Reducing False Positives in Building Fire Detection Systems via Multiple Metal Oxide Sensors

Early and accurate fire detection in indoor environments is essential to ensuring public safety and minimising casualties and property loss. However, conventional fire alarm systems often generate false alarms due to non-fire aerosols such as water vapour and cigarette smoke. In this study, to address this issue, a sensor-based fire detection framework incorporating machine learning techniques has been proposed. A gas sensor matrix composed of multiple BME688 metal oxide sensors was used to collect air quality data representing five environmental conditions, including clean air, water vapour, e-cigarette aerosol, cigarette smoke, and real fire smoke. To enhance classification performance, the collected data have been pre-processed using low-pass filtering and wavelet transformation techniques. Machine learning approaches frequently used in the literature have been applied to classify sensor data. The results indicate that one-dimensional Convolutional Neural Network (1D-CNN) and Long Short-Term Memory (LSTM) approaches are superior to other models. These models have achieved perfect precision and recall scores (1.00) for the cigarette smoke and fire smoke classes, demonstrating their ability to reduce false alarms without compromising sensitivity to real fire events. These findings have confirmed the effectiveness of combining gas sensor arrays with deep learning in developing more reliable and context-aware fire detection systems.