A novel intelligent indoor fire and combustibles detection method based on multi-channel transfer learning strategy with acoustic signals

Abstract

Fire is a ubiquitous hazard that poses a significant threat to life, property, and the environment in both industrial and everyday settings. Therefore, precise and timely detection of fire is crucial in preventing potential harm and minimizing the extent of damages. Currently, however, no fire alarms can detect combustible types. This paper proposed a novel deep-learning method to detect the combustibles of a burning fire utilizing acoustic signals. The method comprises a base model for fire detection and a group model for custom fire combustibles classification. Two models were trained at two stages; the proposed method trains the base model first, in which a single layer of Transformer encoder is utilized for fire detection. The utilization of the base model can aid the classification block in emphasizing the acoustic features of assured combustibles. After that, a novel classification block named Circuits-LSTM-Conformer (CLCFormer) is applied to construct a parallel model for combustible detection. Several experiments were executed to evaluate the efficacy of the proposed method. The outcomes indicate that the proposed method can precisely detect fire combustibles, with an accuracy of 92 %, surpassing all baseline approaches. The presented study presents a method for detecting combustibles that offers a blueprint for high precision, non-contact, convenient, and cost-effective detection of combustibles and provides reliable on-site guidance for firefighting.