Acoustic detection of start of fires in industrial sites – DAFI

The DAFI project presents the development and the evaluation of a detection system of start of fire in industrial sites, using sound modality. This proposal responds to the RA-SIOMRI call for the project, within the framework of the axis 1, by providing an operational and technological response to risk control on industrial sites. The use of the sound modality makes it possible to detect many events, sources of potential risk in industrial sites, such as the shouts of operators in the event of an accident, the impact sound in the event of an accident or the sound of explosion or start of fire. In this project framework, we propose to focus on the detection of the start of fire, a significant source of major disaster in industrial sites.
The consortium for this project is composed of a research laboratory, the LEOST of Gustave Eiffel University, and the company Wavely. For several years, the first of these participants has focused part of his research on the development of automatic surveillance systems using sound modality. This scientific research focuses on the detection of abnormal or critical sounds applied to security and safety contexts, in transport environment. The second is the French startup Wavely, which develops solutions for large industrial companies, combining both the deployment of acoustic sensor networks and the development of AI algorithms. Wavely can process a large amount of acoustic data in real time in order to detect and classify many sound sources in various industrial contexts: construction site, gas leak detection and biodiversity monitoring with TOTAL among others.
Over a period of 24 months, LÉOST and Wavely wish to unite their respective skills to develop an innovative solution, using sound information related to an industrial environment, an operational solution, using an infrastructure already proven on industrial sites, and finally a complete solution by associating connected sensors and fundamental of artificial intelligence in considering Machine Learning and more recently of Deep Learning concepts.
Based on the state of the art of automatic sound patterns recognition, the algorithmic developments will aim to statistically model the so-called “normal” industrial sound environment, to model relevant patterns linked to risky situations and more particularly to model sound patterns relating to the start of fire. For these various models, learning phases will be necessary and will require judiciously to place sensors network in industrial site in order to record the sound environment and secondly to perform safe sound recording of the start of fire. The various phases of testing in the laboratory initially and in site in a second phase will allow the detection system to be tested before it is operated in real conditions.
The final objective will be to have a start of fire detection system that can be deployed on a large set of industrial sites, in particular large SEVESO-type sites, in function possible adaptation procedures depending on the specificities and variability of sound environments.