| To describe the change of aeroengine lubricating oil quantity during flight, a wavelet-LSTM model was proposed by
combining the advantages of wavelet decomposition and the Long Short-Term Memory neural network (LSTM). Inputs of the model are in
the form of multi-variable time series data, consisting of key factors such as engine high-pressure rotor speed, low-pressure rotor speed,
flight altitude, flight attitude, etc, and the output is the lubricating oil quantity sequence. Real flight Quick Access Record (QAR) data were
used, and the data of the descending phase of the flights were selected for modeling due to their high fluctuation in lubricating oil quantity.
The raw data of oil quantity was calibrated to eliminate thermal effects. Eleven factors highly related to lubricating oil quantity were
selected as inputs, including high-pressure rotor speed, flight attitude, flight altitude, flight speed, etc. Wavelet decomposition was applied
to both the input data and the lubricating oil quantity data after thermal calibration to remove noise and reduce data size, to accelerate the
training speed of the subsequent machine learning model. LSTM neural network was adopted in the machine learning step to calculate the
lubricating oil quantity based on input factors. The results show that based on real flight data, the mean squared error between the
calculated lubricating oil quantity and the real lubricating oil quantity is about 0.1 square liters, demonstrating the effectiveness of the
proposed method in modeling the change of lubricating oil quantity during the aircraft descending phase and the applicability of the
proposed method in health monitoring, providing a new method in support of predictive maintenance of engine lubrication system. |