Data Mining of Production Accept Test Data for aircraft electronic engine controllers
(EEC)
Each EEC manufactured is submitted to around 3000 functional tests prior to dispatch to the engine manufacturer. The tests are pass/fail tests and are designed to identify any manufacturing weaknesses, correct the faults and retest. The pass data is stored and little analysis is conducted on the data. Due to the amount of data and number of tests
(typically 2000 tests per unit, 2 lanes per unit, 3 temperature regimes) it is difficult to analyse the data manually. A TSB (formerly DTI) research project (IMPRESS) is investigating the use of data mining techniques to discover whether such techniques can provide engineers with knowledge about the product design and test. This data is unlikely to be conventionally analysed because it is pass data, however there may be some knowledge within the data that can help identify design marginality. Moreover when combined with in-service failure data, patterns maybe detected that can be used to develop a diagnostic tool for service performance of the EEC.
The IMPRESS project is concerned with 3 types of data i.e., PAT, in-service failure data and data stored within the EEC. The overall purpose of the project is to investigate the use of data mining and Bayesian networks of these data sets to aid diagnostic activity at the repair and overhaul facility, develop a prognostic capability, reduce variability during the design phase and aid aircraft troubleshooting.
This mini-project is concerned with evaluating data mining techniques for PAT data.
This would be particularly interesting to students who have knowledge of statistical methods and who could use their knowledge to evaluate which and how data mining could be implemented to increase product design knowledge. It is interesting because conventionally this type of data is not explored except when issues are raised on particular tests. A student would have access to all the PAT data for one product (3000 files available), access to engineers and a researcher working on IMPRESS and access to reports written describing the data thus the project would involve:
initial data processing;
identifying appropriate data mining tools;
designing an experiment for employing the tools;
developing criteria for comparison; and
reporting on conclusions and recommendations.
The project is using an open source software package called WEKA and so far we have started investigating clustering methods
A follow on mini-project could involve investigating any patterns in the PAT data that can be linked to failures in-service (data available on same product) thus developing prognostics capability from the data. This mini-project would involve identifying and experimenting with methods or algorithms that can uncover interesting patterns in the data.
Finally these mini-projects could lead to a PhD that is focused on data mining methods that can identify patterns in complex engineering data to enhance understanding of the causes of failure in complex electronic equipment.