Cocktail LSTM and Its Application Into Machine Remaining Useful Life Prediction

In the industrial field, gearbox and bearing are the key components of machines, and their health statuses are crucial to the reliable operation of machines. Therefore, the remaining useful life (RUL) prediction of gearbox and bearing is of great significance. To accurately predict the long-term RULs of gearboxes and bearings, a novel multihierarchy network based on multiordered neurons, namely, cocktail long short-term memory (C-LSTM), is proposed. First, the 21 time-frequency characteristics are extracted from the collected vibration signals, which are then input into the trained variational autoencoder to construct the health indicator (HI) with the distinct degradation trend. Next, the generated HI points are fed into C-LSTM for predicting the future HI points. With regard to C-LSTM, the existing HI vector is first divided to get the multihierarchy, and different update rules are put forward to extract various trend information from the known HI points based on the hierarchy result, then the future HI points are predicted in sequence until the threshold is exceeded. It then follows that RUL can be calculated. The proposed methodology has been verified using gearbox datasets and the IEEE 2012 bearing dataset, and the comparative results in Score, mean absolute percentage error, mean absolute error, and normalized root-mean-square error show that C-LSTM has a higher comprehensive predictive performance than the traditional prediction methods.

Automated summary

In the industrial field, accurately predicting the remaining useful life (RUL) of gearboxes and bearings is crucial for reliable machine operation. A novel multihierarchy network called cocktail long short-term memory (C-LSTM) is proposed to achieve this. It extracts time-frequency characteristics from vibration signals to construct a health indicator (HI) with distinct degradation trends, and uses C-LSTM to predict future HI points based on the known HI points. The proposed methodology shows higher predictive performance compared to traditional methods in gearbox and bearing datasets.