Self-supervised online learning algorithm for rTPNN to reduce false alarms in fire detection

Commercial fire detectors often suffer from false positive alarms due to benign environmental variations, such as cooking odors, humidity, and smoke. In order to address this issue, we develop a self-adapting fire detection system with Sample & Sensor Selective Self Online Learning Algorithm (4-SOLA). This system is comprised of Recurrent Trend Predictive Neural Network (rTPNN) enhanced with 4-SOLA for a multi-sensor fire detector. The rTPNN model processes sensor data and makes decisions providing a real-time detection while 4-SOLA combines and orchestrates supervised transfer learning based on human feedback and self-learning using unlabeled and precisely selected sensor data. The proposed system is evaluated using two public datasets, namely the NIST indoor fire dataset and the EN 54 test room dataset. Results show that the rTPNN enhanced with 4-SOLA is the best performing model among nine different models, achieving a balanced performance of 0.75 TPR and 0.90 TNR on the NIST dataset, and 0.97 TPR and 0.94 TNR on the EN 54 dataset when human feedback is fully available. The rTPNN enhanced with 4-SOLA system demonstrates great potential for applications beyond fire detection, particularly for time-series systems that require adaptation to constantly updating data.