An enhanced CNN-LSTM remaining

useful life prediction model for aircraft

engine with attention mechanism
Hao Li1 , Zhuojian Wang2 and Zhe Li2
1
Air Force Engineering University, Graduate School, Xi’an, Shaanxi, China
2
Air Force Engineering University, Aeronautics Engineering College, Xi’an, Shaanxi, China

ABSTRACT
Remaining useful life (RUL) prediction is one of the key technologies of aircraft
prognosis and health management (PHM) which could provide better maintenance
decisions. In order to improve the accuracy of aircraft engine RUL prediction under
real flight conditions and better meet the needs of PHM system, we put forward an
improved CNN-LSTM model based on the convolutional block attention module
(CBAM). First, the features of aircraft engine operation data are extracted by multi-
layer CNN network, and then the attention mechanism is processed by CBAM in
channel and spatial dimensions to find key variables related to RUL. Finally, the hidden
relationship between features and service time is learned by LSTM and the predicted
RUL is output. Experiments were conducted using C-MPASS dataset. Experimental
results indicate that our prediction model has feasibility. Compared with other state-
of-the-art methods, the RMSE of our method decreased by 17.4%, and the score of the
prediction model was improved by 25.9%.

Subjects Algorithms and Analysis of Algorithms, Artificial Intelligence, Data Mining and
Machine Learning, Data Science
Keywords Aircraft engine, Remaining useful life, Convolutional block attention module,
Convolutional neural network, LSTM
Submitted 13 June 2022
Accepted 15 August 2022
Published 30 August 2022 INTRODUCTION
Corresponding author The aircraft engine is one of the core equipment of aircraft, which largely determines
Zhuojian Wang,
zhuojianw1974@[Link]
the flight performance and flight safety of aircraft. With the development of maintenance
theory, sensing and communication technology, aircraft prognosis and health management
Academic editor
Ka-Chun Wong (PHM) is increasingly a significant way to improve the reliability and safety of aircraft (Li,
Additional Information and Verhagen & Curran, 2020). PHM collects operational data in real time through sensors, and
Declarations can be found on realizes the functions of aircraft condition monitoring, fault prediction and maintenance
page 16
planning, among which the remaining useful life(RUL) prediction is a vital part of PHM
DOI 10.7717/peerj-cs.1084 system (Che et al., 2019).
Copyright At present, the RUL prediction methods can be roughly divided into two categories. They
2022 Li et al. are methods based on the physical model and methods based on monitoring data (Peng,
Distributed under Ye & Chen, 2019; Singh et al., 2020; Ren et al., 2018; Shankar kumar, Kumaraswamidhas &
Creative Commons CC-BY 4.0 Laha, 2021). The physical model-based method is usually for constructing and estimating
OPEN ACCESS the degradation model of equipment mathematically by summarizing the rules of failure

How to cite this article Li H, Wang Z, Li Z. 2022. An enhanced CNN-LSTM remaining useful life prediction model for aircraft engine
with attention mechanism. PeerJ Comput. Sci. 8:e1084 [Link]
 mechanism (Lei et al., 2018). This method is more suitable for the crack propagation of
metal components, etc. Ye et al. (2018) studied a fatigue life prediction method based on
multi-scale crack propagation model, and the prediction ability of this method was verified
by nickel-based GH4169 alloy. However, for complex systems and equipment, the method
based on physical model is no longer applicable, mainly because in industrial practice,
we don’t fully grasp the failure mechanism of equipment, and the process of establishing
mathematical model is very complicated. Therefore, more studies on RUL prediction
mainly focus on data-based methods, which could be divided into three categories (Muneer
et al., 2021; Sun et al., 2021; Shankar kumar, Kumaraswamidhas & Laha, 2021). One is to
describe the degradation of equipment through stochastic process. Gao et al. (2019) studied
the characteristics of light-emitting diode driving power, and proposed a RUL prediction
method based on Wiener process. Lin et al. (2021) proposed a novel method to model
two-phase degenerate behavior of products based on the nonlinear Wiener process. The
second is to establish RUL prediction model by machine learning. Ordez et al. (2019)
proposed a RUL prediction model which combined support vector regression model
(SVM) and auto-regressive integrated moving average (ARIMA) model. Pan et al. (2020)
put forward a two-stage prediction method based on extreme learning machine (ELM).
Their method could quickly and accurately predict the RUL of rolling bearings. Kamat,
Sugandhi & Kumar (2021) used the unsupervised machine learning for anomaly trend
analysis of bearings, and the semi-supervised method for RUL prediction. The third
category is the deep learning prediction method that has emerged with the development
of artificial intelligence. Wen et al. (2021) established a deep learning architecture based
on bidirectional gated recurrent unit (BGRU), and proposed a domain-adaptive RUL
prediction model. Guo et al. (2017), Kim & Liu (2021) and Khumprom, Grewell & Yodo
(2020), respectively use three network models, deep belief network (DBN), bayesian deep
learning framework, and feedforward neural network(FNN) to establish different RUL
prediction models, and all of them achieved good prediction results.
Aircraft engine monitoring data has the features of large amount of data and high
dimensionality, and deep learning has better feature extraction ability in processing such
data (Mao et al., 2022; Zhao & Wang, 2021), so the prediction method of deep learning
is more suitable for aircraft engine RUL prediction. At present, some scholars have
also conducted relevant research. Hu et al. (2021) built several new DBRNNs with new
network configurations and designed a new customized loss function for RUL prediction
of aircraft engine. Chui, Gupta & Vasant (2021) proposed a RUL prediction algorithm
in combination with RNN and LSTM, used non-dominated sorting genetic algorithm II
(NSGA-II) to optimize it, and achieved better results. Li et al. (2021) proposed a new
Bayesian deep learning nework considering the epistemic uncertainty and aleatoric
uncertainty to solve the prognostic uncertainty problem. Ragab et al. (2021) addressed
the domain shift problem based on contrastive adversarial domain adaptation and got
better results in experiments. However, the above methods are all based on the research
under the stable working condition; without considering the influence of the working
condition change of the aircraft engine on the prediction accuracy of RUL under the real
flight condition, it still faces the following problems in the practical application process:

Li et al. (2022), PeerJ Comput. Sci., DOI 10.7717/peerj-cs.1084 2/19

 1. At present, the prediction models of aircraft engines are all based on a snapshot of
flight data, the advantages of real-time monitoring data of aircraft engines are not fully
utilized;
2. Because the deep learning network can’t be explained, the RUL prediction model can
only be used as a black box, and the problems in the use of aircraft engines can’t be
found through the prediction results;
3. In some RUL prediction models, complex signal processing is needed to extract features
artificially in advance, which fails to give full play to the features of deep model learning.
In order to effectively provide reliable RUL information for PHM system to better
manage the aircraft engines and reduce the maintenance cost, we propose an improved
CNN-LSTM aircraft engine RUL prediction method based on convolutional block attention
module (CBAM). This method firstly learns all the data in one cycle of aircraft engine by
two dimensional convolutional neural network, extracts the potential relationship between
feature sequences to generate a feature map, then deduces the attention weight along the
two dimensions of spatial and channel by CBAM, and multiplies it with original feature
map to adjust the features adaptively. Finally, LSTM has a good learning ability to time
series information to predict aircraft engine. Compared with other methods, this RUL
prediction model has higher prediction accuracy. In addition, our model can analyze the
effect of the original input data on the final prediction result by visualizing the attention
weight, and provide suggestions for the use and maintenance of aircraft engines.
Our work is arranged as follows. Section 2 briefed the basic theories needed to build the
model described respectively. Section 3 introduced the structure of our RUL prediction
model in detail. The experiment is carried out through C-MPASS simulation data, and the
experimental results are discussed in Section 4. In Section 5, a summary is given.

THEORETICAL METHODS
Convolutional Neural Network (CNN)
CNN is a special multi-layer perceptron (MLP), which has the traits of local connection
and weight sharing. Because of its advantages in image processing, it is widely used in
computer vision, image classification and other applications (Zhou, 2020). CNN can be
divided into one dimensional, two dimensional and three dimensional CNN (1-D CNN,
2-D CNN and 3-D CNN). We mainly use 2-D CNN, which usually includes convolution
layer, activation layer and pooling layer. Figure 1 shows a simple convolutional neural
network.
The crucial part of CNN is convolution layer. Convolution layer convolves input data
through local connection and weight sharing, and its function is to extract features by
shifting convolution filters on original data. Convolution filter is also called receptive field.
Convolution filter moves along all dimensions of input data, calculates the weight and the
dot product of input, and then adds bias to realize convolution operation. The specific
convolution operation is:

yn = wm ∗ xn + bm (1)

Li et al. (2022), PeerJ Comput. Sci., DOI 10.7717/peerj-cs.1084 3/19

 Rectified Linear
Unit (ReLU)
Output feature

Input data

Y
X

2-D convolutional layer Activation layer Pooling layer

Figure 1 The network structure of convolutional neural network.
Full-size DOI: 10.7717/peerjcs.1084/fig-1

where wm and bm ∈ Rh×w represent the weight and bias of the m-th filter, respectively,
and their sizes h × w are specified when designing the network, xn is the n-th region of the
input data, and yn is the convolution output of the filter.
As can be seen from Eq. (1), each convolution filter can generate a corresponding feature
map. When multiple filters are specified in a convolution layer, a corresponding number
of feature maps can be generated, and the output of the convolution layer can be obtained
by stacking all the feature maps in the channel dimension.
The role of activation function is to introduce nonlinear factors to enhance the feature
expression ability of the model. At present, the Rectified Linear Unit (ReLU) is the most
versatile and effective activation function (Dubey & Chakraborty, 2021), which can improve
the network sparsity and reduce the network over-fitting. The expression of ReLU is:
(
x if x > 0
ReLU (x) = (2)
0 if x ≤ 0

where x is the input of the activation function.

The role of pooling layer is to gradually shrink the feature space size, decrease the
amount of parameters of the network and speed up the calculation. The pooling layer is
similar to the convolution layer, and it is also performed by filters. The difference is that
the filters of the pooling layer do not perform convolution operation on the input data,
but perform pooling operation, which is also called downsamples. Generally, the pooling
layer can be divided into average pooling layer and max pooling layer. The average pooling
layer is to extract the average value from the filter, and The max pooling layer is to extract
the max value from the filter. The existing studies indicate that the max pooling layer is
preferable in model training (Raj & Kannan, 2022).

Li et al. (2022), PeerJ Comput. Sci., DOI 10.7717/peerj-cs.1084 4/19

 Channel attention Spatial
Input feature module
Refined feature
attention
module

Figure 2 The network structure of CBAM.

Full-size DOI: 10.7717/peerjcs.1084/fig-2

Convolutional block attention module (CBAM)

Attention mechanism is a data processing method in deep learning, which can learn how
different parts of the input affect the output (Chaudhari et al., 2021). CBAM is a kind of
attention mechanism which could be used in CNN. CBAM is designed as a simple and
competent module to be applied to CNN (Woo et al., 2018). Figure 2 shows the structure
of CBAM. CBAM is composed of the Channel Attention Module (CAM) and Spatial
Attention Module (SAM). CAM and SAM can perform attention operations in the channel
and spatial respectively. This structure design enables CBAM to be quickly combined with
existing model, saving parameters and computing resources.
As shown in Fig. 3, the CAM firstly performs global average pooling and global max
pooling on the input data map D(h × w × c) to output two vectors of size 1 × 1 × c, and
inputs them into two-layer MLP respectively which has c/r neurons in the first layer ,where
r is reduction rate and c neurons in the second layer. The parameters of these two-layer
MLP are shared. Next, add up the outputs of MLP by elements. After that, the result will
be activated by sigmoid function to obtain the channel attention Ac . Finally, Ac and D are
multiplied by elements to obtain the weighted output D0 .D0 , and the D0 is input of SAM.
The channel attention mechanism can be expressed by the formula:

Ac (D) = σ (MLP(Pmax ) + MLP(Pavg ))

= σ W1 W0 (Pmax ) + W1 (W0 (Pavg )) (3)

where Pmax and Pavg are the results of global max pooling and global average pooling of
input features, σ is sigmoid activation function, W0 and W1 are the weights of the first and
second layers of MLP, and their sizes are c/r × c and c × c/r respectively.
As shown in Fig. 4, the SAM uses the data map D0 as input which is the output of CAM,
and gets two feature maps with the size of h × w × 1 through max pooling and average
pooling operation in channel, and takes these two feature maps as input of convolution

Li et al. (2022), PeerJ Comput. Sci., DOI 10.7717/peerj-cs.1084 5/19

 Input feature
global average pool

Sigmoid

Channel attention

global max pool Shared MLP

Figure 3 Channel attention module.

Full-size DOI: 10.7717/peerjcs.1084/fig-3

Input feature Convolution layer

Sigmoid

[maxpool,averagepool] Spatial attention

Figure 4 Spatial attention module.

Full-size DOI: 10.7717/peerjcs.1084/fig-4

layer containing a filter. And the output is activated by sigmoid to get spatial attention As .
Finally, D0 is multiplied with As to get the final feature.
The spatial attention can be expressed as:

As (D0 ) = σ (fconv ([maxpool(D0 );avgpool(D0 )])) (4)

where fconv is convolution operation, maxpool and avgpool are max pooling and average
pooling respectively.

Long and short term memory network (LSTM)

Unlike the general forward feedback network, LSTM is a idiosyncratic kind of recurrent
neural network (RNN) (Livieris, Pintelas & Pintelas, 2020). Compared with RNN, LSTM
can analyze the input in time series, and it can solve the problem of long-term dependence
of the series, avoiding the problems of gradient disappearance and gradient explosion, so
it can be used for the analysis of long-term data (Jang et al., 2020).
Figure 5 shows the internal structure of an LSTM unit which can be mainly divided into
three parts: forget gate, input gate, output gate. In each unit, the cell state C and the output
hidden state h will be updated through the 3 internal gates. The forget gate determines the
forgotten message ft according to the input xt and the previous hidden state ht −1 , the input
gate selects the candidate memory message e Ct through the input control it to determine

Li et al. (2022), PeerJ Comput. Sci., DOI 10.7717/peerj-cs.1084 6/19

 Forget gate Input gate Output gate
ht

Ct-1 Ct
╳ ┼

tanh
it
ft ╳ ╳

Ct ot
σ σ tanh σ
ht
ht-1

xt

Figure 5 Network structure of LSTM.

Full-size DOI: 10.7717/peerjcs.1084/fig-5

the updated content of the previous state Ct −1 , and the output gate determines the hidden
state ht that the current cell state passes to the next cell through the output control ot .
The specific calculation process in LSTM network unit is as follows:

ft = σ (fw · [ht −1 ,xt ] + fb ) (5)

it = σ (iw · [ht −1 ,xt ] + ib ) (6)

Ct = tanh(cw · [ht −1 ,xt ] + cb )

e (7)

Ct = ft Ct −1 + it Ct
e (8)

ot = σ (ow · [ht −1 ,xt ] + ob ) (9)

ht = ot tanh(Ct ) (10)

where fw , iw , cw , ow are the weights of ft , it , eCt , ot respectively, and fb , ib , cb , ob are the

biases of ft , it , Ct , ot respectively, tanh is tanh activation function.
e

Li et al. (2022), PeerJ Comput. Sci., DOI 10.7717/peerj-cs.1084 7/19

 CNN block 1
CBAM

CNN block 2 CNN block 3

variable

s
le
c

Spatial attention
cy

Maxpooling
SequenceInput layer

Batch normalization

Channel attention
2D-Convolution

Relu
time

Input data

Batch normalization
RUL

Sigmoid
FCL
LSTM

Figure 6 The network structure of the proposed method.

Full-size DOI: 10.7717/peerjcs.1084/fig-6

PROPOSED METHODOLOGY
Our proposed RUL prediction model contains a variety of deep learning layers. Figure 6
shows the network structure of our model. First, input data is converted into the data
map as shown in Fig. 6, and then the CNN module will extract the features of the input
data firstly. In this process, we use three CNN blocks with the same structure. In addition,
we introduce the batch normalization layer into the CNN network, its function is to
independently normalize the small batch data of all observations in each channel, improve
the training speed and effect of CNN and reduce the influence of initialization parameters
on training process. Then, the CBAM module infers attention, and carries out adaptive
weighting processing. Finally, the obtained feature map is input into LSTM network to
mine the hidden relationship between features and service time. Finally, the predicted RUL
value is obtained through a fully connected layer and activated by sigmoid function.
The above is the design process of our model. For the training, verification and
deployment of the model, our process is shown in Fig. 7. After the design of our model
is completed, the training data set and test data set should be prepare after. Training data
is used as input data to train RUL prediction model and get the trained model. After that
the trained model will be verified with the test data, and the performance will be evaluated
by the specified evaluation method. If the trained model can meet the available standards,
we can deploy the model to the aircraft fleet. After each flight, we collect the real-time
monitoring data for the same preprocessing, and input it into the RUL prediction model.
And then according to the prediction results and the set maintenance threshold, we can

Li et al. (2022), PeerJ Comput. Sci., DOI 10.7717/peerj-cs.1084 8/19

 Preprocessing procedure
Training RUL
Training flight data
prediction model

is performance
Test flight data Trained model evaluation meet the credibility
requirements ?
Yes

Real-time monitoring Deploy RUL Maintenance Yes

Maintenance decision
data prediction model threshold

Figure 7 Realization process of aircraft engine RUL prediction with proposed method.
Full-size DOI: 10.7717/peerjcs.1084/fig-7

make a reasonable decision on engine maintenance, and improve the accuracy of aircraft
engine maintenance activities.

EXPERIMENTAL STUDY
Dataset description
The dataset used in our experiment is the latest Aircraft Engine Run-to-Failure data set
published by NASA Ames Research Center in 2020. The dataset was obtained through
C-MPASS platform which could simulate the working process of turbofan aircraft engine,
its structure is shown in Fig. 8. Compared with the previous run-to-failure trajectories,
it considered the performance degradation behavior of the aircraft engine in real flight
conditions, and recorded all the flight environment parameters, operation data and
degradation data (Chao et al., 2021).
In order to compare with other methods, we use DS02, the most widely used data set in
this data set, which contains the run-to-failure simulation data of nine engines, of which
six (unit =2,5,10,16,18,20) are the training set data and three (unit =11, 14, 15) are the
test set data. The recorded data of each engine contains four scenario descriptors (W ),
13 measured physical properties (Xs ), 18 virtual sensors data (Xv ) and 10 model health
parameters (θ ). Table 1 shows the description of monitoring parameters. The units are
divided into three flight classes (Flight class 1, Flight class 2, and Flight class 3) according to
the length of operation time (short-length flights, medium-length flights, and long-length
flights). Figure 9 shows the working conditions of different flight classes of engines in the
flight envelope, and the green area in the figure is the flight envelope of the engine.
Figure 10 shows the kernel density estimations of scenario descriptors (W ) of each
engine. From the distribution, we can see that the working environment of different
engines is different. Especially, compared with other engines, the working altitude of
unit14 is lower. Under this more realistic condition, the RUL of the aircraft engine is
harder to predict.

Li et al. (2022), PeerJ Comput. Sci., DOI 10.7717/peerj-cs.1084 9/19

 Figure 8 Schematic representation of C-MAPSS model (Chao et al., 2021).
Full-size DOI: 10.7717/peerjcs.1084/fig-8

Experimental process
Data preprocessing
The preprocessing of the monitoring data includes slicing and normalization. Slicing
refers to selecting a part of the recorded data of the aircraft engine as the input data of
RUL prediction. This is because the monitoring data of aircraft engine in one service
cycle contains a large amount of repeated and redundant information. We divide it into
400 copies equally according to the total time in one service cycle, and the data of the
corresponding time is extracted and recombined into input data, which can ensure that the
input data can describe the RUL information of the aircraft engine, reduce the computing
resources and improve the learning speed.
In order to eliminate the training error caused by different measuring units of monitoring
data, we normalize the data monitored by different sensors. The normalization method
adopted in this article is Min-Max normalization, and the calculation formula is:
d − dmin
d=
b (11)
dmax − dmin
dmax , dmin are the maximum and minimum value of the variable, and d, b
d are the
original and normalized values of the variable.

Evaluation metric
In order to quantify the performance of our proposed RUL prediction model, we select the
RMSE and Score to measure the prediction accuracy of the model (Custode et al., 2022).
RMSE is the root mean square error, which is often used as a measure of error in deep
learning models, and can reflect the overall deviation between the predicted RUL and true

Li et al. (2022), PeerJ Comput. Sci., DOI 10.7717/peerj-cs.1084 10/19

 Table 1 The description of aircraft engine measurement data.

Symbol Description Units

1 alt Altitude ft
2 Mach Flight Mach number –
3 TRA Throttle-resolver angle %
◦
4 T2 Total temperature at fan inlet R
5 Wf Fuel flow pps
6 Nf Physical fan speed rpm
7 Nc Physical core speed rpm
◦
8 T24 Total temperature at LPC outlet R
◦
9 T30 Total temperature at HPC outlet R
◦
10 T48 Total temperature at HPT outlet R
◦
11 T50 Total temperature at LPT outlet R
12 P15 Total pressure in bypass-duct psia
13 P21 Total pressure at fan outlet psia
14 P24 Total pressure at LPC outlet psia
15 Ps30 Static pressure at HPC outlet psia
16 P40 Total pressure at burner outlet psia
17 P50 Total pressure at LPT outlet psia
◦
18 T40 Total temp. at burner outlet R
19 P30 Total pressure at HPC outlet psia
20 P45 Total pressure at HPT outlet psia
21 W21 Fan flow pps
22 W22 Flow out of LPC lbm/s
23 W25 Flow into HPC lbm/s
24 W31 HPT coolant bleed lbm/s
25 W32 HPT coolant bleed lbm/s
26 W48 Flow out of HPT lbm/s
27 W50 Flow out of LPT lbm/s
28 epr Engine pressure ratio (P50/P2) –
29 SmFan Fan stall margin –
30 SmLPC LPC stall margin –
31 SmHPC HPC stall margin –
32 NRf Corrected fan speed rpm
33 NRc Corrected core speed rpm
34 PCNfR Percent corrected fan speed pct
35 phi Ratio of fuel flow to Ps30 pps/psi
36 fan_eff_mod Fan efficiency modifier –
37 fan_flow_mod Fan flow modifier –
38 LPC_eff_mod LPC efficiency modifier –
39 LPC_flow_mod LPC flow modifier –
40 HPC_eff_mod HPC efficiency modifier –

(continued on next page)

Li et al. (2022), PeerJ Comput. Sci., DOI 10.7717/peerj-cs.1084 11/19

 Table 1 (continued)
Symbol Description Units
41 HPC_flow_mod HPC flow modifier –
42 HPT_eff_mod HPT efficiency modifier –
43 HPT_flow_mod HPT flow modifier –
44 LPT_eff_mod LPT efficiency modifier –
45 LPT_flow_mod HPT flow modifier –

Figure 9 Flight envelope.

Full-size DOI: 10.7717/peerjcs.1084/fig-9

#10-4
30
4

20
Density

Density

2
10 Unit 2
Unit 5
Unit 10
0 0 Unit 11
1 2 3 4 0.2 0.4 0.6 0.8
Altitude/ft 4 Mach Number Unit 14
#10
0.15 Unit 15
Unit 16
0.06 Unit 18
0.1 Unit 20
Density

Density

0.04
0.05
0.02

0 0
20 40 60 80 100 400 420 440 460 480 500
Throttle Resolver Angle/% Temperature at fan inlet/°R

Figure 10 Kernel density estimations of scenario descriptors of different aircraft engines.

Full-size DOI: 10.7717/peerjcs.1084/fig-10

Li et al. (2022), PeerJ Comput. Sci., DOI 10.7717/peerj-cs.1084 12/19

 RUL. It can be expressed as following formula:
v
u n
u1 X
RMSE = t (Rip − Rit )2 (12)
n
i=1

where Rpi is the ith engine predicted value of RUL , Rt i is the ith engine true value of RUL,
and n is the number of predicted RUL.
Score function is an evaluation function designed by the data set provider. The lower
the score, the better the prediction effect of the model. Its mathematical expression is:
 n i i
 X exp(− Rp − Rt ) − 1 if (Ri − Ri ≤ 0)

p t

13


Score = i=1
n . (13)
X Rip − Rit
) − 1 if (Rp − Rt > 0)
i i

exp(


10


i=1

Parameter settings of the model

In our prediction model, some parameters of network layers need to manually set, e.g.,
the size and number of filters in convolution layer. For C-MPASS data set, the parameter
settings of some network layers in our RUL prediction model are shown in Table 2. Under
this setting, there are about 3 × 106 learnable parameters in our network.
As for the setting of training options, we set the number of iterations to 400, optimize
the network parameters by using the adaptive moment estimation (Adam) optimization
algorithm, and the batch size is six, the initial learning rate is 0.01, and set the gradient
threshold to two to prevent the gradient explosion.

Experimental results and analysis

Comparison and discussion of prediction results
After the training process, input the test set data into the trained model and get the
predicted RUL values. Figure 11 shows our RUL prediction results. As shown in Fig. 11, the
predicted RUL values of our RUL prediction model are generally distributed near the true
values during the whole aircraft engine service life. Therefore, our model can be applied to
predict aircraft engine RUL, and the prediction results can provide reliable suggestions for
the use and maintenance of aircraft engines.
After the RUL prediction result is obtained, the result is evaluated by the evaluation
function provided previously to quantify the model performance. In addition, we also
use MLP, FNN, CNN and CNN-LSTM (Kong et al., 2019) to predict the test data set. The
comparison results are shown in Table 3. Judging from the comparison results in the table,
our improved CNN-LSTM based on CBAM prediction model has better performance. Its
Score is improved by about 25.9% and its RMSE is reduced by 17.4% compared with other
best methods.

Analysis of prediction process

In the process of RUL prediction, different variables have different influences on the
prediction results in the input data map, and this difference can be described by CBAM’s

Li et al. (2022), PeerJ Comput. Sci., DOI 10.7717/peerj-cs.1084 13/19

 Table 2 RUL prediction model parameter setting.

Layer name Parameters

SequenceInput Layer size: (400,42,1)
Convolution_2d layer 1 filter size: (10, 3); number of filters: 10; step size: (10, 3)
Maxpooling_2d Layer 1 filter size: (5, 5); padding: same
Convolution_2d layer 2 filter size: (3, 3); number of filters: 10
Maxpooling_2d Layer 2 filter size: (5, 5); padding: same
Convolution_2d layer 3 filter size: (5, 5); number of filters: 10
Maxpooling_2d Layer 3 filter size: (5, 5); padding: same
Channel attention Layer number of channels: 10; reduction rate: 2
Spatial attention Layer filter size: (3, 3)
Lstm Layer number of hidden units: 128
FullyConnected Layer output size: 1

Unit 11 Unit 14 Unit 15

70 80 70

60 60 Predicted
60 True
50 50
RUL/cycles

RUL/cycles

RUL/cycles
40 40
40
30 30

20 20
20
10 10

0 0 0
0 20 40 60 0 20 40 60 80 0 20 40 60 80
Time/cycles Time/cycles Time/cycles

Figure 11 RUL prediction result of test data set.

Full-size DOI: 10.7717/peerjcs.1084/fig-11

Table 3 Performance comparison with other neural network methods.

Method RMSE Score (×105 )

MLP 8.34 13.43
FNN 7.89 9.21
CNN 7.14 7.60
CNN-LSTM 6.66 7.80
Ours 5.50 5.78

spatial attention in our model. Therefore, we extracted the spatial attention matrix in the
5th, 15th, 25th, 35th, 45th and 55th cycles of unit 11, mapped it into the size of the input
data, and the result is displayed in the form of the hot map in Fig. 12.
In the figure, we are able to distinctly see that the weights of different variables are
obviously different. A larger weight means that the network will focus more on the
changes of the corresponding variables in RUL prediction, while a smaller weight means
the opposite. We found four variables with the largest weights, namely HPT _eff _mod,
HPT _flow_mod, LPT _eff _mod and LPT _flow_mod, and showed the changes of their

Li et al. (2022), PeerJ Comput. Sci., DOI 10.7717/peerj-cs.1084 14/19

 cycle=5 cycle=15 cycle=25
0.6
0.6 0.5
0.5
0.5
0.4

Time

Time

Time
0.4 0.4

0.3 0.3 0.3

0.2
0.2
0.2
Variables Variables Variables
cycle=35 cycle=45 cycle=55
0.6
0.8
0.8
0.5
0.6 0.6
Time

Time

Time
0.4
0.4 0.4
0.3
0.2
0.2
0.2
Variables Variables Variables

Figure 12 Visualization of spatial attention.

Full-size DOI: 10.7717/peerjcs.1084/fig-12

#10 -3
0 1

-5 0.5
HPT_flow_mod
HPT_eff_mod

-10 0

-15 -0.5
Unit 2
-20 -1 Unit 5
0 2 4 6 8 10 12 0 2 4 6 8 10 12 Unit 10
Time/s #10 5 Time/s #10 5 Unit 16
0 0
Unit 18
-0.005 Unit 20
-0.005
LPT_flow_mod
LPT_eff_mod

-0.01
-0.01
-0.015
-0.015
-0.02
-0.02
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time/s #10 5 Time/s #10 5

Figure 13 Traces of the degradation imposed on the the four variables.

Full-size DOI: 10.7717/peerjcs.1084/fig-13

values with the service time in Fig. 13. The results show that in the whole life cycle, the
changes of the values of most variables are strongly correlated with the service time, so
their values also reflect the health of aircraft engines to some extent.

CONCLUSIONS
This article presents an enhanced RUL prediction method of aircraft engine based on
CBAM. It mainly combines CBAM attention mechanism in the CNN-LSTM network. The

Li et al. (2022), PeerJ Comput. Sci., DOI 10.7717/peerj-cs.1084 15/19

 experimental verification is carried out by C-MPASS engine degradation data set, and the
following conclusions are obtained:
1. The CNN-LSTM network with attention mechanism has better feature extraction
ability and time series analysis ability. Compared with the existing methods, it can
analyze all the data of aircraft engine running under different working conditions. The
experimental results indicate that the performance of the proposed prediction model
is good, and most of predicted RUL values are near the true RUL values, which has
high feasibility. Compared with the current methods with high prediction accuracy,
the RMSE of our method decreased by 17.4%, and the score of the prediction model
was improved by 25.9%;
2. CBAM module can effectively find the key variables that reflect the health condition
of aircraft engine. By paying more attention to these key variables, the CNN-LSTM
network can have better prediction accuracy. Furthermore, CBAM makes the model
interpretable to some extent, which can provide some help for sensor arrangement and
troubleshooting of aircraft engines;
3. Our proposed RUL prediction method doesn’t involve any analysis on aircraft engine
composition, fault mechanism, etc., so the model can be deployed and applied quickly.
The RUL information provided by our method can assist the maintenance management
system to make decisions, in order to ensure the reliability of the aircraft engine and
reduce the maintenance cost.

ADDITIONAL INFORMATION AND DECLARATIONS

Funding
This work was supported by the National Natural Science Foundation of China (No.
61873351). The funders had no role in study design, data collection and analysis, decision
to publish, or preparation of the manuscript.

Grant Disclosures
The following grant information was disclosed by the authors:
The National Natural Science Foundation of China: No. 61873351.

Competing Interests
The authors declare there are no competing interests.

Author Contributions
• Hao Li conceived and designed the experiments, performed the experiments, analyzed
the data, performed the computation work, authored or reviewed drafts of the article,
and approved the final draft.
• Zhuojian Wang conceived and designed the experiments, authored or reviewed drafts
of the article, and approved the final draft.
• Zhe Li performed the experiments, performed the computation work, prepared figures
and/or tables, and approved the final draft.

Li et al. (2022), PeerJ Comput. Sci., DOI 10.7717/peerj-cs.1084 16/19

 Data Availability
The following information was supplied regarding data availability:
The data is available at figshare: Li, Hao (2022): N-CMAPSS_DS02-006.h5. figshare.
Dataset. [Link]
The code is available at GitHub: [Link]
[Link].

Supplemental Information
Supplemental information for this article can be found online at [Link]
peerj-cs.1084#supplemental-information.

REFERENCES
Chao MA, Kulkarni CS, Goebel KF, Fink O. 2021. Aircraft engine run-to-failure
dataset under real flight conditions for prognostics and diagnostics. Data 6:5
DOI 10.3390/data6010005.
Chaudhari S, Polatkan G, Ramanath R, Mithal V. 2021. An attentive survey of attention
models. ACM Transactions on Intelligent Systems and Technology (TIST) 12:1–32.
Che C, Wang H, Fu Q, Ni X. 2019. Combining multiple deep learning algorithms for
prognostic and health management of aircraft. Aerospace Science and Technology
94:105423 DOI 10.1016/[Link].2019.105423.
Chui KT, Gupta BB, Vasant P. 2021. A genetic algorithm optimized RNN-LSTM model
for remaining useful life prediction of turbofan engine. Electronics 10(3):285
DOI 10.3390/electronics10030285.
Custode LL, Mo H, Ferigo A, Iacca G. 2022. Evolutionary optimization of spik-
ing neural P systems for remaining useful life prediction. Algorithms 15:98
DOI 10.3390/a15030098.
Dubey SR, Chakraborty S. 2021. Average biased ReLU based CNN descriptor for
improved face retrieval. Multimedia Tools and Applications 80:23181–23206.
Gao Z, Yin X, Zhang B, Chen M, Li B. 2019. A Wiener processbased remaining life
prediction method for light-emitting diode driving power in rail vehicle carriage.
Advances in Mechanical Engineering 11:1–8.
Guo L, Li N, Jia F, Lei Y, Lin J. 2017. A recurrent neural network based health indicator
for remaining useful life prediction of bearings. Neurocomputing 240:98–109
DOI 10.1016/[Link].2017.02.045.
Hu K, Cheng Y, Wu J, Zhu H, Shao X. 2021. Deep bidirectional recurrent neural
networks ensemble for remaining useful life prediction of aircraft engine. IEEE
Transactions on Cybernetics 12:1–13.
Jang B, Kim M, Harerimana G, Kang S-u, Kim JW. 2020. Bi-LSTM model to increase
accuracy in text classification: combining Word2vec CNN and attention mechanism.
Applied Sciences 10(17):5841 DOI 10.3390/app10175841.
Kamat P, Sugandhi R, Kumar S. 2021. Deep learning-based anomaly-onset aware
remaining useful life estimation of bearings. PeerJ Computer Science 7:e795
DOI 10.7717/peerj-cs.795.

Li et al. (2022), PeerJ Comput. Sci., DOI 10.7717/peerj-cs.1084 17/19

 Khumprom P, Grewell D, Yodo N. 2020. Deep neural network feature selection
approaches for data-driven prognostic model of aircraft engines. Aerospace 7(9):132
DOI 10.3390/aerospace7090132.
Kim M, Liu K. 2021. A Bayesian deep learning framework for interval estimation of
remaining useful life in complex systems by incorporating general degradation char-
acteristics. IISE Transactions 53(3):326–340 DOI 10.1080/24725854.2020.1766729.
Kong Z, Cui Y, Xia Z, Lv H. 2019. Convolution and long short-term memory hybrid
deep neural networks for remaining useful life prognostics. Applied Sciences
9(19):4156 DOI 10.3390/app9194156.
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. Mechanical Systems and Signal
Processing 104:799–834 DOI 10.1016/[Link].2017.11.016.
Li G, Yang L, Lee C-G, Wang X, Rong M. 2021. A bayesian deep learning RUL frame-
work integrating epistemic and aleatoric uncertainties. IEEE Transactions on
Industrial Electronics 68:8829–8841 DOI 10.1109/TIE.2020.3009593.
Li R, Verhagen WJ, Curran R. 2020. Toward a methodology of requirements definition
for prognostics and health management system to support aircraft predictive mainte-
nance. Aerospace Science and Technology 102:105877 DOI 10.1016/[Link].2020.105877.
Lin J, Liao G, Chen M, Yin H. 2021. Two-phase degradation modeling and remaining
useful life prediction using nonlinear wiener process. Computers & Industrial
Engineering 160:107533 DOI 10.1016/[Link].2021.107533.
Livieris IE, Pintelas EG, Pintelas PE. 2020. A CNNLSTM model for gold price time-
series forecasting. Neural Computing and Applications 32:1735117360.
Mao P, Lin Y, Xue S, Zhang B. 2022. Remaining useful life estimation of aircraft
engines using differentiable architecture search. Mathematics 10(3):352
DOI 10.3390/math10030352.
Muneer A, Taib SM, Naseer S, Ali RF, Aziz IA. 2021. Data-driven deep learning-based
attention mechanism for remaining useful life prediction: case study application to
turbofan engine analysis. Electronics 10:2453 DOI 10.3390/electronics10202453.
Ordez C, Snchez Lasheras F, Roca-Pardias J, de Cos Juez FJ. 2019. A hybrid ARI-
MASVM model for the study of the remaining useful life of aircraft engines. Journal
of Computational and Applied Mathematics 346:184–191
DOI 10.1016/[Link].2018.07.008.
Pan Z, Meng Z, Chen Z, Gao W, Shi Y. 2020. A two-stage method based on ex-
treme learning machine for predicting the remaining useful life of rolling-
element bearings. Mechanical Systems and Signal Processing 144:106899
DOI 10.1016/[Link].2020.106899.
Peng W, Ye Z, Chen N. 2019. Joint online RUL prediction for multivariate dete-
riorating systems. IEEE Transactions on Industrial Informatics 15:2870–2878
DOI 10.1109/TII.2018.2869429.
Ragab M, Chen Z, Wu M, Foo C-S, Keong KC, Yan R, Li X. 2021. Contrastive adversarial
domain adaptation for machine remaining useful life prediction. IEEE Transactions
on Industrial Informatics 17:5239–5249 DOI 10.1109/TII.2020.3032690.

Li et al. (2022), PeerJ Comput. Sci., DOI 10.7717/peerj-cs.1084 18/19

 Raj FVA, Kannan VK. 2022. Particle swarm optimized deep convolutional neural
sugeno-takagi fuzzy PID controller in permanent magnet synchronous motor. Inter-
national Journal of Fuzzy Systems 24(1):180–201 DOI 10.1007/s40815-021-01126-6.
Ren L, Sun Y, Cui J, Zhang L. 2018. Bearing remaining useful life prediction based on
deep autoencoder and deep neural networks. Journal of Manufacturing Systems
48:71–77 DOI 10.1016/[Link].2018.04.008.
Shankar kumar P, Kumaraswamidhas L, Laha S. 2021. Bearing degradation as-
sessment and remaining useful life estimation based on Kullback-Leibler
divergence and Gaussian processes regression. Measurement 174:108948
DOI 10.1016/[Link].2020.108948.
Singh J, Azamfar M, Li F, Lee J. 2020. A systematic review of machine learning al-
gorithms for prognostics and health management of rolling element bearings:
fundamentals, concepts and applications. Measurement Science and Technology
32(1):012001 DOI 10.1088/1361-6501/ab8df9.
Sun B, Li Y, Wang Z, Ren Y, Feng Q, Yang D. 2021. An improved inverse Gaussian pro-
cess with random effects and measurement errors for RUL prediction of hydraulic
piston pump. Measurement 173:108604 DOI 10.1016/[Link].2020.108604.
Wen B, Xiao MQ, Wang XQ, Zhao X, Li J, Chen X. 2021. Data-driven remaining
useful life prediction based on domain adaptation. PeerJ Computer Science 7:e690
DOI 10.7717/peerj-cs.690.
Woo S, Park J, Lee J-Y, Kweon IS. 2018. CBAM: convolutional block attention module.
In: Computer vision –ECCV 2018. 3–19.
Ye S, Zhang C, yue Zhang P, Zhang X, Tu S, Wang R-Z. 2018. Fatigue life prediction of
nickel-based GH4169 alloy on the basis of a multi-scale crack propagation approach.
Engineering Fracture Mechanics 199:29–40 DOI 10.1016/[Link].2018.05.023.
Zhao Y, Wang Y. 2021. Remaining useful life prediction for multi-sensor systems
using a novel end-to-end deep-learning method. Measurement 182:109685
DOI 10.1016/[Link].2021.109685.
Zhou D-X. 2020. Universality of deep convolutional neural networks. Applied and
Computational Harmonic Analysis 48(2):787–794 DOI 10.1016/[Link].2019.06.004.

Li et al. (2022), PeerJ Comput. Sci., DOI 10.7717/peerj-cs.1084 19/19