Development of a machine-learning approach for identifying the stages of fire development in residential room fires. (December 2021)

• Record Type: Journal Article
• Authors: Fang, Hongqiang; Lo, S.M.; Zhang, Yunjie; Shen, Yixin
• Is Part Of: Fire safety journal. Volume 126(2021)
• Publication Date: 2021-12
• Subjects: Compartment fires -- Stages of fire development -- Machine learning -- GMM-HMM -- Fire scenarios; Fire prevention -- Periodicals; Incendies -- Prévention -- Recherche -- Périodiques; Fire prevention -- Research; Periodicals; 628.92205
• DOI: 10.1016/j.firesaf.2021.103469
• Languages: English
• ISSNs: 0379-7112

Abstract

Abstract: Recognizing the stages of fire development is essential for fire emergency operations. It allows firefighters to predict what will happen next, potential fire spreads, and the likely effect of tactical actions. Currently, firefighters recognize fire stages mainly by observing and judging the signs and symptoms of fire development changing on-site. However, this kind of approach highly relies on firefighters' knowledge and experience, making it difficult to operate. Therefore, a machine learning (ML)-based approach automatically identifying the stages of fire development in residential room fires is proposed in this paper. Modeled by Gaussian Mixture Models and Hidden Markov Models (GMM-HMM), the approach enables identifying the stages of fire development from short-term field temperature collections. To provide adequate data for model training, the two-zone fire model— CFAST and a non-parametric fire design method are applied to generate the temperature observations in various random fire scenarios. Taking the fire in a typical single-story residential construction as a case study, we establish a GMM-HMM-based recognition model with the simulated temperature data. It presents an average of 85% accuracy in identifying the fire stages within the 2 min error range. Moreover, tested with the experimental fire data, the established model also achieves successful recognitions. Highlights: A data-driven approach was proposed for identifying the stages of fire development in residential room fires. Machine learning methods were employed to formulate the recognition model. The data for model training was provided by simulating a large number of fire scenarios.