Aircraft Reliability Prediction Using Bayesian Networks That Combine Fault Data And Design Specifications

Abstract

Reliability engineering focuses on analyzing the properties related with failure times of high valued items. Bayesian Networks (BNs) are an effective way of analyzing causal relations, they allow incorporating expert judgement into mathematical models, and they are able to make probabilistic calculations when only a part of their variables are known. This makes BNs a suitable tool for reliability analysis considering the cost of life testing high-value products. Hence, numerous BNs and Bayesian approaches have been proposed for reliability estimation and prediction analyses. Some of these approaches tend to focus on more on building discrete nodes as BN models for developing a subjective judgement where others are more focused on failure time distribution parameters and mathematical properties of the Bayes Theorem. This thesis proposes a novel BN model for predicting the time to failure distribution of an aircraft fleet, by a bottom to top approach. Our model incorporates both actual failure data and the expert judgement on design and manufacturing qualities of the aircrafts by using BNs. The expert judgement is based on the design life estimations provided by the manufacturer of the aircrafts and these values are transformed into a distribution, reflecting our uncertainty associated with it. Then this prior information is integrated to the Weibull distribution as median parameter and used for obtaining scale parameter. We applied our model to make reliability prediction by using failure data provided by an aircraft fleet operator, after preprocessing the raw data into a structure suitable for reliability analysis. We compared the performance of our model in predicting the reliability of the main systems of the aircrafts to commonly used reliability estimation methods. The proposed model offers a robust approach by giving consistently satisfactory results compared to the purely data-driven approaches and design life estimations. As the sample size increase, the performance of the model becomes very similar to the data-driven approaches. This is expected as the effect of the priors used in the model decreases with as the size of the data increases. We have also used a different prior distribution for shape parameter of the Weibull distribution, compared to standard approaches in the literature, and applied it to the aircraft fleet data.