Remaining useful life prediction of aero-engine enabled by fusing knowledge and deep learning models
Introduction
As the key component of aircraft with high-reliability requirements, the engine usually develops Prognostics and Health Management (PHM) to increase reliability [1]. One important task in PHM is establishing effective approaches to better estimate the remaining useful life (RUL) [2]. Deep learning achieves success in PHM applications because the non-linear degradation characteristics in the collected sensor data can be revealed. As a result, several deep learning approaches have been proposed for accurately predicting RUL [3], [4], [5].
During the long service life of an aero-engine, fatigue, corrosion, and scale accumulation will cause the performance to degrade slowly. The factors affecting the performance degradation of rotating parts are different, but they are reflected in the change of the engine's efficiency and flow capacity, which are collectively referred to as health parameters [6]. The health parameters are usually difficult to obtain directly, but the sensor data such as temperature, pressure, rotor speed, and fuel flow rate can be easily measured. Usually, sensor data is used as a direct input for different novel deep learning models to predict RUL. For instances, dual long short-term memory (Dual-LSTM) model [7], deep convolution neural networks (DCNNs) model [8], bidirectional gated recurrent neural network (BGRU) model [9], and convolutional neural network (CNN) with bi-directional long short-term memory (BDLSTM) model [10] were constructed for accurate RUL prediction. These studies focused on the optimization and innovation of various deep learning architectures. However, the quality and size of the dataset are also key factors in determining the accuracy bottleneck of the deep learning model.
To address the problem that the accuracy of a deep learning model is limited by its structure and the quality of the dataset, scholars have begun to combine the raw data from sensors and other data. For example, many scholars have proposed hybrid prediction methods that combine data-driven and physically based model approaches [11], [12], [13]. It follows that data-driven solutions can be advantageous for enhancing or replacing inaccurate differential or algebraic equations of physically based models. The above approach requires the establishment of physically based models with higher reliability. For aero-engines, the characteristics of rotating components can shift significantly while the blades corrode and wear. Thus, it is difficult to establish high-precision physical models. This problem limits their practical application for combining data-driven and physically based approaches.
When sufficient failure data is available, some examples of recent deep learning models for prognostics and RUL combine both time-domain and frequency-domain features [14], [15], [16]. The method proposed in [17] formulated RUL prediction as a bi-level optimization problem, in which the lower level was used to forecast the time-series data and the upper level was used to predict the RUL by integrating available up-to-date measurements and predicted values by the lower-level prediction. Where the method in [18] first trained a deep belief network to extract the hidden characteristics corresponding to the fault state of a system, and the distance between the degraded state and failed state was used to construct a health indicator. After, the RUL before failure was predicted. Besides, the LSTM-Fusion architecture proposed in [19] fused multi-sensor data with variable time window sizes to capture the short-range (local) and long-range (global) characteristics of the raw and compressed dataset.
In another variation of machine learning, several approaches acknowledge that selected features of the data and sensors have greater effects on the accuracy of the data-driven model [20]. For example, the combination of inputs such as the operation settings provides an unexpected output [21]. Thus, not all the sensor data indicate the health degradation effectively, and it is important to implement sensor selection to achieve accurate predictions. Characteristics from massive source data were learned to adjust the parameters of the neural networks accordingly [22]. Monotonicity and correlation of the signal data were analyzed to achieve sensor selection [23]. Some approaches eliminated sensors that had constant outputs throughout the lifetime of the engine [24,25] because those sensors could not provide any useful information to facilitate the prediction.
The above-mentioned solution ideas utilize the original raw sensor data multiple times for different purposes, but they are not based on the essence of the information. In addition to sensor data, the information represented by the measured quantities of the sensors is also important knowledge that is worth utilizing. In particular, the mechanism of a complex engineering system can also be used for an accurate machine learning model. Therefore, another approach uses knowledge to guide machine learning, i.e., enhancing the input space.
In contrast to the hybrid models mentioned above, we focused on insights based on the combination of knowledge and deep learning models. The dataset of the run-to-failure degradation trajectories not only included measured data but also contained a large amount of information, such as the relationships between the sensors. Making the data reflect the intrinsic features of the data instead of just inputting it into deep learning models is important. For this reason, this paper used knowledge to explore the ontogenetic relationships between sensors to guide the sensor data arrangement and deep learning structure. By developing the method of knowledge representation, knowledge cluster, and deep learning model construction principles, the problem of knowledge utilization is solved for the problem of RUL estimation. Furthermore, the deep learning model constructed by our approach is interpretable. The knowledge cluster results could be illustrated in the physical significance of the sensors and the principle of model construction could explain the reason for the layers of the model. Moreover, the feasibility of our approach was demonstrated by taking NASA C-MAPSS [26] as the study case. The proposed approach outperforms the state-of-art deep learning model. At the same time, different experiment deep learning models prove that knowledge has a positive effect on RUL prediction accuracy.
In the remainder of the paper, the RUL prediction method that combines sensor knowledge and deep learning is proposed in Section 2. The evaluation of this method using a case study is presented in Section 3. Furthermore, ablation studies are utilized to prove the importance of knowledge in Section 4. Then, the discussion of comparison with other literature methods is adopted in Section 5. Finally, the results of our method and future work are concluded in Section 6.
Section snippets
Methodology for fusing knowledge and deep learning
In this section, the proposed RUL prediction method based on knowledge and deep learning is presented. First, the framework of the proposed method is introduced. Then, the representation of the aero-engine sensor knowledge and knowledge transformation are described. Finally, the deep learning model construction principle is introduced.
Case study
This section is divided into four parts to evaluate our method. First, the dataset used in the case and its challenges are explained. Second, the data preprocessing and labeling process are explicated. Then, the evaluation score is presented, and finally, the validity of the application results of our proposed method is analyzed. Moreover, comparisons with results from state-of-the-art methods for the common dataset are presented to demonstrate the superiority of our method.
Ablation studies
This section utilizes two ablation studies to provide further insights into the knowledge importance. First, the impact of the knowledge cluster is evaluated by a test experiment. Second, the impact of the uncertainties of the knowledge is analyzed.
Discussion
For complex engineered systems, the application of deep learning is limited due to the quality of the input data and model uninterpretability. Our approach proposed a novel direction to break through these limitations by embedding knowledge. Compared with other knowledge embedding methods, the differences and innovations of the proposed approach are:
- •
The ways of using knowledge are different. Usually, the knowledge is used at the data pre-processing stage, such as obtaining additional data
Conclusions
In this paper, a novel RUL prediction approach that fuses knowledge and deep learning models is proposed. The proposed approach utilized the knowledge of sensor relationships to guide the sensor data arrangement and hybrid deep learning model construction. This approach improves the accuracy of remaining useful life prediction in reliability engineering and solves the problem of model interpretability.
- 1)
Fusing knowledge and deep learning models is feasible.
The effectiveness of the proposed
CRediT authorship contribution statement
Yuanfu Li: Conceptualization, Data curation, Formal analysis, Methodology, Software, Validation, Visualization, Writing – original draft, Writing – review & editing. Yao Chen: Conceptualization, Investigation, Writing – review & editing. Zhenchao Hu: Conceptualization, Data curation, Validation. Huisheng Zhang: Conceptualization, Resources, Writing – review & editing, Supervision, Funding acquisition, Project administration.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work was supported by the National Natural Science Foundation of China [grant numbers 51876116 and 51906138], National Fundamental Research Project [grant numbers JCKY2019204B009, JCKY2020208B004, JCKY2020208B036, MKF20200020], and the National Science and Technology Major Project [grant numbers 2017-I-0002-0002].
References (48)
- et al.
A dual-LSTM framework combining change point detection and remaining useful life prediction
Reliab Eng Syst Saf
(2021) - et al.
Remaining useful life estimation in prognostics using deep convolution neural networks
Reliab Eng Syst Saf
(2018) - et al.
Prediction of remaining useful life based on bidirectional gated recurrent unit with temporal self-attention mechanism
Reliab Eng Syst Saf
(2022) - et al.
Estimating model inadequacy in ordinary differential equations with physics-informed neural networks
Comput Struct
(2021) - et al.
Hybrid physics-informed neural networks for main bearing fatigue prognosis with visual grease inspection
Comput Ind
(2021) - et al.
Potential, challenges and future directions for deep learning in prognostics and health management applications
Eng Appl Artif Intell
(2020) - et al.
Multi-bearing remaining useful life collaborative prediction: a deep learning approach
J Manuf Syst
(2017) - et al.
A hierarchical scheme for remaining useful life prediction with long short-term memory networks
Neurocomputing
(2022) - et al.
A deep belief network based health indicator construction and remaining useful life prediction using improved particle filter
Neurocomputing
(2019) - et al.
Prediction of remaining useful life of multi-stage aero-engine based on clustering and LSTM fusion
Reliab Eng Syst Saf
(2021)
Data-driven remaining useful life prediction via multiple sensor signals and deep long short-term memory neural network
ISA Trans
A multimodal and hybrid deep neural network model for remaining useful life estimation
Comput Ind
Remaining useful life estimation of engineered systems using vanilla LSTM neural networks
Neurocomputing
Generalized dilation convolutional neural networks for remaining useful lifetime estimation
Neurocomputing
Back-propagation suppression study based on intake configuration optimization for an air-breathing pulse detonation engine
Aerosp Sci Technol
A sequence-to-sequence approach for remaining useful lifetime estimation using attention-augmented bidirectional LSTM
Intell Syst Appl
LSTM-based multi-layer self-attention method for remaining useful life estimation of mechanical systems
Eng Fail Anal
Alarm-based predictive maintenance scheduling for aircraft engines with imperfect remaining useful life prognostics
Reliab Eng Syst Saf
Physics-guided convolutional neural network (PhyCNN) for data-driven seismic response modeling
Eng Struct
Machine learning for reliability engineering and safety applications: review of current status and future opportunities
Reliab Eng Syst Saf
PHM-oriented integrated fusion prognostics for aircraft engines based on sensor data
IEEE Sens J
Prognostics and health management for maintenance practitioners—review, implementation and tools evaluation
Int J Progn Health Manag
Deep convolutional neural network based regression approach for estimation of remaining useful life
Cited by (27)
A novel data augmentation framework for remaining useful life estimation with dense convolutional regression network
2024, Journal of Manufacturing SystemsA dual attention LSTM lightweight model based on exponential smoothing for remaining useful life prediction
2024, Reliability Engineering and System SafetySensor-aware CapsNet: Towards trustworthy multisensory fusion for remaining useful life prediction
2024, Journal of Manufacturing SystemsOptimizing prior distribution parameters for probabilistic prediction of remaining useful life using deep learning
2024, Reliability Engineering and System SafetyA physics-informed autoencoder for system health state assessment based on energy-oriented system performance
2024, Reliability Engineering and System Safety