Application of Rare Event Anticipation Techniques to Aircraft Health Management Stephane Alestra 1,a , Cristophe Bordry 2,b , Cristophe Brand 1,c , Evgeny Burnaev 1,3,4,d , Pavel Erofeev 1,3,4,e , Artem Papanov 1,3,f , and Cassiano Silveira-Freixo 2,g 1 DATADVANCE, Pokrovsky blvd. 3, Moscow, 109028, Russia 2 AIRBUS, St. Martin du Touch, 316 route de Bayonne, 31060 Toulouse Cedex 9, France 3 IITP RAS, Bolshoy Karetny per. 19, Moscow, 127994, Russia 4 PreMoLab, MIPT, Institutsky per. 9, Dolgoprudny, 141700, Russia a stephane.alestra@datadvance.net, b christophe.bordy@airbus.com, c christophe.brand@datadvance.net, d evgeny.burnaev@datadvance.net, e pavel.erofeev@datadvance.net, f artem.papanov@datadvance.net, g cassiano.freixo@airbus.com Keywords: Predictive Maintenance, Rare Event Anticipation, Aircraft Health Management. Abstract. Generally faults in complex technical systems (such as aircrafts) can be considered as rare events. In this paper we apply classification techniques to problem of rare events anticipation and demon- strate the approach to predictive maintenance of aircrafts through the real-world test cases from aircraft operations based on the data granted by AIRBUS. Introduction In recent years the concept of predictive maintenance in complex technical systems is becoming more and more popular. The baseline idea of the concept is to use information given by system being in-service (usually measurements of some sensors) in order to determine its condition and finally predict when main- tenance should be performed. This approach is less expensive comparing to traditional ones (routine or time-based) because maintenance actions are performed only when they are actually needed. There are several examples of successful application of this approach in different areas: US navy cost reduction for maintenance [1], increase of safety and reliability of distributed power systems [2] and automated search for faults in power systems [3]. In this paper, aircraft maintenance problem is examined. Traditional maintenance approach in this field is to perform scheduled activities, such as structural inspections or electronic tests, in order to de- tect failures and eliminate causes of them. This approach misses all faults which appear between these scheduled events, so if they become obvious then emergency activity will be performed in order to restore aircraft functions as quickly as possible. Thus traditional approach also implies unscheduled maintenance events which are usually expensive. Trying to reduce maintenance costs, AIRBUS is interested with predictive maintenance concept as a methodology of failure anticipation and warning monitoring function to decide whether a operability- related failure is present in the aircraft before a fault actually occurs. The goal of the ongoing project in AIRBUS is to develop a full support automated system for the early warnings for possible costly faults. Classical statistical approaches are ineffective for low frequency and high consequence events because of their rarity. In this paper we try to adapt existing classification methods for rare event prediction. Rare Event Anticipation Problem Statement From mathematical point of view the rare event anticipation problem can be formulated in the following way [4]. We observe in real time (with some frequency, possibly not uniform) starting $GYDQFHG 0DWHULDOV 5HVHDUFK 9RO SS 7UDQV 7HFK 3XEOLFDWLRQV 6ZLW]HUODQG GRLZZZVFLHQWLILFQHW$05 $OO ULJKWV UHVHUYHG 1R SDUW RI FRQWHQWV RI WKLV SDSHU PD\ EH UHSURGXFHG RU WUDQVPLWWHG LQ DQ\ IRUP RU E\ DQ\ PHDQV ZLWKRXW WKH ZULWWHQ SHUPLVVLRQ RI 773 ZZZWWSQHW ,' DRAFT