Neurocomputing 462 (2021) 195–207

Contents lists available at ScienceDirect

Neurocomputing
journal homepage: www.elsevier.com/locate/neucom

A data-driven degradation prognostic strategy for aero-engine under

various operational conditions
Cunsong Wang a,b,c, Zhenghong Zhu b, Ningyun Lu c,d,⇑, Yuehua Cheng c, Bin Jiang c,d
a
Institute of Intelligent Manufacturing, Nanjing Tech University, Nanjing 210009, China
b
Department of Mechanical Engineering, York University, Toronto M3J1P3, Canada
c
College of Automation Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China
d
State Key Laboratory of Mechanics and Control of Mechanical Structures, Nanjing University of Aeronautics and Astronautics, Nanjing, 211106, China

a r t i c l e i n f o a b s t r a c t

Article history: Practical degradation prognostics for aero engines is difficult owing to the degradation features covered
Received 18 January 2021 by the continuous switching among various operational conditions. A novel degradation prognostics
Revised 9 April 2021 strategy for aero engines under various operating conditions is proposed in this study. Specifically, to
Accepted 25 July 2021
remove the influence of different operational conditions, degradation features, which are hidden in
Available online 30 July 2021
Communicated by Zidong Wang
raw data, are extracted. This strategy can realize health state estimation, degradation trend prediction
and remaining useful life (RUL) estimation. First, a k-means algorithm and three defined indicators are
combined to distinguish different operational conditions, extract preferably monotonic and trendable
Keywords:
Degradation prognostics
degradation features. A linear logistic regression model is utilized to construct a synthesized health index
RUL estimation (SHI) library. Second, a deep forest classifier (DFC) and long short-term memory (LSTM) are utilized to
Various operating conditions establish an offline health state estimation model and a degradation trend prediction model. Finally, a
Aero-engine dynamic time warping algorithm (DTW) is adopted to obtain a new SHI for online RUL estimation based
on the two aforementioned offline models. Verification results using the NASA Prognostics Center dataset
show that the proposed strategy for aero engines under various operating conditions is effective and
feasible.
Ó 2021 Elsevier B.V. All rights reserved.

1. Introduction tivariate RUL estimation. The univariate method requires run-to-

failure data and relies on certain prior knowledge to construct a
Aerospace systems such as aircraft engines are extremely com- synthesized health index (SHI) for RUL estimation according to a
plicated and generally expected to operate reliably and safely dur- predefined failure threshold (FT) [13–15]. The direct method needs
ing their service period [1]. However, performance degradation is not set an FT and relies on a certain number of high-quality sam-
inevitable in any engineering systems owing to man–machine- ples to learn one mapping between the system condition-
environment interference and component aging. Specifically, the monitoring (CM) data and the system end of life (EOL) for RUL esti-
continuous and frequent switching of operational conditions accel- mation [16–18]. Compared with the first two methods, the multi-
erates the degradation process and will likely result in faults or variate method emerges recently with great promise and attracts a
failures [2,3]. Degradation prognostics and remaining useful life lot of attention [19–21]. This method does not require SHIs nor an
(RUL) estimation are critical issues in the maintenance manage- FT but relies on a certain number of high-quality run-to-failure
ment of engineering systems [4–6]. samples to train and compose multivariate classifiers as well as
In recent years, with the development of sensor technology, time series for RUL estimation. Although effective in different
computer science and artificial intelligence, data-driven prognos- aspects, none of these methods are perfect.
tics technology [7–10] has gradually become an essential tool for Implicitly, the aforementioned RUL estimation methods are
implementing RUL estimation. Data-driven RUL estimation meth- applicable only to systems under static operation conditions. How-
ods can be classified into the following three categories [11,12]: ever, in real-world tasks, systems are often operated under various
1) univariate RUL estimation; 2) direct RUL estimation and 3) mul- conditions. For instance, an aero-engine often works with varying
loading, usage or rotatory speeds. Studies on the degradation prog-
⇑ Corresponding author at: College of Automation Engineering, Nanjing Univer- nostics of engineering systems with various operational conditions
sity of Aeronautics and Astronautics, Nanjing 211106, China. are limited at present. Moreover, few researchers have focused on
E-mail address: [email protected] (N. Lu). data-driven fault diagnostics or degradation prognostics for

https://doi.org/10.1016/j.neucom.2021.07.080
0925-2312/Ó 2021 Elsevier B.V. All rights reserved.
C. Wang, Z. Zhu, N. Lu et al. Neurocomputing 462 (2021) 195–207

Nomenclature

LASSO Least Absolute Shrinkage and Selection Operator

ACRONYMS LPC Low-Pressure Compressor
BLSTM Bidirectional Long Short-Term Memory Network LPT Low-Pressure Turbine
DBN Deep Belief Network LSTM Long Short-Term Memory
DCNN Deep Convolution Neural Network MODBNE Multi-Objective Deep Belief Networks Ensemble
DFC Deep Forest Classifier RF Random Forest
DTW Dynamic Time Warping RUL Remaining Useful Life
FCM Fuzzy C-means RVR Relevance Vector Regression
HPC High-Pressure Compressor SHI Synthesized Health Index
HPT High-Pressure Turbine SVM Support Vector Machine
MLP Multilayer Perceptron SVR Support Vector Regression

mechanical bearings under varying working conditions [22–24], 3. The proposed strategy solves the delay of time series sequences
where transfer learning technique is used [25–27]. Several and warps a new SHI for a new aero engine to estimate RUL in
improved or hybridized transfer learning techniques, such as real time.
improved joint distribution adaptation (IJDA) [28], deep convolu-
tional transfer learning network (DCTLN) [29], deep transfer learn- This study applies the aero engine data under various opera-
ing network (DTL) based on a sparse autoencoder (SAE) [30] and tional conditions from the NASA Prognostics Center to verify the
fine-grained adversarial network-based domain adaptation proposed strategy. The data contain six different operational condi-
(FANDA) [31], have been applied. These literatures aim to develop tions and three main operational parameters (i.e., altitude, Mach
methods for transferring and applying learned knowledge from number, and sea-level temperature) are used to distinguish opera-
one working condition to another. Such methods cannot be directly tional conditions. The remainder of this paper is organized as fol-
performed on systems that often switch operating conditions lows. Section 2 presents the details of the offline data
online. manipulation and degradation prediction modeling. Optimal linear
Several pioneering works have been conducted on data-driven model parameter matching and new SHI construction based
degradation prognostics for aero engines under various operating dynamic time warping (DTW) are proposed for online RUL estima-
conditions. For instances, Tang et al. [2], Bian et al. [32] and Son tion in Section 3. Section 4 presents the verification results using
et al. [33] respectively adopted some statistical methods, such as the NASA data for aero engines under various operating conditions.
the semi-Markov decision process (SMDP), continuous-time Mar- Finally, conclusions are drawn in Section 5.
kov chain, and Wiener process, to describe the dynamic nature of
operational conditions for RUL estimation. Huang et al. [34] pro-
posed a prognostics framework based on the bidirectional long
short-term memory (BLSTM) networks to handle both sensor data 2. Offline degradation prognostic modeling
and operational condition data for RUL estimation. Li et al. [35] uti-
lized deep convolution neural networks (DCNN) to estimate RUL. 2.1. Key ideas
Zhang et al. [36] employed a multiobjective evolutionary algorithm
with the training deep belief network (DBN) training technique to The proposed strategy under various operational contains two
evolve multiple DBNs for RUL estimation under varying operation stages, that is, offline stage (Fig. 1) and online stage (Fig. 2). It
conditions. However, none of the above works have focused on can realize health state estimation, degradation trend prediction
health state estimation and degradation trend prediction. Ideally, and RUL estimation.
an application-oriented degradation prognostics should contain During the offline degradation modeling stage, first, the k-
three main contents (i.e., health state estimation, degradation means algorithm is utilized to divide the aero-engine training sam-
trend prediction and RUL estimation) [37,38]. ples into multiple operational conditions according to three opera-
In our previous work [6], we successfully applied a data-driven tional parameters. The training samples data is preprocessed using
strategy to areo-engine under static operation conditions. Also, we Z-score normalization, which can eliminate the influence of contin-
utilized the framework of multivariate RUL estimation methods uous operational condition switching. Second, three indicators (i.e.,
and considered the unlabeled, unbalanced condition monitoring entropy indicator, correlation indicator and trend indicator) are
data. In this latter work, we focus on developing a data-driven defined to handle the measurement uncertainty and extract prefer-
degradation prognostic strategy for aero-engine under various ably monotonic and trendable degradation features which are hid-
operational conditions, which can integrate the functions of uni- den in raw data. The linear logistic regression method is adopted to
variate, direct and multivariate RUL estimation methods. The main build the SHI library. Finally, a deep forest classifier (DFC) is used to
contributions of this paper are summarized below: develop the offline health state estimation model, which can han-
dle the uncertainty of different health states. LSTM is adopted to
1. The proposed strategy introduces an information entropy indi- obtain the offline degradation trend prediction model, which can
cator combined with correlation and trend indicators [6] to deal with vanishing and exploding gradient problems.
handle measurement uncertainties and extract preferably During the online RUL estimation stage, the operational condi-
monotonic and trendable degradation features, which are hid- tions of a new aero-engine unit are first distinguished. Then, the
den in raw data. DTW algorithm is applied to select a superior linear model param-
2. The proposed strategy effectively eliminates the influence of eter from the SHI library and warp a new SHI which can solve the
continuous operational condition switching and realizes three delay of time series sequences. Finally, the well-trained health
main prognostic contents under different operational state estimation and degradation trend prediction models are com-
conditions. bined to estimate RUL.
196
C. Wang, Z. Zhu, N. Lu et al. Neurocomputing 462 (2021) 195–207

Fig. 1. Offline stage of the proposed degradation prognostic framework.


2.2. Data formatting and preprocessing xijkðnÞ ¼ xijkðnsÞ xjkðnÞ ;
jkðnÞ

Aero-engine is a complicated system that needs to deploy mul-

where
tiple sensors to monitor its health state. The collected available
X
K ðn Þ X
I
xjkðnÞ ¼ K ð1nÞI xijkðnÞ ;
training sample data and corresponding operational condition ð1Þ
kðnÞ¼1 i¼1
parameters are denoted by X ðI
J
K Þ and hðI
E
K Þ, respec- vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u X
K ðnÞ X
tively. Two three-dimensional data matrices exist, where I is the u I
 2
sjkðnÞ ¼ tK ð1nÞI xijkðnÞ  xjkðnÞ
total number of aero-engine units ði ¼ 1; 2; 3; . . . ; IÞ; J is the total kðnÞ¼1 i¼1
number of sensors ð j ¼ 1; 2; 3; . . . ; J Þ; E is the operational condition
parameters ðe ¼ 1; 2; 3; . . . ; EÞ; K i represents the number of samples where xijkðnÞ is the value of the sensor-measuring variate j of sample
th
in the entire life cycle of the i engine unit and i at cycle k in operating condition n, and coefficients xjkðnÞ and sjkðnÞ
K ¼ ½K 1 ; . . . ; K i ; . . . ; K I . are respectively the mean and variance of sensor-measuring variate
Fig. 3 shows the process of operational condition identification. j for all training samples in operating condition n, respectively.
First, the three-dimensional data hðI
E
K Þ is expanded into two- Therefore, the normalized three-dimensional time series data
dimensional data hðKI
EÞ according to the direction of E [39],  ðI
J
K Þ arranged in the original can be obtained.
X
P
where KI ¼ Ii¼1 K i . Then, similar to [16], the k-means algorithm
is utilized to divide all the training sample data into N operating
conditions according to operational condition parameters 2.3. Degradation feature selection and SHIs library establishment
hðKI
EÞ. The centers of different operating conditions are defined
as on ¼ fon1 ; on2 ; on3 gðn ¼ 1; 2; 3; . . . ; N Þ. Finally, each sensor signal Combined with correlation indicator qij for sensor-measuring
xijkðnÞ at different working conditions is normalized respectively variate j in sample i (as shown in Eq. (2)) and trend indicator T j
as follows: for sensor-measuring variate j (as shown in Eq. (3)), which is pro-
posed in our previous work [6], the information entropy indicator

197
C. Wang, Z. Zhu, N. Lu et al. Neurocomputing 462 (2021) 195–207

value respectively means that the feature is positively or negatively

correlated with time.
PI
eðqij Þ
Tj ¼ i¼1
I
where 8
ð3Þ
< 1ð x > 0 Þ
>
eðxÞ ¼ 0:5ðx ¼ 0Þ
>
:
0ðx < 0Þ

where T j 2 ½0; 1. If the value is equal to 0 or 1, it means that sensor-

measuring variate j has the monotonic upward or downward
trends.
P    
 KI  
k¼1 p xjk log p xjk
Hj ¼ ð4Þ
I
where p represents the probability of 
xijk that occurs in the training
samples.
Next, according to our previous work [6] and engineering expe-
rience, the three constraint conditions of the degradation feature
selection process are set as follows:
 
 
qij  P 0:4; T j  0 or 1; Hj P 0:01 ð5Þ

If the sensor-measuring variate j meets the Eq. (5), it will be

selected. Assume that F sensor-measuring variates are selected.
The new three-dimensional time series data can be written as
_
Fig. 2. Online stage of the proposed degradation prognostic framework. X ðI
F
K Þ.
Finally, a linear logistic regression model is adopted for each
sample i to construct the one-dimensional SHI yik , as shown in
Eq. (6). This model can guarantee that the range of the constructed
Hj for variate j is first defined in Eq. (4). This indicator can describe SHI yik is from 1 to 0.
the information content covered in a variate and uncertainties in
X
F
_
the feature selection process. yijk ¼ ain þ bijðnÞ x ijkðnÞ ð6Þ
j¼1
PK i 2    
6 k¼1 dxijk
qij ¼ 1    ð2Þ where aiðnÞ ; bijðnÞ ¼ aiðnÞ ; bi1ðnÞ ; bi2ðnÞ ; bi3ðnÞ ; . . . ; biF ðnÞ are the F þ 1
K i K 2i  1 model parameters of operational condition n for sample

i; aiðnÞ 2 aið1Þ ; aið2Þ ; aið3Þ ; . . . ; aiðNÞ and
n o
where dxijk is the difference between sensor-measuring variate xijk bijðnÞ 2 bijð1Þ ; bijð2Þ ; bijð3Þ ; . . . :; bijðNÞ . The parameter set of each sample
and its cycle; qij 2 ½1; 1 and the positive or negative correlation is saved in the SHI library, which is composed of I different

Fig. 3. The operational condition identification process.

198
C. Wang, Z. Zhu, N. Lu et al. Neurocomputing 462 (2021) 195–207

 
aiðnÞ ; bijðnÞ . Therefore, the new three-dimensional time series data 9: If all the classification accuracies exceed x, output the
can be written as Y ðI
1
K Þ. optimal forest ensemble structure; Otherwise return to
Step 3.
2.4. Health state estimation modeling

 ðI
J
K Þ in the SHI library is

Similar to our previous work [6], X
used to establish the offline health state estimation model. In this
modeling, an FCM (Fuzzy C-means) algorithm is used for the 2.5. Degradation trend prediction modeling
 P 
unfolded two-dimensional data Y ð Z
1Þ Z ¼ KI ¼ Ii¼1 K i to In this strategy, multistep degradation prediction can be real-
assign health state labels Tag ðzÞ. The health state label assignment ized through a well-trained univariate sequence-to-sequence
process can be considered as minimizing objective function L in Eq. regression LSTM predictor. The LSTM network is well-suited for
(7). Subsequently, this optimization problem is solved by itera- time series prediction, as this structure can deal with long time
tively computing Eqs. (8) and (9). lags of unknown duration between important events. In addition,
Min fLg an LSTM network is developed to solve vanishing and exploding
fmc ;lc g gradient problems which are often encountered in a recurrent neu-
X
4 X ral network (RNN). The offline degradation trend prediction mod-
Z
 
L¼ lc ðyz Þ 2 kyz  mc k2 ð7Þ
eling is presented briefly in Algorithm 2.
c¼1 z¼1
X
4
Algorithm 2 Offline Degradation Trend Prediction Modeling
lc ðyz Þ ¼ 1; z ¼ 1; 2; . . . ; Z
c¼1 Input: Y ðI
1
K Þ
Output: A well-trained LSTM network model
PZ  
lc ðyz Þ 2 yz Process:
mc ¼ Pz¼1  2 ð8Þ 1: Set several main parameters of LSTM network;
z¼1 lc ðyz Þ
Z
2: for i ¼ 1; 2; 3; . . . ; I do
3: Set the sequence from yi ð1Þ to yi ðK i  1Þ as the input of the
1
lc ðyz Þ ¼ P  2 ð9Þ model and the sequence from yi ð2Þ to yi ðK i Þ as the input of
C kyz mc k the model;
k¼1 kyz mk k
4: end for
where c ¼ 1; 2; . . . ; C represents the health state (i.e., normal, mild, 5: Train the network and calculate its six intermediate
moderate and near-to-failure); lc ðyz Þ is the membership degree of parameters [43] during the forward propagation;
yz in health state c; and mc represents the center of health state c. 6: Adopt L2 regularization and Adam methods [44] to
Fig. 4 shows that the DFC [40] is applied to train the offline prevent overfitting and promote the convergence during
health state estimation model and save the optimal forest struc- the resilient back-propagation;
ture based on unfolded two-dimensional data Y ð Z
1Þ and corre- 7: Save the structure until the LSTM network implements
sponding health state labels Tag ðzÞ. The DFC is actually a convergence.
multiple complete random forests (represented by full lines) [41]
and random forests (represented by imaginary lines) [42] ensem-
ble classification algorithm. This classifier has fewer hyperparame-
ters and satisfactory robustness and level automatic extension
than deep neural network algorithms. Algorithm 1 gives the proce- 3. Online RUL estimation
dures of the offline health estimation modeling.
3.1. New aero-engine data
Algorithm 1 Offline Health State Estimation Modeling
Input: Y ð Z
1Þ with health state labels Tag ðzÞ The new aero engine (testing data) should be preprocessed
Output: An optimal forest ensemble model before online RUL estimation is implemented. The new aero engine
Process: data and corresponding operational condition parameters are
   
1: Set several main parameters of DFC; assumed as X ^ ^I
F
K
^ and ^
h ^I
E
K
^ , respectively, where K
^^
i
2: Define the classification accuracy x of offline modeling
is incomplete compared with the complete run-to-failure cycle
process;
K i . First, Euclidean distance is applied to the operational condition
3: Select 10 samples randomly to verify the training process ^ to match an
and other samples for training; parameters ^ h^ ^ ðe ¼ 1; 2; 3; . . . ; EÞ of sample ^i at cycle k
iek

4: Train the multiple forest models in the first level, compute optimal one from operational condition centers
out-of-bag estimation and transfer the generated feature on ðn ¼ 1; 2; 3; . . . ; N Þ, as shown in Eq. (10):
n        o
vectors to the second level;
min d ^h^iek^ ; o1 ; d ^h^iek^ ; o2 ; . . . ; d ^h^iek^ ; on ; . . . ; d ^h^iek^ ; oN
5: Extend and train the new level composed of multiple n

forest models;   rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

 2  2  2
^
where; d h^ ^ ; on ¼ ^h^ ^  on1 þ ^h^ ^  on2 þ ^h^ ^  on3
6: Calculate the performance precision of this new level; iek iek iek iek

7: If the performance precision of the extended level is not ð10Þ

improved, continue the next Step; Otherwise return to Step  
5 and extend a new level automatically; h^iek^ ; on represents the Euclidean distance between ^
where d ^ h^iek^ and
8: Use the acquired DFC structure for validation samples to
on . Therefore, sensor-measuring data ^
x^ijk^ðnÞ can be categorized as
calculate classification accuracies;
operational condition n according to the above calculation denoted
by ^x^if k^ðnÞ .
199
C. Wang, Z. Zhu, N. Lu et al. Neurocomputing 462 (2021) 195–207

Fig. 4. Offline health state estimation model training framework.

Then, the DTW [45] algorithm is utilized to measure the simi-

Algorithm 3 Online Data Manipulation larity between the one-dimensional SHI y ^^ii and yi . The optimal
   
^ ^I
J
K
Input: X ^ and ^
h ^I
E
K
^ SHI serial number i of sample ^i is selected, as Eq. (11). According
_ _ _
^^ii . In prac-
to [16], the selected SHIs must be longer than sequence y
Output: The new online sequence Y I
I
K tice, this condition may not be reasonable owing to the delay of
time series sequences. However, the DTW algorithm can solve
Process:
the problem effectively, as it can warp two sequences and align
1: for ^i ¼ 1; 2; 3; . . . ; ^I do
them in time.
2: for k^ ¼ 1; 2; 3; . . . ; K ^ i do
            
3: Use Eq. (10) to match an operational condition for ^^i1 ; u^i1 ðy1 Þ ; d /^i2 y
min d /^i1 y ^^i2 ; u^i2 ðy2 Þ ; . . . ; d /^iI y
^^iI ; u^iI ðyI Þ
i
measuring variate ^ x^if k^ according to the operational
ð11Þ
parameters;
4: end for where i represents the selected optimal serial number for
     
5: Select the optimal operational condition n for ^
x^if k^ and y ^^ii ; u^ii ðyi Þ is the Euclidean distance between /^ii y
^^ii ; d /^ii y ^^ii and
denote it as ^
x^if k^ðnÞ ; u^ii ðyi Þ; /^ii and u^ii denote the transformed functions of the warping
paths using DTW.
6: end for  
Finally, /^ii y ^^ii and u^ii ðyi Þ are compared to select the most sim-
7: for ^i ¼ 1; 2; 3; . . . ; ^I do
ilar sequence of u^ii ðyi Þ for the online RUL estimation. Assume that
8: ^ ¼ 1; 2; 3; . . . ; K
for k ^ i do _ _ _ _ _
9: for i ¼ 1; 2; 3; . . . ; I do Ki 2 K1; K2; K3; . . . ; KI is the cycle of the highly similar sequence.
 
10: Use aiðnÞ ; bijðnÞ to fuse sensor-measuring variate The online sequence for the RUL estimation can be denoted as
_ _
^
x^if k^ðnÞ ; 11: end for three-dimensional data Y ^I
1
K . The detailed steps of the
12: end for
online data manipulation are summarized in Algorithm 3.
13: end for
14: Get I different one-dimensional SHIs for each sample and
  3.2. RUL estimation
denote it as Y^ ^I
I
K
^ ;

15: for ^i ¼ 1; 2; 3; . . . ; ^I do _
Similar to [6], sequence y^i of sample ^i is utilized to predict its
16: for i ¼ 1; 2; 3; . . . ; I do degradation trend using the established offline degradation trend
17: Use DTW to warp the two sequences y ^^ii and yi ; prediction model. Furthermore, the predicted values are used to
18: end for evaluate health states according to the offline health state estima-
19: Select the optimal SHI serial number i according to Eq. tion model. The FT of the online sample data is set to 0. Once the
(11); 20: end for predicted value reaches the FT, the final RUL can be estimated.
21: for ^i ¼ 1; 2; 3; . . . ; ^I do The evaluation criterion of this strategy can be defined as Eqs.
 
22: Compared /^ii y ^^ii and u^ii ðyi Þ to select the highly similar (12) and (13) [46]:
sequence; 8 ^  
>
> X I
 ai
r^

23: end for >

> e 1
 1; r^i < 0
>
<^
24: Generate the online sequence for RUL estimation i¼1
_ _ score ¼   ð12Þ
Y ^I
1
K . >
> X^I r^
>
>
i
>
: e
a2
 1; r^i P 0
^i¼1

vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
Second, the sensor-measuring variates of each sample are fused u X b
u1 I 2
into I different one-dimensional SHIs according to the operating
  RMSE ¼ t ^r ð13Þ
bI ^ ^i
condition parameters aiðnÞ ; bijðnÞ of the above established library. i¼1

The fused one-dimensional SHIs of all the samples can be consid- where score represent the score of this strategy; RMSE is the root
 
^ ^I
I
K
ered as three-dimensional data Y ^ . mean square error of the RUL estimation process;
200
C. Wang, Z. Zhu, N. Lu et al. Neurocomputing 462 (2021) 195–207

r^i ¼ RUL^i e  RUL^i a is the difference value of sample ^i between the

estimated RUL value RUL^i e and its corresponding actual RUL value
RUL^i a. Parameters a1 and a2 are set to 10 and 13 [46], respectively,
in this study. The lower score and RMSE of the above equations
show that the degradation prognostics strategy is superior.

4. Verification

4.1. Data description

The available aero-engine dataset (i.e., ‘‘FD002” [47] acquired

from the NASA Prognostics Center is applied to verify the proposed
data-driven prognostic strategy under various operating condi-
tions. The dataset is generated by simulating the gradual degrada-
tion of the aero-engine high-pressure compressor (HPC) under six
Fig. 5. Aero-engine sensor layout [6].
operating conditions ðN ¼ 6Þ. The available dataset mainly includes
260 run-to-failure training X ð260
21
K Þ and 259 incomplete is used to normalize each sensor-measuring variate xijkðnÞ according
 
^ 259
21
K
testing X ^ as well as the corresponding real RUL val- to its operational condition. Compared with raw data, a normalized
sensor-measuring variate can present the incipient degradation
ues RUL^i a of the testing samples. Each training or testing sample
trend, as shown in Fig. 8. Normalization based on different opera-
contains three operational condition parameters (hð260
3
K Þ
  tional conditions can stimulate the following degradation feature
or ^
h 259
3
K ^ ). Furthermore, 21 sensorimeasuring variates
selection.
are described in Table 1.
The general layout of the sensors is presented in Fig. 5. Sensor-
4.3. Results of offline degradation feature selection and SHIs library
measuring variates 1 to 9 rely on hardware measurement, whereas
construction
variates 10 to 21 depend on soft measurement. The simulated
aero-engine model contains a combustor, a nozzle and five rotating
Similar to our previous work [6], the 20-step moving average
components. The five components are a fan, a low-pressure turbine
technique is adopted to first smooth the normalized sensor-
(LPT), a low-pressure compressor (LPC), an HPC and a high-
measuring data. Next, correlation indicator qij , trend indicator T j
pressure turbine (HPT).
and information entropy indicator Hj are calculated using Eqs.
(2)–(4). A sensor-measuring variate that meets all the conditions
4.2. Results of offline data preprocessing in Eq. (5) is considered as a useful degradation feature for the next
offline degradation prognostic modeling.
Given the influence of multiple operating condition switching, Fig. 9 gives four main cases for the correlation indicator of each
the degradation trend of a sensor-measuring variate is often cov- measuring sensor in the training samples. For example, several cor-
ered (as shown in Fig. 6). Generally, applying prognostics methods relation indicators of one sensor-measuring variate are less than
to the raw data of sensors is difficult. 0.4 as shown in Fig. 9(a), which demonstrates that the degradation
First, the three operational condition parameters hð260
3
K Þ trend of this measuring sensor is not obvious; All the correlation
of all the training samples are expanded into two-dimensional data indicators of one sensor-measuring variate remain over 0.4, as
hð53759
3Þ. Next, the k-means algorithm is utilized to distin- shown in Fig. 9(b), which represents a superior degradation fea-
guish six different operating conditions. Fig. 7 shows the clustering ture; All the correlation indicators of one measuring-sensor variate
results of the operating condition parameters and the specific val- are not still greater or less than 0, as shown in Fig. 9(c), which
ues of the clustering centroids are given in Table 2. Finally, Eq. (1) shows that the sensor is not a monotonic feature; All the correla-
Table 1 tion indicators of sensor-measuring variate are still equal to 0.5,
The sensor-measuring variate description [46]. as shown in Fig. 9(d), which indicates that the sensor-measuring
value remains constant and is an invalid feature.
No. Symbol Description Unit
Therefore, sensor-measuring variates 2, 3, 4, 11, 15 and 17 are
1 T2 Total temperature at fan inlet °R
chose initially according to the first two constraints of Eq. (5). Only
2 T24 Total temperature at LPC outlet °R
3 T30 Total temperature at HPC outlet °R the information entropy indicator of sensor-measuring variate 17
4 T50 Total temperature at LPT outlet °R fails to meet the third constraint of Eq. (5). In conclusion, sensor-
5 P2 Pressure at fan inlet psia measuring variates 2, 3, 4, 11 and 15 are selected as degradation
6 P15 Total pressure in bypass-duct psia features. The new three-dimensional time series data is written
7 P30 Total pressure at HPC outlet psia _
8 Nf Physical fan speed rpm as X ð260
5
K Þ.
9 Nc Physical core speed rpm
 
10 epr Engine pressure ratio (P50/P2) –
Finally, an SHI library which contains 260 different aiðnÞ ; bijðnÞ
 
11 Ps30 Static pressure at HPC outlet psia
is established by Eq. (6), where aiðnÞ ; bijðnÞ contains six model
12 phi Ratio of fuel flow to Ps30 pps/psi
13 NRf Corrected fan speed rpm parameters. The three-dimensional time series data
14 NRc Corrected core speed rpm _
15 BPR Bypass Ratio – X ð260
5
K Þ is transferred into Y ð260
1
K Þ.
16 farB Burner fuel–air ratio –
17 htBleed Bleed Enthalpy –
18 Nf_dmd Demanded fan speed rpm 4.4. Results of degradation prognostic modeling
19 PCNfR_dmd Demanded corrected fan speed rpm
20 W31 HPT coolant bleed lbm/s Similar to our previous work [6], data Y ð260
1
K Þ is
21 W32 LPT coolant bleed lbm/s
unfolded into two-dimensional data Y ð53759
1Þ and the FCM
201
C. Wang, Z. Zhu, N. Lu et al. Neurocomputing 462 (2021) 195–207

Fig. 6. Raw sensor data of five random training samples. (a) Sensor-measuring variate 2. (b) Sensor-measuring variate 9.

algorithm is utilize to assign four different health state labels (i.e., 4.5. Results of online RUL estimation
normal, mild, moderate and near-to-failure) using Eqs. (7)–(9).
 
Next, the main parameters (i.e., the number of different forests ^ 259
5
K
For the testing sample data X ^ , its operational con-
in each individual level, the number of decision trees in each indi-  
vidual forest, the split fraction for cascade training, accuracy toler- dition parameters ^ h 259
3
K^ are first utilized to match an
ance for the cascade growth and training accuracy) are set to 2, optimal operational condition n for each sample at any cycle using
101, 0.2, 0 and 90%, respectively. The DFC is trained to build the off- Eq. (10). Furthermore, the sensor-measuring variates of each test-
line health state estimation model according to Algorithm 1. Fig. 10 ing sample are fused into 260 different one-dimensional SHI
shows the training process of the modeling. In this figure, ‘‘1”, ‘‘2”,  
^ 259
260
K
Y ^ based on operating condition parameters
‘‘3” and ‘‘4” represent the four different health states. The offline  
health state estimation model demonstrates high classification aiðnÞ ; bijðnÞ using Eq. (6).
accuracy. Second, the DTW algorithm is adopted to measure the similarity
Finally, similar to our previous work [43], the main parameters ^^ii and yi . The optimal SHI serial
between the one-dimensional SHI y
(i.e., hidden units, factor for dropping the learning rate, and factor
^
number i of sample i is calculated further by Eq. (11). For example,
for L2 regularization) of the LSTM network are initialized as 200,
0.3, and 0.1, respectively. A multistep univariate predictor is the optimal SHI serial numbers of testing samples # 9 and # 254
are 9 and 252, respectively. For testing samples # 9 and # 254,
trained according to Algorithm 2 for the offline degradation trend  
modeling. sequences (/^ii y ^^ii and u^ii ðyi Þ), which are aligned in time, are
shown in Fig. 11.
 
^^ii and u^ii ðyi Þ are compared to select the highly sim-
Third, /^ii y
ilar sequence (as shown in Fig. 12) for the following online RUL
estimation.
Finally, the defined FT, offline health state estimation and
degradation trend prediction models are combined to realize three
main prognostic contents of each new smoothed online SHI. As
shown in Fig. 13(a), the predicted starting point of testing sample
# 9 is the 161th cycle, which is at the near-to-failure degradation
state. When the predicted degradation trend arrives at the 171th
cycle, the testing sample will be at the failure state. As shown in
Fig. 13(b), the predicted begining point of testing sample # 254
is the 131st cycle, which is at the moderate degradation state.

Table 2
Centroids of different operational conditions.

Condition Altitude (ft) March number Temperature (°F)

1 10.003 0.2505 100
2 25.003 0.6205 60
3 42.003 0.8405 100
4 0.0015 0.0005 100
5 20.003 0.7005 100
Fig. 7. Clustering centroids of different operating conditions. 6 35.003 0.8405 100

202
C. Wang, Z. Zhu, N. Lu et al. Neurocomputing 462 (2021) 195–207

Fig. 8. Normalized data of five random training samples. (a) Sensor-measuring variate 2. (b) Sensor-measuring variate 9.

Fig. 9. Correlation indicator of all training samples. (a) Sensor-measuring variate 1. (b) Sensor-measuring variate 6. (c) Sensor-measuring variate 11. (d) Sensor-measuring
variate 18.

When the predicted degradation trend arrives at the 149th and the several popular prognostic methods according to Eqs. (12) and
169th cycles, the testing sample will be at the near-to-failure (13). The comparisons illustrate that the RMSE and score of the
degradation and failure states. proposed degradation prognostic strategy are lower than those of
The estimated RUL of all the testing samples is calculated and other methods. The lower RMSE and score indicate that the pro-
showed in Fig. 14. Table 3 shows the performance comparison of posed strategy is better than several other prediction strategies.

203
C. Wang, Z. Zhu, N. Lu et al. Neurocomputing 462 (2021) 195–207

Fig. 10. Offline health state estimation model training. (a) Training sample # 251. (b) Training sample # 252.

Fig. 11. Warping the testing sample SHI and its similar SHI. (a) Testing sample # 9. (b) Testing sample # 254.

Fig. 12. New SHI for online RUL estimation. (a) Testing sample # 9. (b) Testing sample # 254.

204
C. Wang, Z. Zhu, N. Lu et al. Neurocomputing 462 (2021) 195–207

Fig. 13. Online degradation prognostic. (a) Testing sample # 9. (b) Testing sample # 254.

Fig. 14. Online RUL estimation of testing samples(a) Estimated RUL results. (b) Absolute error for RUL estimation.

Table 3
Performance comparisons with other methods the proposed strategy can build new online SHIs and realize three
Method RMSE Score main prognostic contents online. The verification results based on
the NASA aero-engine data under various operating conditions can
Strategy in this paper 20.30 3:2963
103
reveal the feasibility and effectiveness of the proposed strategy.
BLSTM [34] 25.11 4:7930
103
The RMSE and score comparisons of several popular methods indi-
DCNN [35] 22.36 1:0412
104
RF [36] 30.01
cate that the proposed strategy is better than the others and
8:4068
104
enables an aero-engine to realize predictive maintenance. The pro-
DBN [36] 27.12 9:0316
103
posed degradation prognostic strategy can be suitable for similar
LASSO [36] 37.13 2:7692
105
engineering system under the condition of system hardware.
MLP [36] 28.78 4:9815
105
SVM [36] 52.99 3:1648
105
This study assumed that only one aero engine degradation com-
SVR [36] 42.00 5:8990
105
ponent under several known conditions was present. However, in
RVR [36] 31.30 1:7423
104 practice, multiple aero engine degradation components under
MODBNE [36] 25.05 5:5853
103 unknown conditions may exist. Thus, degradation generated by
multiple components under unknowns conditions should be con-
sidered in future research.

5. Conclusions
CRediT authorship contribution statement
In this study, a degradation prognostic strategy combined with
a DFC, LSTM and DTW was proposed and applied to an aero-engine
Cunsong Wang: Methodology, Software, Validation, Writing -
system under various operating conditions. This strategy can dis-
original draft. Zhenghong Zhu: Writing - review & editing. Nin-
tinguish different operating conditions successfully, extract degra-
gyun Lu: Writing - review & editing, Supervision, Funding acquisi-
dation features efficiently, and construct an SHI library. Moreover,
tion. Yuehua Cheng: Writing - review & editing. Bin Jiang: Writing
205
C. Wang, Z. Zhu, N. Lu et al. Neurocomputing 462 (2021) 195–207

- review & editing, Project administration. Cuimei Bo: Funding [18] X. Li, Q. Ding, J. Sun, Remaining useful life estimation in prognostics using deep
convolution neural networks, Rel. Eng. Syst. Safe. 172 (2018) 1–11.
acquisition, Project administration.
[19] E. Ramasso, M. Rombaut, N. Zerhouni, Joint prediction of continuous and
discrete states in time-series based on belief functions, IEEE Trans. Cybern. 43
(1) (2013) 37–50.
Declaration of Competing Interest [20] E. Rammasso, R. Gourtveau, Remaining useful life estimation by classification
of predictions based on a neuro-fuzzy system and theory of belief functions,
The authors declare that they have no known competing finan- IEEE Trans. Rel. 63 (2) (2015) 555–566.
[21] K. Javed, R. Gouriveau, N. Zerhouni, A new multivariate approach for
cial interests or personal relationships that could have appeared prognostics based on extreme learning machine and fuzzy clustering, IEEE
to influence the work reported in this paper. Trans. Cybern. 45 (12) (2015) 2626–2639.
[22] B. Yang, Y. Lei, F. Jia, N. Li, Z. Du, A polynomial kernel induced distance metric
to improve deep transfer learning for fault diagnosis of machines, IEEE Trans.
Acknowledgments Ind. Electron. (2019), https://doi.org/10.1109/TIE.2019.2953010.
[23] Y. Song, Y. Li, L. Jia, M. Qiu, Retraining strategy based domain adaption network
for intelligent fault diagnosis, IEEE Trans. Ind. Inform. (2019), https://doi.org/
This work was supported in part by the National Key Research 10.1109/TII.2019.2950667.
and Development Program of China under Grant [24] J. Jiao, M. Zhao, J. Lin, Unsupervised adversarial adaptation network for
2019YFB1705800, the National Natural Science Foundation of intelligent fault diagnosis, IEEE Trans. Ind. Electron. (2019), https://doi.org/
10.1109/TIE.2019.2956366.
China under Grant 61873122, the Research Fund of State Key Lab- [25] W. Lu, B. Liang, Y. Cheng, D. Meng, J. Yang, T. Zhang, Deep model based domain
oratory of Mechanics and Control of Mechanical Structures (Nan- adaptation for fault diagnosis, IEEE Trans. Ind. Electron. 64 (3) (2016) 2296–
jing University of Aeronautics and Astronautics) under Grant 2305.
[26] Z. Zhang, H. Chen, S. Li, Z. An, J. Wang, A novel geodesic flow kernel based
MCMS-I-0121G03 and Six Talent Climax Foundation of Jiangsu
domain adaptation approach for intelligent fault diagnosis under varying
under Grant TD-JNHB-009. working condition, Neurocomputing 376 (2020) 54–64.
[27] L. Wen, L. Gao, X. Li, A new deep transfer learning based on sparse auto-
encoder for fault diagnosis, IEEE Trans. Syst. Man Cyb. 49 (1) (2017) 136–144.
References [28] W. Qian, S. Li, P. Yi, K. Zhang, A novel transfer learning method for robust fault
diagnosis of rotating machines under variable working conditions,
[1] H. Hanachi, C. Mechefske, J. Liu, A. Banerjee, Y. Chen, Performance-based gas Measurement 138 (2019) 514–525.
turbine health monitoring, diagnostics, and prognostics: A survey, IEEE Trans. [29] L. Guo, Y. Lei, S. Xing, T. Yan, N. Li, Deep convolutional transfer learning
Rel. 67 (3) (2018) 1340–1363. network: A new method for intelligent fault diagnosis of machines with
[2] D. Tang, J. Cao, J. Yu, Remaining useful life prediction for engineering systems unlabeled data, IEEE Trans. Ind. Electron. 66 (9) (2019) 7316–7325.
under dynamic operational conditions: A semi-Markov decision process-based [30] C. Sun, M. Ma, Z. Zhao, S. Tian, R. Yan, X. Chen, Deep Transfer learning based on
approach, Chinese J. Aeronuat. 32 (3) (2019) 627–638. sparse autoencoder for remaining useful life prediction of tool in
[3] Y. Li, N. Lu, X. Wang, B. Jiang, Islanding fault detection based on data-driven manufacturing, IEEE Trans. Ind. Electron. 15 (4) (2019) 2416–2425.
approach with active developed reactive power variation, Nerocomputing 337 [31] Z. Chai, C. Zhao, A Fine-Grained Adversarial Network Method for Cross-Domain
(2019) 97–109. Industrial Fault Diagnosis, IEEE Trans. on Autom. Sci. and Eng. 17 (3) (2020)
[4] M. Mahdavi, H. Monsef, R. Romero, Reliability effects of maintenance on TNEP 1432–1442.
considering preventive and corrective repairs, IEEE Trans. Power Syst. 32 (5) [32] L. Bian, N. Gebraeel, J. Kharoufeh, Degradation modeling for real-time
(2017) 3768–3781. estimation of residual lifetimes in dynamic environments, Lie Trans. 47 (5)
[5] Z. Zhu, Y. Xiang, M. Li, W. Zhu, K. Schneider, Preventive maintenance subject to (2015) 471–486.
equipment unavailability, IEEE Trans. Rel. 68 (3) (2019) 1009–1020. [33] K. Le Son, M. Fouladirad, A. Barros, E. Levrat, B. Iung, Remaining useful life
[6] C. Wang, N. Lu, Y. Cheng, B. Jiang, A data-driven aero-engine degradation estimation based on stochastic deterioration models: A comparative study,
prognostic strategy, IEEE Trans. Cybern. (2019), https://doi.org/10.1109/ Realib. Eng. Syst. Safe. 112 (2013) 165–175.
TCYB.2019.2938244. [34] C. Huang, H. Huang, Y. Li, Bidirectional LSTM prognostics method under
[7] K. Liu, N. Gebraeel, J. Shi, A data-level fusion model for developing composite multiple operational conditions, IEEE Trans. Ind. Electron. 66 (11) (2019)
health indices for degradation modeling and prognostic analysis, IEEE Trans. 8792–8802.
Autom. Sci. Eng. 10 (3) (2013) 652–664. [35] X. Li, Q. Ding, J. Sun, Remaining useful life estimation in prognostics using deep
[8] H. Li, D. Pan, C.L.P. Chen, Reliability modeling and life estimation using an convolution neural networks, Rel. Eng. Sys. Safe. 172 (2018) 1–11.
expectation maximization based wiener degradation model for momentum [36] C. Zhang, P. Lim, A. Qin, K. Tan, Multiobjective deep belief networks ensemble
wheels, IEEE Trans. Cybern. 45 (5) (2015) 969–977. for remaining useful life estimation in prognostics, IEEE Trans. Ind. Electron. 28
[9] V. Atamuradov, K. Medjaher, F. Camci, P. Dersin, N. Zerhouni, Railway point (10) (2017) 2306–2318.
machine prognostics based on feature fusion and health state assessment, IEEE [37] Y. Lei, Intelligent Fault Diagnosis and Remaining Useful Life Prediction of
Trans. Instrum. Meas. 64 (3) (2017) 2276–2285. Rotating Machinery, Butterworth-Heinemann, Oxford, U.K., 2016.
[10] M. Xia, T. Li, T. Shu, J. Wan, C.W. de Silva, Z. Wang, A two-stage approach for [38] V. Atamuradov, K. Medjaher, P. Dersin, B. Lamoureux, N. Zerhouni, Prognostics
the remaining useful life prediction of bearings using deep neural networks, and health management for maintenance practitioners-review,
IEEE Trans. Ind. Inform. 15 (6) (2019) 3703–3711. implementation and tools evaluation, Int. J. Pro. Heal. Manage. 8 (60) (2018)
[11] K. Javed, R. Gouriveau, N. Zerhouni, State of the art and taxonomy of 1–31.
prognostics approaches, trends of prognostics applications and open issues [39] D. Louwerse, A. Smilde, Multivariate statistical process control of batch
towards maturity at different technology readiness levels, Mech. Syst. Signal processes based on three-way models, Chem. Eng. Sci. 55 (7) (2000) 1225–
Process. 94 (2017) 214–236. 1235.
[12] M. Kordestani, M. Saif, M. Orchard, R. Razavi-Fai, K. Khorasani, Failure [40] Z. Zhou, J. Feng, Deep forest, Nat. Sci. Rev. 6 (1) (2019) 74–86.
prognosis and applications: A survey of recent literature, IEEE Trans. Rel. [41] W. Fan, H. Wang, P. Yu, S. Ma, Is random model better? on its accuracy and
(2019), https://doi.org/10.1109/TR.2019.2930195. efficiency, in: Proc. 3th IEEE Int. Conf. Data Min. Melbourne, FL, USA. 2003, pp.
[13] P. Wang, B. Youn, C. Hu, A generic probabilistic framework for structural health 51–58..
prognostics and uncertainty management, Mech. Syst. Signal Process. 28 (2) [42] L. Breiman, Random forest, Mach. Learn. 45 (2001) 5–32.
(2012) 622–637. [43] C. Wang, N. Lu, S. Wang, Y. Cheng, B. Jiang, Dynamic long short-term memory
[14] F. Yang, M. Habibullah, T. Zhang, Z. Xu, P. Lim, S. Nadarajan, Health index- neural-network-based indirect remaining-useful-life prognosis for satellite
based prognostics for remaining useful life predictions in electrical machines, lithium-ion battery, Appl. Sci. 8 (11) (2018) 2078.
IEEE Trans. Ind. Electron. 63 (4) (2016) 2633–2644. [44] D. Kingma, J. Ba, Adam: A method for stochastic optimization, in: Proc. 3rd Int.
[15] A. Chehade, S. Bonk, K. Liu Sensory-based failure threshold estimation for Conf. Learn. Represent. San Diego, CA, USA, 2015..
remaining useful life prediction, IEEE Trans. Rel. 66 (3) (2017) 939-949.. [45] H. Sakoe, S. Chiba, Dynamic programming algorithm optimization for spoken
[16] T. Wang, J. Yu, D. Siegel, J. Lee A similarity-based prognostics approach for word recognition, IEEE Trans. Acoust. Speech 26 (1) (1978) 43–49.
remaining useful life estimation of engineered systems, in: Proc. Int. Conf. [46] A. Saxena, K. Goebel, D. Simon, N. Eklund, Damage propagation modeling for
Prognos. Heal. Manage. 2008, pp. 1–6.. aircraft engine run-to-failure simulation, in: Proc IEEE Int. Trans. Prognostics
[17] R. Khelif, B. Chebel-Morello, S. Malinowski, E. Laajili, F. Fnaiech, N. Zerhouni, Health Manage. Denver, CO, USA, 2008, pp. 1–9.
Direct remaining useful life estimation based on support vector regression, [47] NASA. (Sept. 2019.) Prognostics data repository. [Online]. Available: https://ti.
IEEE Trans. Ind. Electron. 64 (3) (2016) 2276–2285. arc.nasa.gov/tech/dash/groups/pcoe/prognostic-data-repository/.

206
C. Wang, Z. Zhu, N. Lu et al. Neurocomputing 462 (2021) 195–207

Cunsong Wang received the B.S. degree and M.S. degree Yuehua Cheng received the Ph.D. degree from the
in automation and control theory and control engi- Nanjing University of Aeronautics and Astronautics,
neering from Nanchang Hangkong University, Nan- Nanjing, China, in 2012.
chang China, in 2013 and 2016, respectively. He is She is currently a Professor with the College of
currently pursuing the Ph.D degree in control theory Automation Engineering, Nanjing University of Aero-
and control engineering with the College of Automation nautics and Astronautics, Nanjing, China. Her current
Engineering, Nanjing University of Aeronautics and research interests include tolerant control and life pre-
Astronautics. His current research interests include diction for satellite attitude control systems.
data-driven fault prognosis and health management.

Zhehong Zhu received Ph.D. degree from Shanghai Jiao Bin Jiang received the Ph.D. degree in automatic control
Tong University, Shanghai, China and the University of from Northeastern University, Shenyang, China, in 1995.
Toronto, Ontario, Canada. He has been Post-Doctoral Fellow, a Research Fellow,
Dr. Zhu is currently a Member, College of the Royal and a Visiting Professor in Singapore, France, USA and
Society of Canada; a Professor and Tier 1 York Research Canada, respectively. He is currently a Chair Professor of
Chair in space technology; the inaugural academic Cheung Kong Scholar Program in Ministry of Education
Director of research commons, office of vice-president and the Vice President of Nanjing University of Aero-
research and innovation, York University, Toronto, nautics and Astronautics, Nangjing, China.
Canada. His current interests include dynamics and He has been the Principal Investigator on several pro-
control, tethered spacecraft system, electrodynamic jects of the National Natural Science Foundation of
tether propulsion, multibody dynamics, space robotics, China. He has authored eight books and over 200 ref-
visual servo control, computational solid mechanics and ereed international journal articles and conference
multi-physics modeling, space debris mitigation and removal, and multifunctional articles. His current research interests include fault diagnosis and fault-tolerant
Nano CFRP composites for spacecraft. control and their applications in aircrafts, satellites and high-speed trains.
Dr. Jiang received the 2nd Class Award of the National Natural Science of China in
2018. He is the Chair of Control Systems Chapter in IEEE Nanjing Section, a member
Ningyun Lu received the Ph.D. degree from Northeast- of the IFAC Technical Committee on Fault Detection, Supervision, and Safety of
ern University, Shenyang, China, in 2004. Technical Processes. He currently serves as Associate Editor or Editorial Board
From 2002 to 2005, she was a Research Associate and a Member for a number of journals such as the IEEE Transactions On Cybernetics, IEEE
Post-Doctoral Fellow with the Hong Kong University of Transactions On Control Systems Technology, International Journal of Control,
Science and Technology, Hong Kong. She is currently a Automation and Systems, Neurcomputing, Journal of Franklin Institute, Acta
Full Professor with the College of Automation Engi- Automatica Sinica, Control and Decision, and Systems Engineering and Electronics
neering, Nanjing University of Aeronautics and Astro- Technologies.
nautics, Nanjing, China. Her research interest includes
data-driven fault prognosis and diagnosis and its
applications to various industrial processes.

207