應用混合式卷積神經網路和注意力機制門控循環單元模型之集成學習於剩餘使用壽命預測

Remaining useful life prediction using ensemble learning of hybrid convolutional neural network and attention-based gated recurrent unit model

許心耀(Hsin-Yao Hsu)

指導教授 ： 陳牧言

共同指導教授 ： 范敏玄(Min-Hsuan Fan)

國立臺中科技大學/資訊與流通學院/資訊管理系碩士班/碩士(2020年)

摘要

隨著智能製造(Intelligent Manufacturing)的發展，需要長時間運行過程中，用智能工程維護預測和健康管理(Prognostics Health Management ,PHM)，能確保是否具有可靠性(Reliability)和安全性(Security)。在未預期的發生情況下，對機械的組件而產生連鎖反應，導致失靈、損壞和故障問題。本研究用於渦輪風扇發動機(Turbofan Engine)的感測訊號，以數據驅動(Data-Driven)的方法對退化的特性進行建模，並推斷剩餘使用壽命(Remaining Useful Life, RUL)。由於訊號的雜訊影響，以至於難以預測退化(Degradation)的過程。在本論文中，通過訊號的特徵篩選(Feature Selection)和標準化(Normalization)，用卷積神經網路(Convolutional Neural Network, CNN)提取局部特徵(Local Feature)，結合門控循環單元(Gated Recurrent Unit, GRU)做時間序列(Time Series)的預測，以注意力機制(Attention Mechanism)來強化性能，並用集成學習(Ensemble Learning)來做多模型的優化，來提高模型的泛化能力。本研究用NASA研究中心為C-MPASS資料集做航空發動機(Aero-Engine)的剩餘使用壽命預測。實驗研究將不同的模型進行比較，結果表明提出的方法均優於其他方法。

關鍵字

預測與健康管理 ； 剩餘使用壽命 ； 深度學習網路 ； 集成學習 ； 注意力機制

並列摘要

As the development of Intelligent Manufacturing, it is necessary to use of the intelligent engineering maintenance and prognostics health management (PHM) during long-term operation, which can ensure reliability and security. In case of unanticipated event and chain reactions, leading to malfunction, damage and failure. This study used for sensing signal of turbofan engine in data-driven approach to model the degradation characteristics and inferred remaining useful life (RUL). Owing to the complexity of noise, it is difficult to predict the degradation process. In this paper, through feature selection and normalization, then using the convolutional neural network (CNN) to extract local features, combined with the gated recurrent unit (GRU) for prediction of time series, using attention mechanism to enhance the performance, and using ensemble learning for optimize multiple models to improve the generalization ability of the model. The experimental study, predicting the remaining useful life of aero-engine by NASA research Center’s C-MAPSS data set. The experimental study compares the performance to different models, and the results show that the proposed method outperform the other methods.

並列關鍵字

Prognostic and health manage ； remaining useful life ； deep neural network ； ensemble learning ； attention mechanism

參考文獻

[1]Atamuradov, V., Medjaher, K., Dersin, P., Lamoureux, B., & Zerhouni, N. (2017). Prognostics and health management for maintenance practitioners-Review, implementation and tools evaluation. International Journal of Prognostics and Health Management, 8(060), 1-31.

[2]Ali, J. B., Chebel-Morello, B., Saidi, L., Malinowski, S., & Fnaiech, F. (2015). Accurate bearing remaining useful life prediction based on Weibull distribution and artificial neural network. Mechanical Systems and Signal Processing, 56, 150-172.

[3]Al-Salah, T. H., Zein-Sabatto, S., & Bodruzzaman, M. (2012, March). Decision fusion software system for turbine engine fault diagnostics. In 2012 Proceedings of IEEE Southeastcon (pp. 1-6). IEEE.

[4]Al-Dulaimi, A., Zabihi, S., Asif, A., & Mohammadi, A. (2019). A multimodal and hybrid deep neural network model for remaining useful life estimation. Computers in Industry, 108, 186-196.

[5]A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Desmaison, L. Antiga, A. Lerer. Automatic differentiation in PyTorch. In NIPS Workshop(2017) .