Data mining within the aviation industry is becoming more important every day. Operators realise that maintenance costs can be reduced and fleet availability increased by analysing their data. Combining their maintenance records, flight recorded data and health monitoring information from the aircraft makes it possible to build new reliability models which can predict when components will fail. Thus, the research focused on how maintenance and flight recorded data of aircraft can be transformed into valuable information in order to find (flight) parameters which can predict component reliability, by using predictive data mining techniques within a model.
The first part focused on exploring the available data and selecting possible options. These options were researched according to three of the CRISP-DM methodology: data acquisition, data understanding, data preparation and data mining.
The output was that three data sources can be used:
- Flight Data Recorders (FDR)
- Aircraft Health Monitoring (AHM) sensor data
- AHM maintenance messages
The second part focused on how the selected data sources can be transformed into valuable information and which predictive data mining techniques are suitable for predictive maintenance on component reliability. This part was the most challenging one as huge amounts of data of all three sources must be obtained, prepared and analysed. As the time frame of my research was limited, not enough FDR data could be obtained in order to get a reliable output and therefore will not be further discussed. However, my colleagues who have been involved in the overall predictive maintenance research are still working on this part.
For the maintenance messages, the data preparation phase appears most challenging since only 57% of the original data was left after this phase. An approach I developed using matrix calculations shows good result to filter out the noise in the maintenance logs. A tool is developed to visualise the messages over intervals which presents a clear picture on when messages have started to appear and/or changed.
The matrix similarity analysis has shown good results on how to determine significant patterns which can be used as input in the online failure prediction technique. After testing, the matrix similarity method can be applied together with the online failure prediction technique. This will provide airlines with a reliable overview of the maintenance messages and a recommendation on when to replace a component based on its expected failure.
The AHM sensor data seems the most promising data source, since it poses no foreseen challenges so far. The techniques used for this kind of data though, rely heavily on the presence of multiple of the same component on the aircraft, and thus gets less reliable if more than one of these components is malfunctioning.
Overall it can be concluded that the developed prediction techniques in my thesis are very suitable since the overall goal of predictive maintenance of components is to find out the differences/anomalies between two groups: failure flights and non-failure flights. Furthermore, the techniques have shown excellent results.