The POC (Proof of Concept) consisted of two phases of training and testing - Phase 1 proving out the development of a predictive model, and Phase 2 proving out the quick turnaround in model improvement. Applying a data-driven approach, along with collaboration with the airline maintenance team, Amygda developed a model in half the expected delivery time, at a 78% detection rate, improving to 92% with rapid iteration.
The solution was built on the back of 3 years of Flight Operations and Quality Assurance (FOQA) data for the entire fleet at 4Hz, equating to approximately 650,000 flights, accompanied by maintenance data. Through data-driven labelling techniques, Amygda was able to re-label and date the incidents pertaining to ACMP faults. By modelling the operation of the hydraulic system, we generated effective system cycles which allowed us to engineer and trend homogenous features and detect impending failures. Amygda's approach, along with feature engineering, helped us not only predict ACMP faults, but pinpoint exactly which of the hydraulic ACMPs was at fault. Our solution differentiated between real faults and planned maintenance activities, which were previously leading to false positives.