Research On Data Driven Remaining Useful Life Prediction For The Gears Of Air Turbine Starter

The air turbine starter is the core of the civil aircraft engine starting system.As the key power transmission element in the ATS,the gears can smoothly complete the mechanical power transmission,which is the precondition of the reliable operation of the ATS.Due to long-term work in a complex condition,ATS gears have a high failure rate.Therefore,it is of great engineering and theoretical significance to accurately predict the remaining useful life(RUL)of ATS gears,and then to formulate scientific condition-based maintenance plans for gears.The main research work done in this thesis is as follows:Firstly,the physical structure of the ATS and typical gear failure modes are studied.This part deeply analyzes the physical structure and starting principle of ATS,focusing on the common failure modes of ATS gears,which lays the foundation for the follow-up research on RUL prediction.Secondly,the Long Short-Term Memory model guided by attention mechanism is proposed to accurately predict the RUL of ATS gears.On the basis of in-depth analysis of the gears’ vibration signal,four time-domain characteristics that can best reflect its operating conditions are selected from the vibration signal,namely root mean square,kurtosis,variance and margin indicators,which is to construct a new data set and serves as the input to the RUL prediction network.The RUL prediction network is based on LSTM,and two attention layers are introduced to build the RUL prediction model.Among them,the first introduced self-attention layer is set before the LSTM layer to improve the weight of the inputs;the second introduced attention layer is set after the LSTM layer to improve the weight of the output results of the LSTM at different times,thereby improving the accuracy of the RUL prediction of the ATS gears.Finally,experimental validation and evaluation tests are carried out.The accelerated fatigue test environment of ATS gears is designed,and the hardware and software conditions are configured to carry out accelerated fatigue experiments on gears.Through the gears information management and data acquisition system,we set the appropriate sampling frequency and sampling duration,and have completed acceleration fatigue experiment of 4sets of gears.The performance of the proposed method is verified by the obtained data,and comparative experiments are set up.The experimental results show that the proposed ATS gears RUL prediction network performance meets the requirements.