Créée il y a 70 ans, l’ENAC, École Nationale de l’Aviation Civile, est la plus importante des Grandes Écoles ou universités aéronautiques en Europe. Elle assure la formation initiale et le perfectionnement des cadres et des principaux acteurs de l’aviation civile : ingénierie, navigation aérienne, pilotage, gestion aéroportuaire, recherche, expertises, développement à l’international.
L’ENAC est un établissement public à caractère scientifique, culturel et professionnel – grand établissement (EPSCP-GE), sous tutelle de la DGAC (Direction Générale de l’Aviation Civile), Direction du Ministère de la Transition Écologique et Solidaire. L’ENAC comprend une direction générale localisée à Toulouse et 8 sites en France.
In a so-called ultra-safe industry like aviation, it is now widely acknowledged that relying on accidents to enhance safety is not sufficient. In order to have access to feedback from operations, including less critical situations than accidents or significant incidents, reporting systems have been introduced in aviation at several levels, leading to databases of spontaneously reported events that might be of interest from a safety viewpoint. ECCAIRS and ASRS are examples of such databases at regional level in respectively Europe and North America. In addition to these regional initiatives, a number of aviation organizations have their own internal reporting systems allowing them to collect events considered safety relevant by the reporter. These reports constitute a unique source of qualitative information on operations that are nowhere else available. They might come in addition to FDM data or other sources of structured data collected in a systematic way but based on ‘sensored’ parameters only. In other words, they are a very valuable source of information to develop insights on the safety of operations overall. However, the events reported through these channels do not provide a complete and accurate picture of real operations. Indeed, first of all, reporting is not mandatory for events that are not critical enough to fall under the ICAO Annex 13 or other mandatory reporting scopes (e.g. EASA regulations or national/regional laws). Therefore, the appreciation of whether an event might be of interest to safety and is worth reporting is left to potential reporters and may vary from one individual to another as it may vary from one organization to another. Besides, the way an event is reported relies on the perception, understanding and analysis of the individual who reports and might differ from objective facts. Eventually, the database of reported events provides a partial a biased view of real operations. A major challenge to still use this data but for its actual value is to know more about the extent to which it represents operations.
The objective of this research is to explore through statistical approaches the representativeness of the sample of the events reported to find ways to qualify the safety concerns it highlights and adjust the safety management strategy accordingly. Yet, the representativeness of reported events may vary from one type of event to another, from one airline to another, from one period of time to another, from one aircraft type to another... In other words, characterizing the parameters/dimensions along which to define the representativeness is part of the research effort, as well as developing or adapting methods to assess the probability for a given type of event to be reported for a given type of aircraft...
The first step consists in determining a model to detect the occurrence of an event based on exhaustive sources of data such as PFR reports or time series. The outputs can either be the type of event detected (or absence thereof), or a vector of probabilities associated to each event. The probability of a given event being reported can then be inferred by confronting the outputs to the actual presence/absence of a report. Given the high number of available data, complex methods such as Random Forests or Neural Networks can be considered to derive a probability of being reported to each PFR report, associated to the available information.
As a second step, the algorithm can be applied to the spontaneously reported events and occurrences databases. The output of the algorithm can then be interpreted as a proportion of such events being reported in the "whole population" of safety related events. Techniques of calibration used in survey sampling can then be applied to correct the possible biases of reports pertaining to different types of events, as a way to improve the representativeness of the events.
- Knowledge and experience in machine learning
- PhD in applied mathematics, especially Probability or Statistics would be appreciated
- knowledge of machine learning language like R or Python
- Quick learning
- Strong communication skills, ability to interact with people having diverse profiles and interests
- Advanced spoken and written English
- Strong interest in research
- Interest in aviation
Contrat ENAC - 24 mois à compter du 1er Septembre