Mike Gerdes

Decision trees and genetic algorithms for condition monitoring forecasting of aircraft air conditioning

Unscheduled maintenance of aircraft can cause significant costs. The machine needs to be repaired before it can operate again. Thus it is desirable to have concepts and methods to prevent unscheduled maintenance. This paper proposes a method for forecasting the condition of aircraft air conditioning system based on observed past data. Forecasting is done in a point by point way, by iterating the algorithm. The proposed method uses decision trees to find and learn patterns in past data and use these patterns to select the best forecasting method to forecast future data points. Forecasting a data point is based on selecting the best applicable approximation method. The selection is done by calculating different features/attributes of the time series and then evaluating the decision tree. A genetic algorithm is used to find the best feature set for the given problem to increase the forecasting performance. The experiments show a good forecasting ability even when the function is disturbed by noise.

Effects of condition-based maintenance on costs caused by unscheduled maintenance of aircraft

PurposeThis paper analyses the effects of condition-based maintenance based on unscheduled maintenance delays that were caused by ATA chapter 21 (air conditioning). The goal is to show the introduc ...

Fuzzy condition monitoring of recirculation fans and filters

A reliable condition monitoring is needed to be able to predict faults. Pattern recognition technologies are often used for finding patterns in complex systems. Condition monitoring can also benefit from pattern recognition. Many pattern recognition technologies however only output the classification of the data sample but do not output any information about classes that are also very similar to the input vector. This paper presents a concept for pattern recognition that outputs similarity values for decision trees. Experiments confirmed that the method works and showed good classification results. Different fuzzy functions were evaluated to show how the method can be adapted to different problems. The concept can be used on top of any normal decision tree algorithms and is independent of the learning algorithm. The goal is to have the probabilities of a sample belonging to each class. Performed experiments showed that the concept is reliable and it also works with decision tree forests (which is shown during this paper) to increase the classification accuracy. Overall the presented concept has the same classification accuracy than a normal decision tree but it offers the user more information about how certain the classification is.