An IoT-Based LSTM System for Early Fire Detection in Wood Waste Warehouses

Abstract

High global fire frequencies and the inherent vulnerability of industrial wood waste storage demand advanced monitoring solutions to prevent catastrophic loss. This research develops an Internet of Things-based early fire detection system specifically designed for wood waste warehouses. The system is grounded in recurrent neural network theory, utilising multivariate Long Short-Term Memory architectures to model long-term temporal dependencies between thermal and gas variables. Following experimental research design and the iterative Spiral Model, the study utilises a technical sample of 5,280 sensor records and a user sample of 30 warehouse personnel. Data gathered in a simulated warehouse setting were processed using a 15-step sliding window and Synthetic Minority Over-sampling Technique to generate a balanced training set of 52,800 sequences. The Long Short-Term Memory model achieved a weighted average accuracy of 98.6% and an F1-score of 98.1%. The system delivered alerts with a best-case end-to-end latency of 3.23 seconds, and statistical tests confirmed no significant difference in usability scores between different personnel roles. The detection logic misclassified 0 Danger events as Normal states, proving the reliability of the multivariate temporal approach. This study provides a novel application of sliding-window multivariate recurrent networks to the specific slow-evolving thermal signatures of dense organic waste; consequently, warehouse operators should integrate temporal predictive modelling into fire safety frameworks to reduce response latencies below the levels of traditional threshold-based alarms.