Abstract
Remaining Useful Life (RUL) prediction is an essential aspect of Prognostics and Health Management (PHM), facilitating the assessment of mechanical components’ health statuses and their times to failure. Currently, most deep learning-based RUL prediction methods can achieve accurate RUL point estimations. However, due to sample variability and degradation randomness, point estimations may contain uncertainties. To obtain both RUL prediction values and their corresponding uncertainty estimations, this paper proposes a novel hybrid-driven prediction method that effectively combines an Asymmetric Dual-Channel AutoEncoder and the Nonlinear Wiener Process (ADCAE-NWP). To achieve comprehensive feature extraction, two feature extraction channels are parallelly combined in the encoder. Moreover, to reduce the space-time overhead of the model training process, an asymmetric form of the autoencoder is composed by using only the fully connected layer in the decoder. Subsequently, the ADCAE model is trained to construct health indicators in an unsupervised manner. Finally, the RUL Probability Density Functions (PDFs) are calculated using the NWP. RUL predictions containing uncertainty estimations are obtained by calculating expectations over confidence intervals. The proposed model is experimentally validated and compared on two datasets, and the results demonstrate that the proposed scheme achieves better prediction performance than competing approaches.
Similar content being viewed by others
Data Availability
The Train wheel data that support the findings of this study are available from Shenhua Group but restrictions apply to the availability of these data, which were used under licence for the current study, and so are not publicly available. Data are however available from the authors upon reasonable request and with permission of Shenhua Group. The public milling data can be found in https://www.nasa.gov/content/prognostics-center-of-excellence-data-set-repository.
Abbreviations
- ADCAE-NWP:
-
Asymmetric Dual-Channel AutoEncoder and Nonlinear Wiener Process
- BiLSTM:
-
Bidirectional LSTM
- CNN:
-
Convolutional Neural Network
- FHT:
-
First Hitting Time
- FPT:
-
First Passage Time
- HI:
-
Health Indicator
- LSTM:
-
Long Short-Term Memory
- MAPE:
-
Mean Absolute Percent Error
- PDF:
-
Probability Density Function
- RMSE:
-
Root Mean Square Error
- RNN:
-
Recurrent Neural Network
- RUL prediction:
-
Remaining Useful Life prediction
- STFT:
-
Short-Time Fourier Transform
- TWDS:
-
Train Wheel Detection System
References
Lei Y, Li N, Guo L, Li N, Yan T, Lin J (2018) Machinery health prognostics: A systematic review from data acquisition to RUL prediction. Mech Syst Signal Process. 104:799–834
Zeming L, Jianmin G, Hongquan J, Xu G, Zhiyong G, Rongxi W (2018) A similarity-based method for remaining useful life prediction based on operational reliability. Appl Intell. 48(9):2983–2995
Fernandes M, Corchado JM, Marreiros G (2022) Machine learning techniques applied to mechanical fault diagnosis and fault prognosis in the context of real industrial manufacturing use-cases: A systematic literature review. Appl Intell. p 1–35
Mao W, He J, Zuo MJ (2019) Predicting remaining useful life of rolling bearings based on deep feature representation and transfer learning. IEEE Trans Instrum Meas. 69(4):1594–1608
Serradilla O, Zugasti E, Rodriguez J, Zurutuza U (2022) Deep learning models for predictive maintenance: a survey, comparison, challenges and prospects. Appl Intell. p 1–31
Wang L, Cao H, Xu H, Liu H (2022) A gated graph convolutional network with multi-sensor signals for remaining useful life prediction. Knowl Based Syst. 252:109340
Zhu J, Chen N, Peng W (2018) Estimation of bearing remaining useful life based on multiscale convolutional neural network. IEEE Trans Ind Electron. 66(4):3208–3216
Wu JY, Wu M, Chen Z, Li XL, Yan R (2021) Degradation-aware remaining useful life prediction with LSTM autoencoder. IEEE Trans Instrum Meas. 70:1–10
Huang Y, Huang Z, Yu J, Dai X, Li Y (2022) Short-term load forecasting based on IPSO-DBiLSTM network with variational mode decomposition and attention mechanism. Appl Intell. p 1–18
An Q, Tao Z, Xu X, El Mansori M, Chen M (2020) A data-driven model for milling tool remaining useful life prediction with convolutional and stacked LSTM network. Measurement. 154:107461
Yu W, Pi D, Xie L, Luo Y (2021) Multiscale attentional residual neural network framework for remaining useful life prediction of bearings. Measurement 177:109310
Wang B, Lei Y, Li N, Wang W (2020) Multiscale convolutional attention Network for predicting remaining useful life of machinery. IEEE Trans Ind Electron. 68(8):7496–7504
Chen C, Lu N, Jiang B, Xing Y, Zhu ZH (2021) Prediction Interval Estimation of Aeroengine Remaining Useful Life Based on Bidirectional Long Short-Term Memory Network. IEEE Trans Instrum Meas. 70:1–13
Xia M, Zheng X, Imran M, Shoaib M (2020) Data-driven prognosis method using hybrid deep recurrent neural network. Appl Soft Comput. 93:106351
Li X, Zhang W, Ding Q (2019) Deep learning-based remaining useful life estimation of bearings using multi-scale feature extraction. Rel Eng Syst Saf. 182:208–218
Shi ZY, Chehade A (2021) A dual-LSTM framework combining change point detection and remaining useful life prediction. Rel Eng Syst Saf. 205:1–10
Malhotra P, Tv V, Ramakrishnan A, Anand G, Vig L, Agarwal P, Shroff G (2016) Multi-sensor prognostics using an unsupervised health index based on LSTM encoder-decoder. Proc. 1st ACM SIGKDD Work. Mach Learn Progn Heal Manag San Fransisco, CA, USA
Yu W, Kim Y II, Mechefske C (2019) Remaining useful life estimation using a bidirectional recurrent neural network based autoencoder scheme. Mech Syst Signal Process. 129:764–780
Xue B, Xu F, Huang X Xu Z, Zhang X (2022) Improved similarity based prognostics method for turbine engine degradation with degradation consistency test. Appl Intell. p 1–21
Feng T, Li S, Guo L, Gao H, Chen T, Yu Y (2022) A Degradation-Shock Dependent Competing Failure Processes Based Method for Remaining Useful Life Prediction of Drill Bit Considering Time-shifting Sudden Failure Threshold. Rel Eng Syst Saf. p 108951
Cosme LB, D’Angelo MF, Caminhas WM, Yin S, Palhares RM (2018) A novel fault prognostic approach based on particle filters and differential evolution. Appl Intell. 48(4):834–853
Pei H, Hu C, Si X, Zheng J, Zhang Q, Zhang Z, Pang Z (2019) Remaining useful life prediction for nonlinear degraded equipment with bivariate time scales. IEEE Access 7:165166–165180
Zhang JX, Hu CH, He X, Si XS, Liu Y, Zhou DH (2018) A novel lifetime estimation method for two-phase degrading systems. IEEE Trans Reliab. 68(2):689–709
Pitchforth DJ, Rogers TJ, Tygesen UT, Cross EJ (2021) Grey-box models for wave loading prediction. Mech Syst Signal Process. 159:107741
Obando DR, Martinez JJ, Bérenguer C (2021) Deterioration estimation for predicting and controlling RUL of a friction drive system. ISA Trans. 113:97–110
Yan T, Lei Y, Li N, Wang B, Wang W (2021) Degradation modeling and remaining useful life prediction for dependent competing failure processes. Rel Eng Syst Saf. 212:107638
Pang Z, Si X, Hu C, Du D, Pei H (2021) A Bayesian inference for remaining useful life estimation by fusing accelerated degradation data and condition monitoring data. Rel Eng Syst Saf. 208:107341
Yu W, Shao Y, Xu J, Mechefske C (2022) An adaptive and generalized Wiener process model with a recursive filtering algorithm for remaining useful life estimation. Rel Eng Syst Saf. 217:108099
Pei H, Si XS, Hu CH, Zheng JF, Li TM, Zhang JX, Pang ZN (2021) An adaptive prognostics method for fusing CDBN and diffusion process: application to bearing data. Neurocomputing 421:303–315
Li N, Gebraeel N, Lei Y, Bian L, Si X (2019) Remaining useful life prediction of machinery under time-varying operating conditions based on a two-factor state-space model. Rel Eng Syst Saf. 186:88–100
Shi H, Yang J, Si J (2020) Centralized maintenance time prediction algorithm for freight train wheels based on remaining useful life prediction. Math Probl Eng
Duan Y, Li H, He M, Zhao D (2021) A BiGRU autoencoder remaining useful life prediction scheme with attention mechanism and skip connection. IEEE Sens. J. 21(9):10905–10914
Si XS, Wang W, Hu CH, Zhou DH, Pecht MG (2012) Remaining useful life estimation based on a nonlinear diffusion degradation process. IEEE Trans Reliab. 61(1):50–67
Liu J, Li Q, Han Y, Zhang G, Meng X, Yu J, Chen W (2019) PEMFC residual life prediction using sparse autoencoder-based deep neural network. IEEE Trans Transp Electrific. 5(4):1279–1293
Liu H, Liu Z, Jia W, Lin X (2020) Remaining useful life prediction using a novel feature-attention-based end-to-end approach. IEEE Trans Ind Informat. 17(2):1197–1207
Kitaev N, Kaiser L, Levskaya A (2020) Reformer: The efficient transformer. In: Proc the ICLR Conf
Zhou H, Zhang S, Peng J, Zhang S, Li J, Xiong H, Zhang W (2021) Informer: Beyond efficient transformer for long sequence time-series forecasting. Proc. the AAAI Conf. 35(12):11106–11115
Duan Y, Li H, Zhang N (2022) Mechanical health indicator construction and similarity remaining useful life prediction based on natural language processing model. Meas Sci Technol. 33(9):094008
Acknowledgements
This research is supported by the Technology Innovation Project of Shenhua Group (Grant No. SHGF-17-56) and National Key Research and Development Program of China (Grant number 2019YFB2102500).
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflicts of interest
The authors have no relevant financial or non-financial interests to disclose.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix A Key hyperparameter learning search
Appendix A Key hyperparameter learning search
This section explores the effects of Transformer layers, attention heads, Transformer nodes, and LSTM nodes on the RUL prediction results. We use the violin plot to display the evaluation metrics for RUL prediction visually. The violin plot can well reflect the data distribution, where the white dots represent the median value, the thick black bars indicate the interquartile range, the thin black bars represent the 95% confidence interval, and the edge contours indicate the density of the data distribution.
1.1 A.1 Transformer layers
The number of Transformer network layers affects the feature extraction capability and model complexity of the attention channel. The prediction results with different Transformer layers are shown in Fig. 13. When the Transformer layer is 1, the evaluation metrics take a wide range of values, and the performance of prediction results is volatile. When the number of Transformer layers is 3 and 6, the performance of the evaluation metrics is very close and reaches the best. To simplify the complexity of the model and improve the computational efficiency, we finally choose the number of Transformer layers to be 3.
1.2 A.2 Attention heads
Attention heads can be understood as the model extracting critical information about the time series from different perspectives. For a certain number of hidden nodes, an increase in the number of attention heads decreases the number of attention vector dimensions, which is equal to Transformer nodes divided by attention heads. In Fig. 14, it can be observed that the RMSE and score of RUL prediction results are higher when the attention head is 2 or 8. When the attention head is selected as 4, the median performance of RMSE and score is lower. The median accuracy is slightly higher than the other two.
1.3 A.3 Transformer nodes
The number of Transformer nodes affects the attention dimension calculated for each attention head. As seen in Fig. 15, the median performance of the evaluation metrics score and accuracy is better when the Transformer nodes are set as 128. However, compared to the smaller Transformer node value, 64, the prediction performance does not significantly improve. It may mean that increasing the dimension of attention does not bring a significant RUL prediction advantage.
1.4 A.4 LSTM nodes
The LSTM is used as the temporal feature extraction channel in the proposed model, complementing the attention channel to achieve long-dependency information extraction and information supplementation. In Fig. 16, we can observe that the median of the evaluation metrics is very close and reaches the best when the number of LSTM nodes is 50 or 100. When the LSTM nodes are set as 25, the prediction result has a large range of evaluation metrics, and the stability of the prediction model is poor. Considering the issue of computational efficiency, we selected smaller LSTM nodes for experiments and statistics.
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Duan, Y., Liu, Z., Li, H. et al. A hybrid-driven remaining useful life prediction method combining asymmetric dual-channel autoencoder and nonlinear Wiener process. Appl Intell 53, 25490–25510 (2023). https://doi.org/10.1007/s10489-023-04855-3
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10489-023-04855-3