Diagnostics and Prognostics of safety critical systems using machine learning, time and frequency domain analysis

Abstract: The prime focus of this thesis was to develop a robust Prognostic and Diagnostic Health Management module (PDHM), capable of detecting faults, classifying faults, fault progression tracking and estimating time to failure. Priority was to obtain as much accuracy as possible with the bare minimum amount of sensors as possible. Algorithms like k-Nearest Neighbors (k-NN), Linear and Non- Linear regression and development of rule engine to identify safe operating limits were deployed. The entire solution was developed using R (v 3.5.0). The accuracy of around 98% was obtained in diagnostics. For Prognostics, our ability to predict time to failure more accurately increases with time. Some balance must be there between learning horizon and predicting horizon in order to get good predictions with reasonable time left to hit catastrophic failure. In conclusion, the PDHM module works just as desired and makes Predictive maintenance, smart replacement and crisis prediction possible ensuring the safety and security of people on board and assets.