Time-Efficient Aircraft Fault Isolation Procedures with NLP techniques
Technical talk | English
Technical talk | English
Theatre 16: Track 4
Wednesday - 12.35 to 13.15 - Technical
Aircraft fault isolation and troubleshooting are time-consuming tasks. Since after a flight operation, several failures related by chains of events are raised, the time needed to complete maintenance actions can be improved by identifying the origin of the failure by applying automated root-cause-analysis (RCA) and reducing the number of tasks required.
The main purpose of reducing the maintenance tasks is to increase the availability of the fleet. In addition to the associated person-hour costs of the maintenance, commercial operators of civil aviation suffer the additional monetary loss of an aircraft that is not generating revenue by transporting passengers and/or cargo. On the other side, military aircraft operators have a reduced operational efficiency due to the fact that not all its weapons or troops are mission-ready, leading to a reduced power-deployment capability.
Machine learning has demonstrated a great capability in different domains, with remarkably good performance. However, this capability has not really been proven in aircraft diagnostic and prognostic domain, where expert-based algorithms, engineering-driven physical models or probabilistic approaches are more trusted by health management engineers to determine aircraft or systems health. In this paper a new Machine learning technique is introduced to deal with RCA: a Knowledge Graph built with Natural Language Processing (NLP). In addition, this technique is compared with an already existing probabilistic approach, the Bayesian Network, leading to identification of significant advantages in the use of the Knowledge Graph over traditional approaches to RCA in terms of effort and time.
All the faults appearing during the flight are registered in the Post Flight Report (PFR) as Fault Codes. Each of these Fault Codes is related to a single aircraft fault isolation (AFI) procedure document, that explains to technicians how to identify accurately the equipment that failed, known as a Line Replaceable Unit (LRU). LRUs are equipments with a specific functional role in the aircraft that are identified unequivocally by an alpha-numeric number known as a Functional Identification Number (FIN).
The algorithm will be implemented for the fuel system. In this way, the nodes of the graph will be all the FIN related with the fuel system fault codes. The set of all the nodes can easily be gathered from a database of the aircraft manufacturer containing a list with all the FINs. Nevertheless, the information about the connection between nodes and its direction falls into what is known as ‘expert knowledge’. In order to acquire this information in an automated manner, NLP is applied to AFI documents.
One of the strengths of the NLP approach is that it requires no retraining and the same algorithm can be used for all the aircraft models of the portfolio of the company just by changing the AFI documents and re-learning the DAG. Another benefit comparing to the Bayesian approach is that is that the data is less biased. In the Bayesian approach, due to the fact that the algorithm uses historical data, any fault code that is not present in the database used to build the network, will not appear in it.
With the NLP approach, the global performance of the algorithm measured over six years of data reflects a potential to reduce by a 30% the time needed to complete fault isolation in aircraft maintenance. In this way, it is demonstrated how useful it can be to mix the ‘expert knowledge’ with data analytics tools.