Research On Remaining Life Prediction Method Of Gear Based On ON-LSTM Network And Meta Learning Reinforcement

Gear is a key component of mechanical equipment which located in the gearbox.Compared with shaft,bearing,shell and other parts of gear box,gear has higher failure probability.Fatigue pitting is one of the most common failure forms of gear.Gear failure may affect the operation of mechanical equipment,and even cause the damage of machine and even casualties.Thus,the remaining useful life(RUL)estimation of the gear has important significance.The RUL prediction of gear is mainly based on the health index extracted by vibration signal of gear.The trend of the health index is used to predict the RUL.At present,the data-driven method has limited ability in long-term prediction.Therefore,this thesis uses machine learning to carry out research on the methods of longterm prediction of RUL and rapid short-term prediction,and designs a gear failure RUL system based on the two prediction methods.Firstly,this thesis proposes a gear RUL prediction method based on ordered neuron long short-term memory(ON-LSTM)network.The proposed methodology consists of two parts: Extract the health index by computing frequency-domain features of raw signals;The ON-LSTM network model is constructed for generating the target output of the RUL prediction.Unlike the traditional LSTM neural network,the developed model integrating tree structures into LSTM to use the sequential information between neurons,so it has enhanced predictive ability.In comparative experiments,the scores of ONLSTM is the best compared with LSTM,GRU,DLSTM and DNN;and ON-LSTM successfully fulfils 23 tasks while LSTM just fulfils 5 tasks in long-term prediction.Secondly,to solve the problem of long training time caused by the retraining of the small sample prediction model,a gear residual life prediction method based on metalearning and multi-step error correction was proposed.It is found that the Model-Agnostic Meta-Learning(MAML)can train a set of initialization parameters with good generalization performance for the prediction model.Under the condition of ensuring accuracy,the parameters are continuously updated based on the historical data,which greatly reduces the training time.In order to further improve the prediction accuracy of model,an error correction method based on imitative learning is proposed to reduce the accumulated errors.The experimental results show that the method have the advantages of shorter training time and higher accuracy compared with other methods.Finally,a gear failure RUL system based on two RUL prediction methods proposed in this thesis is designed,which includes information management,signal analysis,life prediction functional modules.The RUL of the gear is predicted with the experimental data,and verify the availability of the system.