A deep bi-directional long short-term memory model for automatic rotating speed extraction from raw vibration signals

Abstract

The rotating speed information is significant for condition-based monitoring of rotating machines which are often operated under varying speed conditions. To automatically extract rotating speed from vibration signals, a deep learning model named many-to-many-to-one bi-directional long short-term memory (MMO-BLSTM) model is proposed. The proposed model consists of two parts: (1) the many-to-many BLSTM part (BLSTM part) and (2) the many-to-one LSTM part (LSTM part). The BLSTM part learns speed related information from vibration signals in both forward-time and backward-time directions. The final speed is successively extracted via the LSTM part from the information learned by the BLSTM part. The proposed MMO-BLSTM model is trained via a supervised pre-training and fine-tuning strategy. The performance of the proposed model is validated with an internal combustion engine dataset, a rotor system dataset and a fixed-shaft gearbox dataset. The results show that the proposed model achieves a higher speed extraction accuracy than some reported models.