Multimodal sensor data fusion and event recognition in IoT

Abstract

In the context of smart homes that are resource-constrained and are managed by an IoT network, a fire event detection system is a very important safety problem because of the critical high false alarm rate, lack of contextual awareness, and lack of modelling capability of the temporal evolution of the fire conditions of the single-sensor-based systems. This paper suggests a novel hybrid multimodal approach called RF–LSTM–ANN which combines the Random Forest (RF) model for feature selection, the Long Short-Term Memory (LSTM) model for temporal dependency modeling and the Artificial Neural Network (ANN) model for nonlinear binary classification. Heterogeneous sensor data including temperature, gas, smoke, and environmental data is collected from three publicly available real-world sources, resulting in a multimodal sensor dataset of 47,234 samples with 18 features. A feature-level fusion strategy with weighted importance score derived from RF is used and then the data is divided into 70% training, 15% validation and 15% test sets. The proposed framework outperforms all baseline models by at least 2.4% in terms of test accuracy with 96.8%, precision of 95.9%, recall of 96.3%, F1-score of 96.1% and ROC–AUC of 0.982. Results from the ablation studies validate the individual contributions and significance of each component and the class-conditional feature importance analysis demonstrates that temperature, gas concentration, and smoke level are the most important fire predictors. The proposed framework is a scalable, reproducible and interpretable solution to real time fire event detection and can be easily applied to intelligent safety systems in smart IoT environments.