Multi-head self-attention bidirectional gated recurrent unit for end-to-end remaining useful life prediction of mechanical equipment

Abstract

Abstract In order to reduce error accumulation caused by multistep modeling and achieve a generally accurate model, this paper proposes an end-to-end remaining useful life (RUL) prediction model based on a multi-head self-attention bidirectional gated recurrent unit (BiGRU). Taking multivariable samples with long time series as the model input and multistep RUL values as the model output, the BiGRU model is constructed for continuous prediction of RUL. In addition, single-head self-attention models are applied for time series and variables of samples before or after the BiGRU, which can be fused into a multi-head attention BiGRU. Aeroengines and rolling bearings are selected to testify the effectiveness of the proposed method from the system level and component level respectively. The results show that the proposed method can achieve end-to-end RUL prediction efficiently and accurately. Compared with single-head models and individual deep learning models, the prediction mean square error of the proposed method is reduced by 20%–70%.