Research on an Early Detection Method for Residential Indoor Fires Based on the LSTM-FNN Algorithm

Abstract

To address the issues of delayed response and high false alarm/miss rates in traditional residential indoor fire detection methods, an LSTM-FNN model integrating Long Short-Term Memory (LSTM) and Fuzzy Neural Network (FNN) was developed to enable early identification of smoldering and flaming fire stages. The model utilizes the LSTM module to extract temporal dynamic features of smoke, carbon monoxide concentration, and temperature. It combines the FNN module to handle feature uncertainty and incorporates an attention mechanism for feature selection. Standard fire scenarios were constructed using Pyrosim software to generate training data. A physical experimental platform was established to comprehensively validate the model's early warning capability and generalization performance. Results demonstrate that the LSTM-FNN model achieves remarkably low Mean Squared Error (MSE) values of 0.0009, 0.0002, and 0.0002 during no-fire, smoldering, and flaming phases, respectively. Its Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) metrics significantly outperform standalone models. Validation experiments on typical combustibles—beech wood, cotton rope, polyurethane, and n-heptane—confirm its effective capture of early-stage fire characteristics. This method offers superior early warning timeliness compared to traditional approaches, significantly reducing false alarm and missed detection rates while enhancing the timeliness and reliability of early fire detection.