success, benefited from exploration of sensor data. However, existing works are mostly designed for specific building structures and seldom account for false alarms. Such flaws affect their generalization in broader environments and robustness in practical use. In this paper, we propose a decentralized learning-based approach for building fire estimation. The estimation is conducted in a decentralized scheme, where building zones conduct fire estimation respectively. The approach is built on variational inference technique, and can infer the location and dynamic intensity of the fires concurrently. Specifically, the fire estimation problem is interpreted in a latent variable model (LVM), with latent variables for fire intensity features and location features. The solution is then derived from variational inference (VI), and approximated via data learning on a novel deep network (namely FireEst-VAE). The network can produce accurate estimation of the fire intensity, while performing false alarm identification as a bonus effect. We evaluate the decentralized approach on full-scale experimental and simulation-generated data from a multi-compartment residence. The results illustrate that the proposed approach performs well on a wide range of fire scenarios with one or multiple fire locations, while preventing false alarms from common nuisance sources. The decentralized learning-based approach for fire estimation shows great potential in smart firefighting for large-scale buildings. • A decentralized learning approach based on variational inference for fire estimation. • Integrating interpretability from model methods and efficiency from learning methods. • The decentralized scheme benefits edge computing and generalization in practical use. • Promising performances on fire intensity estimation and false alarm identification.