A novel dual attention mechanism combined with knowledge for remaining useful life prediction based on gated recurrent units

Improving Remaining Useful Life (RUL) prediction accuracy in Prognostic and Health Management (PHM) is the primary pursuit of researchers. Deep learning provides opportunities to address critical reliability engineering and safety analysis challenges. However, the performance could be improved due to the quantity and quality of sensor data. Knowledge is available to enhance the deep learning effect to improve RUL prediction performance. This paper proposes a novel RUL prediction approach that implements an encoder/decoder architecture with Gated Recurrent Units (GRUs) and dual attention mechanism combined with domain knowledge. The dual attention mechanism includes time and knowledge attention: knowledge attention reviews the information across the feature axis using domain knowledge to decide which sensor data is more critical. And time attention extracts the relatively necessary multiple time steps from the essential features. Compared to other deep learning models with attention using the NASA open dataset C-MAPSS, the approach reached first place in the overall ranking, demonstrating the reliability of utilizing domain knowledge in deep learning training. The results indicate that the proposed approach not only produces a superior accuracy of estimation accuracy but also benefits the efficiency and interpretability of the prediction.

Automated summary

Improving Remaining Useful Life (RUL) prediction accuracy is the main focus in Prognostic and Health Management (PHM). Deep learning offers opportunities to address reliability engineering and safety analysis challenges, but performance can be enhanced with better sensor data quality and quantity. This paper proposes a novel RUL prediction approach that combines an encoder/decoder architecture with Gated Recurrent Units (GRUs) and a dual attention mechanism, leveraging domain knowledge. Compared to other attention-based deep learning models, the approach achieved first place in overall ranking on the NASA C-MAPSS dataset, demonstrating the reliability of incorporating domain knowledge in deep learning training. The results indicate superior estimation accuracy, efficiency, and interpretability of the proposed approach.