Multi-dimensional recurrent neural network for remaining useful life prediction under variable operating conditions and multiple fault modes

Data-driven remaining useful life (RUL) prediction approaches, especially those based on deep learning (DL), have been increasingly applied to mechanical equipment. However, two reasons limit their prognostic performance under variable operating conditions. The first one is that the existing DL-based prognostic models usually ignore the utilization of operating condition data. And, the other is that most DL-based prognostic models focus on enhancing the nonlinear representation learning ability by stacking network layers, and lack exploration in extracting diverse features. To break through the limitation of prediction accuracy under variable operating conditions, this paper presents a novel multi-dimensional recurrent neural network (MDRNN) for RUL prediction under variable operating conditions and multiple fault modes (VOCMFM). Different from existing DL prognostic models, MDRNN can simultaneously model and mine multisensory monitoring data and operating condition data to achieve RUL prediction under VOCMFM. In MDRNN, parallel bidirectional long short-term memory and bidirectional gated recurrent unit pathways are constructed to automatically capture degradation features from different dimensions. Two prognostic benchmarking datasets of aircraft turbofan are adopted to validate MDRNN. Experimental results demonstrate that MDRNN can perform the prediction tasks under VOCMFM well and surpass many state-of-the-arts.