Fire Detectors Based on Chemical Sensor Arrays and Machine Learning Algorithms: Calibration and Test

Ana Maria Solorzano

Santiago Marco Colás Jordi Fonollosa Magrynà

Conventional fire alarms are based on smoke detection. However, fires often emit many volatiles before they emit smoke. This has opened the door to the development of fire detection systems based on chemical sensors, which can provide a faster response. However, these systems are still prone to false alarms due to interference. This reason prevents the commercialization and standardization of fire detectors based on gas sensors. The use of pattern recognition techniques may be the key to mitigating this limitation. In this thesis, two fire detectors based exclusively on gas sensors, from different technologies, have been developed, which provide a fire alarm based on machine learning algorithms. The detectors were exposed to standardized fires and to several relevant interferences in domestic and industrial environments. The results confirm the ability to detect fires at an early stage of their development and the rejection of most interferences. In addition, two methodologies are presented for reducing the calibration costs of gas sensor clusters for fire detection, bearing in mind that the experiments to evaluate the detectors are carried out in a standard fire room and are very long and expensive. The first proposed methodology combines data from a standard fire room and data from experiments carried out on a small scale, faster and less expensive. The results show that the performance of prediction models can be improved with data fusion. The second cost-reduction methodology compensates for the need for individual calibration models for each sensor array (due to sensor variability) by rejecting sensor variability and providing general calibration models.

Currently there is not a commercial detector based exclusively on gas sensors. The advantages of gas detection in fires encourage the development of gas detectors for commercial applications. The thesis "Fire Detectors based on Chemical Sensor Arrays and Machine Learning Algorithms: Calibration and test" presented for the Pioneer award 2020 was framed in the European Project SAFESENS. The project has as objective the co-integration of gas sensor technologies that enable and enhance safety and security in fires. The EU project involved several companies and universities. Some of the companies are AMS, Imec and Bosch, with experience in the gas sensors field; the Minimax Company, with the know-how of fires; and IBEC with experience in machine learning. The commercial and industrial interest in the results stands out with the participation of the biggest fire solutions industries such as Bosch and Minimax. The discrimination power of the machine learning algorithms relies (among other factors) on the selection of the best set of sensors for fire detection. In this way, besides the gas sensor array developed in the EU project, I designed and tested other gas sensor arrays using commercial and available sensors. This comparison provides a real hint of the possibility to build gas sensor arrays with the current and commercial gas sensor technologies, resulting in a fast development of a new product for the fire industry. On the other hand, it is important to train algorithms that provide fast and real time detection. For that reason, I built two different methodologies, one based on Partial Least squares -PLS-DA- and one on Support Vector Machines -SVM-. The use of these algorithms allows the validation and certification of the system for commercial and safety applications. Additionally, my thesis was focused since the beginning on the industrial application of the results. For that reason, I also developed two different methodologies for the reduction of calibration and training costs. It is well known that building robust machine learning algorithms requires a large number of samples. The performance of fire experiments is a time consuming task. In that way, I developed a methodology that allows the fusion of different fire experiment setups (data fusion) to speed up the test time. This methodology reduces the experimental costs by a factor of 20. The second methodology presented is a solution for the calibration cost. Without this methodology, every single gas sensor array detector requires to be individually calibrated. With our proposed methodology, only one general (global) calibration model is built for all the gas sensor arrays developed. This second methodology reduces the calibration costs, making the calibration procedure scalable (mass production).

Conventional Fire Alarms; Smoke Detection; Fire Volatiles; Chemical Sensors; Faster Response; False Alarms; Interference; Commercialization; Standardization; Gas Sensors; Pattern Recognition Techniques; Machine Learning Algorithms; Standardized Fires; Domestic Environments; Industrial Environments; Early Stage Fire Detection; Interference Rejection; Calibration Costs Reduction; Gas Sensor Clusters; Standard Fire Room; Small Scale Experiments; Data Fusion; Prediction Models Performance; Individual Calibration Models; Sensor Variability; General Calibration Models.