Ensemble-Learning-Based Multi-sensor Fire Detection Technology with CO2 Background Noise Adaptation

CO₂ is a significant fire detection parameter. However, its effectiveness can be compromised by high background concentrations, leading to false alarms or inaccurate detection. To mitigate the effects of uncertain ambient background concentrations on the performance of CO₂ fire detectors, this study developed an environmentally adaptive multi-sensor fire detection model based on an ensemble learning approach. First, an ensemble learning model consisting of base classifiers of gated recurrent unit (GRU) and convolutional neural network (CNN) was designed to learn and classify CO₂ and smoke concentrations under fire and non-fire conditions. Second, the fire detection performance of the multi-sensor ensemble learning model was validated using accuracy, precision, recall, and F1 score metrics. Finally, its performance was compared with those of multi-sensor fire detection models constructed with a CNN and GRU network. The results indicate that the multi-sensor fire detection model based on ensemble learning outperformed the fire detection models based on CNN and GRU. Moreover, the proposed model became more self-adaptive to the environment of the installation space as the number of CO₂ background concentration signals used for training increased. The fire recognition accuracy of the multi-sensor fire detection model exceeded 99% with 150 sets of CO₂ background concentration signals supplemented to the training dataset.