Genetic algorithms and decision trees for condition monitoring and prognosis of A320 aircraft air conditioning

Unscheduled maintenance is a large cost driver for airlines, but condition monitoring and prognosis can reduce the number of unscheduled maintenance actions. This paper discusses how condition monitoring can be introduced into most systems by adopting a data-driven approach and using existing data sources. The goal is to forecast the remaining useful life (RUL) of a system based on various sensor inputs. Decision trees are used to learn the characteristics of a system. The data for the decision tree training and classification are processed by a generic parametric signal analysis. To obtain the best classification results for the decision tree, the parameters are optimised by a genetic algorithm. A forest of three different decision trees with different signal analysis parameters is used as a classifier. The proposed method is validated with data from an A320 aircraft from Etihad Airways. Validation shows that condition monitoring can classify the sample data into ten predetermined categories, representing the total useful life (TUL) in 10% steps. This is used to predict the RUL. There are 350 false classifications out of 850 samples. Noise reduction reduces the outliers to nearly zero, making it possible to correctly predict condition. It is also possible to use the classification output to detect a maintenance action in the validation data.