Research On Prediction Of Remaining Useful Life For The Gears Of Air Turbine Starter Based On LSTM

The air turbine starter is the key component of the engine starting system,and gears are significant parts of ATS.Due to the severe working environment,gears have a high failure rate,which is the key factor affecting the normal operation of the ATS.Accurate ATS gears remaining useful life prediction can provide theoretical direction for preventative maintenance,and be beneficial to make reasonable maintenance plan.Aim at the problem of ATS gears remaining useful life prediction,the main work and contributions of research are as follows:First of all,the physical structure of the ATS and typical gear failure modes have been studied.This part study the internal structure and working mechanism of the typical ATS deeply,as well as the common ATS gear failure,which provides foundation for the subsequent research on the prediction of ATS gears remaining useful life.Secondly,the remaining useful life prediction hybrid model of ATS gears based on Long Short-Term Memory network was established.The prediction model based on LSTM realizes gears remaining useful life prediction through analyzing and selection the time domain feature of ATS gear vibration signals、constructing degradation characteristics evaluation index、confirming model parameters、training and testing model.Compared with traditional models BPNN and RNN,the proposed model enhances the prediction accuracy and reduces the online calculation time and cost.Subsequently,the gear deep degradation feature extraction method has been proposed.the gear vibration signals have been decomposed into several IMFs,which were evaluated and selected by the method representative IMF selection index to obtain the representative IMF.Compared with various IMF selection algorithms,the proposed method has superiorities that reducing calculation amount and boosting prediction accuracy.Finally,verify the performance of the proposed method.Collecting the data from accelerated gear degradation test platform,and compared the real degradation result with results of different prediction methods,indicating the proposed method has better prediction accuracy and shorter training time.