Remaining useful life prediction of rolling bearing based on multi-head attention embedded Bi-LSTM network

Accurate rolling bearing remaining useful life (RUL) prediction is significant to evaluate machine health state. The condition monitoring signals of rolling bearings contain amounts of redundant information, which is detrimental to the analysis of degradation process. To meet this challenge, a novel data-driven method is proposed in this paper. Firstly, several time and frequency domain features of vibration signals are extracted and selected by comprehensive evaluation, then a non-linear degradation indicator (NDI) is constructed to describe the degradation process. Next, an attention mechanism is given to the network named Multi-Head Attention Bidirectional-Long-Short-Term-Memory (MHA-BiLSTM), which can adaptively weaken the impact of redundant information. Finally, NDI is fed into MHA-BiLSTM, which is used to fit the degradation curve and predict the RUL of rolling bearings. The proposed method is evaluated on two datasets from testbed and real wind farm. The comparison experimental results reveal the superiority of proposed method.