Scribd
Upload
19 views3 pages

Aircraft Engine Remaining Useful Life Prediction U

Uploaded by

hazel shiny
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
19 views3 pages

Aircraft Engine Remaining Useful Life Prediction U

Uploaded by

hazel shiny
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 3

Aircraft Engine Remaining Useful Life Prediction Using Machine Learning

Michael Kimollo, Xudong Liu


University of North Florida
John E. Mathews Jr. Computer Science
UNF Dr., Jacksonville, FL 32224
[email protected], [email protected]

Introduction before failure, offering a detailed understanding of engine


The aviation industry faces a critical challenge: ensuring the lifespan and facilitating proactive maintenance planning.
safety and efficiency of operations while minimizing the risk
of costly and potentially dangerous engine failures. Tradi- Dataset
tional maintenance approaches, often based on fixed sched- The dataset utilized in this study is sourced from NASA
ules or simple usage thresholds, prove inadequate in ad- (NASA 2023) and encompasses multivariate time series data
dressing this challenge. They can lead to unscheduled down- from 100 unique aircraft engine units. Each unit’s opera-
time, unnecessary maintenance actions, and inefficient re- tional behavior is captured through readings from 21 sen-
source allocation, ultimately impacting safety, operational sors and observations of three(3) distinct operational set-
performance, and profitability. tings. These engines, belonging to the same type, initiate
Various studies have been performed to study the estima- their respective time series with varied levels of initial wear
tion of the Remaining Useful Life (RUL) for aircraft en- and manufacturing differences.
gines. Saxena (Saxena et al. 2008) introduced a model for
damage propagation in aircraft gas turbine engine modules, Results and Analysis
providing valuable data for PHM’08. Zheng’s LSTM model Experiment 1: Classification
(Zheng et al. 2017) leveraged sequence information, demon-
strating improved accuracy across PHM datasets. Lim and In the classification experiment, the focus was solely on
Fellows (Lim et al. 2014) extended the Kalman Filter ensem- evaluating the efficiency of the Long Short-Term Mem-
ble with SKF, addressing non-linear degradation patterns ory (LSTM) model in predicting engine failure within one
and broadening its application. A parallel study (Li et al. month.The optimal LSTM configuration involved two lay-
2020) introduced GRU networks for automatic feature ex- ers: the initial layer comprised 100 units, followed by a
traction from time series data, achieving superior RUL pre- second layer with 50 units. A dropout layer with a rate of
diction performance.
This paper delves into the development and evaluation of
robust RUL prediction models for aircraft engines. It em-
ploys the NASA C-MAPSS dataset (NASA 2023), under-
goes data preprocessing step, and applies a piece-wise RUL
function to generate the RUL column to be predicted. By im-
plementing and comparing various machine learning mod-
els, the project identifies the most accurate and generalizable
approach for predicting RUL.

Methodology
In this study, predictive models for aircraft engine RUL
involved two problem formulations, each addressing spe-
cific operational requirements and decision-making con-
texts. The first is a binary classification task, predicting Figure 1: Training vs Validation Loss and Accuracy for
whether an engine will fail within the next 30 days or not, LSTM Classifier
simplifying the prediction into identifying immediate atten-
tion needs. The second treats RUL prediction as a regres- 0.2 was applied between each of these layers. The classi-
sion problem, estimating the remaining operational cycles fier achieved impressive performance metrics: 96% on pre-
Copyright © 2024 by the authors. cision, 0.88 recall, and 0.92 F1-score.
This open access article is published under the Creative Commons Figure 1 shows the learning curves, including the loss and
Attribution-NonCommercial 4.0 International License. accuracy of the LSTM binary classifier during training.
Experiment 2: Regression results. It consisted of two convolutional layers and two max
In this experiment, three different types of models were built pooling layers. This model achieved a training RMSE of
and tested to predict the actual RUL of the engine, and the 11.09 and a test RMSE of 14.02, accompanied by an R-
results were compared. square value of 0.88.
Additionally, other CNN models were explored, includ-
Regression Models In this category, various regression ing CNN-2, which comprised 1 convolutional layer and 1
machine learning models, including linear regression, ran- max pool layer. Furthermore, the CNN-1+LSTM model was
dom forest, k-nearest neighbors, and others, were trained to composed of 1 convolutional layer followed by 2 LSTM lay-
predict the RUL. Different models exhibited varying perfor- ers. The corresponding performances of CNN-2 and CNN-
mances, as shown in Table 1. Random Forest had the best 1+LSTM are detailed in Table 3.
overall performance in this category, with an RMSE of 15.6
on the training set and 46.3 on the test set. However, the
practicality of the random forest model was challenged by Table 3: Performance Comparison of selected CNN Models
the high test set RMSE. Model Train RMSE Test RMSE
CNN-4 11.909235 14.023282
Table 1: Train and Test RMSE for Regression Models
CNN-2 11.957053 16.370277
Model Train RMSE Test RMSE
CNN-1+LSTM 12.477310 14.625450
Decision Tree 0.000000 69.070572
Extra Tree 0.000000 46.190967 The initial observations indicate that the proposed CNN
Forest 15.626659 46.369789 model outperforms most of the reported values from other
research studies, as indicated in Table 4.
XGB 28.174315 48.496991
KNR 40.501531 48.955520
SVM Reg 43.472257 48.873759
LReg 44.660360 48.399484
Ada Reg 47.671437 51.666636

Deep Recurrent Models In this category, various deep


recurrent models, including Simple Recurrent Neural Net-
works (RNN), LSTM, GRU, and Bidirectional LSTM, were
experimented. These models were chosen due to their ef-
fectiveness with time series data, considering the time- Figure 2: Final Predictions on the Test Set
dependent nature of sensor values. Among these models,
GRU and Bidirectional LSTM demonstrated superior per-
formance. Bidirectional LSTM showed superiority in train-
Table 4: Comparison of various CNN Models in literature
ing RMSE, while GRU outperformed in test RMSE, as sum-
marized in the figure below: Model Test RMSE R2 MAE
Proposed 14.02 0.88 8.80
Table 2: Performance Comparison of Deep Recurrent Mod- Zheng et al. (2017) 16.42 NaN NaN
els
Li et al. (2020) NaN 0.78 13.45
Model Train RMSE Test RMSE
Bidirectional 12.292627 15.144974 The proposed CNN model showcased remarkable robust-
GRU 12.880411 14.782531 ness in final predictions, as illustrated in Figure 2. The
model’s predictions exhibited a high degree of accuracy in
RNN 13.034745 14.831060 closely tracking the actual values on the test set, thereby af-
LSTM 13.884804 15.078082 firming its efficacy in predicting the RUL of the engine.

Deep Convolutional Models In this category, various 1D


and 2D CNN underwent training, and their performances
were systematically compared. Among these models, CNN-
4, considered as the best CNN model, exhibited exceptional
References
Li, L.; Zhao, Z.; Zhao, X.; and Lin, K.-Y. 2020. Gated
recurrent unit networks for remaining useful life prediction.
IFAC-PapersOnLine 53(2):10498–10504.
Lim, P. .; Goh, C. K.; Tan, K. C. .; and Dutta, P. . 2014. Es-
timation of remaining useful life based on switching kalman
filter neural network ensemble. Annual Conference of the
PHM Society 6(1).
NASA. 2023. C-mapss jet engine simulated
data. https://data.nasa.gov/Aerospace/
CMAPSS-Jet-Engine-Simulated-Data/
ff5v-kuh6.
Saxena, A.; Goebel, K.; Simon, D.; and Eklund, N. 2008.
Damage propagation modeling for aircraft engine run-to-
failure simulation. In 2008 International Conference on
Prognostics and Health Management, 1–9.
Zheng, S.; Ristovski, K.; Farahat, A.; and Gupta, C. 2017.
Long short-term memory network for remaining useful life
estimation. In 2017 IEEE International Conference on Prog-
nostics and Health Management (ICPHM), 88–95.

You might also like