Research On Bearing Fault Early Warning Of Traction Motor For High Speed EMU

With the arrival of the "13th Five-Year Plan" ending battle,Chines high-speed rail network continues to expand,the speed is increased again,and the transportation demand is increasing.The problem of ensuring the operation safety of electric multiple units(EMU)is attracted much attention.As the core of EMU's power system,the stability of traction motor is the key to the safe operation of EMU.And the traction motor bearing is of uppermost priority.Nowadays EMU alarms the failure of the traction motor bearing by real-time monitoring and predicting the change of its temperature.The research of bearing fault early warning is still the focus of prognostics health management(PHM).In the real environment,the temperature of the operating bearing is influenced by different kinds of complicated factors such as present working condition,external environment,component performance and so on.Many researches on vibration signals based on experiments have great limitations in practical applications.Currently EMU adopts a fixed threshold for the bearing fault early warning,which having high false alarm rate.In the operation data,the normal bearing temperature data and the bearing fault data are seriously unbalanced,which is difficult to directly classify the faults.In response to these problems,a temperature prediction model based on the operation data of EMU was put forward to predict the bearing temperature.And the predicted temperature was used as a dynamic threshold to assist in the early warning of failures.It can alarm the abnormal condition of the bearing in time,improve the safety factor of EMU and reduce the operation and maintenance cost.The specific research content has the following aspects.(1)Analyze the characteristics of EMU operation data and the correlation among them.And a Multi-Task Learning and Attention Mechanism Based Long Short-Term Memory(MTL-AM-LSTM)model is proposed to predict the bearing temperature of traction motor in actual situation.The attention mechanism is integrated into the Long Short-Term Memory network,and the temperature at different locations of the same traction motor is learned jointly to build the axial temperature prediction model including multiple influence factors and spatio-temporal correlation of bearings.In addition,two losses of the model are optimized.The quantile loss function is applied to solve the problem that the real value is more concerned than the predicted value in practice.And a joint loss based on gradients is put forward to automatically assign different weights to each task,which can avoid the loss of different scale,the influence of different convergence speed and the low efficiency of manual weight adjustment.(2)According to the different trend of bearing temperature and the fuzziness of different running states,the operation state of EMU is identified automatically by fuzzy membership function.Then combined with the MTL-AM-LSTM model in different states,a temperature prediction model of the traction motor bearing based on multiple-model fusion is generated.Meanwhile,through the prediction model of bearing temperature studied in this paper,the temperature curve of traction motor bearing in EMU is fitted.The predicted temperature is regarded as the dynamic threshold to assist the current fixed threshold method in fault early warning.Finally,in the actual operation datasets of EMU,the prediction effect of the MTL-AM-LSTM model and the effectiveness of bearing fault early warning based on the dynamic threshold is verified.The experimental results show that our method performs the best with other models and the model optimization is effective.To a certain extent,the method of dynamic threshold reduces the false positive rate.