Resumen de Fire detectors based on chemical sensor arrays and machine learning algorithms: calibration and test
In some types of fire, namely, smoldering fires or involving polymers without flame, gases and volatiles appear before smoke is released. Most of the fatalities registered for fires, are caused due to the intoxication of the building occupants over the burns. Nowadays, conventional fire detectors are based on the detection of smoke or airborne particles. In smoldering fires situations, conventional fire detectors triggers the alarm after the release of toxic emissions. The early emission of gas in fires opens the possibility to build fire alarm systems with shorter response times than widespread smoke-based detectors. Actually, the sensitivity of gas sensors to combustion products has been proved for many years. However, already early works remarked the challenge of providing reliable fire detection using chemical sensors. As gas sensors are not specific, they can be calibrated to detect large variety of fire signatures. But, at the same time, they are also potentially sensitive to any activity that releases volatiles when being performed. Cross-sensitivity to water vapor and other chemical compounds make gas-based fire alarm systems prone to false positives. For that reason, the development of reliable and robust fire detectors based on gas sensors relies in pattern recognition and Machine Learning algorithms to discriminate fire from nuisance sensor signatures.
The presented PhD. Thesis explore the role of pattern recognition algorithms for fire detection using detectors based exclusively in chemical sensors. Two prototypes based on different types of gas sensors were designed. The sensor selection was performed to be sensitive to combustion products and to capture other volatiles that may help to discriminate fire and nuisances. Machine Learning algorithms for the prediction of fire were trained using standard fire tests stablished in EU norm 54. Additionally to those test experiments that may induce false alarms were also performed. Two approaches of machine learning algorithms were explore. The first prediction algorithms is based on Partial Least Squares Discriminant Analysis and the second set of algorithms are based on Support Vector Machines.
Additionally, two new methodologies for cost reduction are presented. The first methodology build fire detection algorithms using the combination of Standard fire test and a reduced version of those experiments. The reduced version were performed in a small chamber. The smaller setup allows the performance of experiments in a shorter period of time. In consequence, the number of experiments to test the models increase and also the robustness of the prediction algorithms. The second methodology built general calibration models using replicates of the same sensor array. The use of different units rejects the variance between sensor arrays and allows the construction of general calibration models. The use of a single model to calibrate sensor arrays systems allows the mass production and resulting in the reduction of costs production.
