Anomaly prediction on an autopilot system

Transport

Context

Some anomalies appear from time to time in autopilot systems which irritate pilots when they are flying the aircraft. An autopilot system is monitored by hundreds of sensors.

Need

The customer wanted to know firstly which sensors were most closely correlated with anomalies and secondly, for each anomaly, whether they could be predicted on a flight preceding the one where the anomaly might occur.

Solution

DiagFit was first used to find correlations between the 400+ data sensors and the anomalies; all sensors were ranked in order of correlation importance. 20 sensors with a correlation of more than 80% were selected to build a model predicting anomalies before they occurred.

Results

Classified sensors have enabled our customers to find the root causes of anomalies. The aim was to feed these root causes to its R&D team, to analyse potential changes to the system and to avoid such anomalies in the future. The predictive model was successful in predicting the occurrence of anomalies on flights preceding flights where anomalies could occur, which was the objective set by the customer.

 

The graph above shows the data acquired by a sensor over a period of 1.5 hours. The red areas show the anomalies that our model had predicted one flight earlier.