Recurrent Neural Network Based on Statistical Recurrent Unit for Remaining Useful Life Estimation

In industry, efficient predictive maintenance tools can reduce maintenance costs and increase the safety and reliability of the monitored equipment, since they can anticipate equipment failures. In particular, making the efficient Remaining Useful Life (RUL) estimation of machinery is important to lead to appropriate maintenance actions. Traditional RUL approaches depend on prior knowledge of the equipment degradation process to predict RUL. However, in most cases, the accurate physical or expert models are not available. Following that, this paper proposes a Recurrent Neural Network (RNN) model architecture based on Statistical Recurrent Unit (SRU) for RUL estimation. The proposed architecture is able to extract hidden partners from multivariate time series sensor data with multiple operation condition faults and degradation. SRU outperforms other complex deep learning methods, since it obtains a multitude of past viewpoints by linear combinations of few averages. The proposed model is compared to state-of-the-art RUL approaches. Experimental results, using a turbofan aero engine data set, reveal that the proposed architecture outperforms state-of-the-art RUL approaches in most tests.