Engine fault diagnosis and modelling of related impact on engine performance

An engine fault diagnosis tool was created. The tool uses Support Vector Machine to identify root causes of in-service events to aircraft engines. The tool was developed in Python and it allowed to automatically identify new events root causes from past events. In-service events analysed regard v2500 Airbus equipped fleet between 2010 and 2018. Accuracy of the tool was tested on new events occurred in 2018 with a precision of 87 % in identifying the correct root cause for repetitive events. Digital flight recorder data DFDR were used to study impact of events on engine performance. Multiple regression analysis was applied to DFDR data to study anomalous variations during events. Regression curves were obtained for EPR , using N1, N2, EGT, Altitude and Fuel Flow as independent variables. Accuracy of regression analysis was proved by analysis of variance, residual plots, coefficient of determination, adjusted coefficient of determination and F-test. Lowest coefficient of determination was 0.91. Classes of events defined by root causes were studied, identifying in-flight engine behaviour for events as Dry Film Lubricant degradation, Variable Stator Vane loss of control, Bearing compartment distress, Stress Corrosion Cracking, Blade Outer Air Seal distress and faults in Fuel Metering Unit, Pressure Burner, Handling Bleed Valves, Fuel and Oil systems. A study of in-service stalls was conducted, defining conditions, root causes and consequences on aircraft missions and engine performance. Electronic Crew Alerting Management ECAM reports were used to identify affected systems and risk assessment analysis for in-service events. Primary Flight Report PFR was used as proof of ECAM warnings triggered during events and crew operations. Flight Crew Operating Manuals FCOM were studied to compare actual pilots decisions and expected decisions in events caused by pilots errors. An engine mechanical map to locate probability of events to determined engine components was produced. A model to calculate thrust from on-board recorded data was analysed.