sensors

Article
A Remaining Useful Life Prognosis of Turbofan Engine Using
Temporal and Spatial Feature Fusion
Cheng Peng 1,2 , Yufeng Chen 1 , Qing Chen 1 , Zhaohui Tang 2, * , Lingling Li 1 and Weihua Gui 2

1 School of Computer, Hunan University of Technology, Zhuzhou 412007, China; [email protected] (C.P.);
[email protected] (Y.C.); [email protected] (Q.C.); [email protected] (L.L.)
2 School of Automation, Central South University, Changsha 410083, China; [email protected]
* Correspondence: [email protected]

Abstract: The prognosis of the remaining useful life (RUL) of turbofan engine provides an important
basis for predictive maintenance and remanufacturing, and plays a major role in reducing failure
rate and maintenance costs. The main problem of traditional methods based on the single neural
network of shallow machine learning is the RUL prognosis based on single feature extraction, and
the prediction accuracy is generally not high, a method for predicting RUL based on the combination
of one-dimensional convolutional neural networks with full convolutional layer (1-FCLCNN) and
long short-term memory (LSTM) is proposed. In this method, LSTM and 1- FCLCNN are adopted
to extract temporal and spatial features of FD001 andFD003 datasets generated by turbofan engine
respectively. The fusion of these two kinds of features is for the input of the next convolutional neural
networks (CNN) to obtain the target RUL. Compared with the currently popular RUL prediction
models, the results show that the model proposed has higher prediction accuracy than other models in
RUL prediction. The final evaluation index also shows the effectiveness and superiority of the model.

Keywords: remaining useful life (RUL); long short-term memory (LSTM); one-dimensional convo-



lutional neural networks with full convolutional layer (1-FCLCNN); temporal and spatial features;


turbofan engine
Citation: Peng, C.; Chen, Y.; Chen,
Q.; Tang, Z.; Li, L.; Gui, W. A
Remaining Useful Life Prognosis of
Turbofan Engine Using Temporal and
1. Introduction
Spatial Feature Fusion. Sensors 2021,
21, 418. https://doi.org/10.3390/
Turbofan engine is a highly complex and precise thermal machinery, which is the
s21020418 “heart” of the aircraft. About 60% of the total faults of the aircraft are related to the tur-
bofan engine [1]. The RUL prediction of the turbofan engine will provide an important
Received: 14 December 2020 basis for predictive maintenance and pre maintenance. In recent years, because of the
Accepted: 6 January 2021 rapid development of machine learning and deep learning, the intelligence and work effi-
Published: 8 January 2021 ciency of turbofan engines have been greatly improved. However, as large-scale precision
equipment, its operation process cannot be separated from the comprehensive influence of
Publisher’s Note: MDPI stays neu- internal factors such as physical and electrical characteristics and external factors such as
tral with regard to jurisdictional clai- temperature and humidity [2]. The performance degradation process also shows temporal
ms in published maps and institutio- and spatial characteristics [3], which provide necessary data support and bring challenges
nal affiliations. for the RUL prediction of the turbofan engine. At the same time, the data generated
by the operation process of turbofan engine have the characteristics of nonlinearity [4],
time-varying [5], large scale and high dimension [6], which results in the failure of effective
Copyright: © 2021 by the authors. Li-
feature extraction, and the non-linear relationship between the extracted features and the
censee MDPI, Basel, Switzerland.
RUL cannot be mapped, which are the key problems to be solved urgently.
This article is an open access article
Many models have been developed to predict the RUL of turbofan engines. Ah-
distributed under the terms and con- madzadeh et al. [7] divided the predicting methods into four categories, including experi-
ditions of the Creative Commons At- mental, physics-based, data driven, and hybrid methods. Experimental type relies on prior
tribution (CC BY) license (https:// knowledge and historical data, but the operating conditions and operating environment of
creativecommons.org/licenses/by/ the equipment are uncertain, which leads to large prediction accuracy error, and cannot
4.0/). be promoted in complex scenarios. The physical model uses the physical and electrical

Sensors 2021, 21, 418. https://doi.org/10.3390/s21020418 https://www.mdpi.com/journal/sensors

Sensors 2021, 21, 418 2 of 20

characteristics of the equipment to construct accurate mathematical equations to describe

the degradation law of the equipment and predict its remaining life. Usually, it is difficult
to obtain the physical model for large precision equipment such as turbofan engines, and its
application is restricted. Data driven is independent of the failure mechanism of equipment,
its key is to monitor and extract effective performance degradation data. The method lacks
an analysis of the uncertainty of the predicted results, and a large amount of historical
data is needed to build a high-precision model. Hybrid model is a new prediction method
which combines two or more neural network models, which has become the mainstream
research trend. Among them, the hybrid model composed of CNN [8–10] and LSTM is the
most common one in the field of RUL prediction of turbofan engine. CNN has a strong
feature extraction ability, which cannot only extract local abstract features, but also pro-
cess the data with multiple working conditions and multiple faults [11–13], especially the
one-dimensional CNN can be well applied to the time series analysis generated by sensors
(such as gyroscope or accelerometer data [14–16]). It can also be used to analyze signal with
fixed length period (such as audio signal). Zhang et al. [17] adopted a fully convolutional
neural network for feature self-learning and reduced training parameters; a weighted
average method was used to denoise the prediction results, and the bearing-accelerated life
experiment verified the effectiveness of the proposed method. Yang et al. [18] proposed an
intelligent RUL prediction method based on the dual CNN model architecture to predict
the turbofan engine RUL, the first CNN model determines the initial failure point, and
the second CNN model is used for RUL prediction. This method does not require any
feature extractor. The original vibration signal can be received, and useful information can
be retained as much as possible. The prediction results and evaluation indicators prove
the effectiveness and superiority of the method. Hsu [19] applied several deep learning
methods to assess the status of aircraft engines in operation, and to classify the stages of
operational degradation so as to predict the functional remaining lifespan of components.
Li et al. [20] designed a new data-driven method using deep convolutional neural network
(DCNN) for prediction, time windows are used for sample preparation to better extract
features, and experiments based on the C-MAPSS data set have confirmed the effectiveness
of this method. In fact, for the turbofan engine dataset (C-MAPSS), recent studies have
used CNN to extract its features and achieved good results, while one-dimensional CNN
can extract sensor data from the dataset and the full convolutional layer can reduce training
parameters and weights.
Long short-term memory network (LSTM) is a kind of time recurrent neural network
(RNN), which can solve the problems of gradient disappearance and gradient explosion
in RNN. For LSTM, the special "three gate structure" enables it to capture a long range of
dependence and process time series data. While, the RUL prediction of turbofan engine
needs to process the time series data, and LSTM can obtain the optimal features of the
time series data generated by the turbofan engine, and can also mine rules in time series.
Zhang et al. [21] applied a method based on LSTM, which is specifically used to discover
the underlying patterns embedded in the time series, so as to track the system performance
degradation, thereby predicting RUL. Kong et al. [22] utilized polynomial regression to
obtain health indicators, and then combined them with CNN and LSTM neural networks to
extract spatiotemporal features. Song et al. [23] proposed a hybrid health prediction model
that combines the advantages of the autoencoder neural network and the bidirectional
long-term short-term memory (BLSTM) neural network, using the autoencoder as a feature
extraction tool, and the BLSTM captures the characteristics of the bidirectional long-range
dependence of features. The above methods are all tested on the C-MAPSS data set to verify
the effectiveness and accuracy; however, there are common problems such as complex
training process and low prediction accuracy.
We found that the data set of turbofan engine is composed of multiple time series, the
data in different data sets contain different noise levels, so it is necessary to normalize the
original data, which will eliminate the influence of noise, and realize data centralization to
enhance the generalization ability of the model. At the same time, it is difficult to capture
 Sensors 2021, 21, x FOR PEER REVIEW 3 of 21
Sensors 2021, 21, 418 3 of 20

We found that the data set of turbofan engine is composed of multiple time series,
the data in different data sets contain different noise levels, so it is necessary to normalize
multi fault mode and multi-dimensional feature data in different operating environments.
the original data, which will eliminate the influence of noise, and realize data centraliza-
It istion
alsotonecessary to use multi scene and multi time point data to extract effective features
enhance the generalization ability of the model. At the same time, it is difficult to
to improve prediction
capture multi fault mode accuracy, traditional methods
and multi-dimensional featurecannot
data in extract temporal
different operating anden-spatial
features simultaneously and effectively fuse them. In addition, single
vironments. It is also necessary to use multi scene and multi time point data to extract neural network
model is difficult to extract enough effective information in the face
effective features to improve prediction accuracy, traditional methods cannot extract of multiple working
conditions
temporaland and multiple types simultaneously
spatial features of features. and effectively fuse them. In addition, sin-
gleThe main
neural contributions
network of this paper
model is difficult include:
to extract enough(1) use LSTM
effective to extract
information in thethe temporal
face of
multiple working
characteristics of theconditions and multiple
data sequence, and types
learnof features.
how to model the sequence according to
the target TheRUL
maintocontributions
provide accurateof thisresults.
paper include: (1) use LSTM to extract
(2) A one-dimensional the temporal layer
full-convolutional
characteristics of the data sequence, and learn how to model
neural network is adopted to extract spatial features, and through dimensionality the sequence according to
reduction
the target RUL to provide accurate results. (2) A one-dimensional full-convolutional
processing, the parameters and computational complexity of the training process are greatly layer
neural network is adopted to extract spatial features, and through dimensionality reduc-
reduced. (3) The spatiotemporal features extracted by the two models are fused and used
tion processing, the parameters and computational complexity of the training process are
as the input of the one-dimensional convolutional neural network for secondary processing.
greatly reduced. (3) The spatiotemporal features extracted by the two models are fused
Comparing
and used as this
themethod
input ofwith other mainstream
the one-dimensional RUL prediction
convolutional methods,
neural network forthe score and
second-
errorary processing. Comparing this method with other mainstream RUL prediction methods, the
control of the method proposed in this article are better than others, which proves
feasibility
the scoreand
andeffectiveness
error control of of the
thismethod
method. proposed in this article are better than others,
The proves
which rest ofthethis article and
feasibility is arranged as of
effectiveness follows: Part 2 is the basic theory, mainly
this method.
introducing theofmodel
The rest structure
this article of neural
is arranged network
as follows: and
Part 2 isevaluation indicators.
the basic theory, mainlyThe in- third
parttroducing the of
is the focus model
this structure of neural
article, mainly networkthe
including and evaluation
proposed indicators.
model The third
structure, algorithm,
part isprocess,
training the focusimplementation
of this article, mainly
flow.including
The fourththepart
proposed
is themodel structure,
experiment andalgorithm,
result analysis,
training process, implementation
and the last part is summary and prospect. flow. The fourth part is the experiment and result
analysis, and the last part is summary and prospect.
2. Basic Theory
2.1.2.Convolutional
Basic Theory
Neural Network
2.1. Convolutional Neural Network
Convolutional neural network has been widely used in image recognition, complex
Convolutional neural network has been widely used in image recognition, complex
data processing [24]. The convolutional neural network consists of the input layer, convo-
data processing [24]. The convolutional neural network consists of the input layer, con-
lutional layer, pooling layer, full connection layer, and output layer. Its basic components
volutional layer, pooling layer, full connection layer, and output layer. Its basic compo-
arenents
shown in Figure 1.
are shown in Figure 1.

Figure
Figure 1. 1. Traditionalconvolutional
Traditional convolutional neural
neuralnetwork.
network.

The The convolutionallayer

convolutional layerincludes
includes convolution
convolution kernel,
kernel,convolutional
convolutionallayerlayerparame-
parameters,
ters, and activation function. The main function is to extract features from
and activation function. The main function is to extract features from the input data. the input data. Each
Each element in convolution kernel has its own weight coefficient and bias. The convo-
element in convolution kernel has its own weight coefficient and bias. The convolution
lution kernel regularly scans the input features, and performs matrix element multipli-
kernel regularly scans the input features, and performs matrix element multiplication and
cation and summation on the input features in the receptive field and superimposes the
summation on the input features in the receptive field and superimposes the deviation [25],
deviation [25], the core step is to solve the error term. To solve the error term, we first
theneed
coretostep is to which
analyze solve node
the error
needsterm. To solve the
to be calculated anderror
whichterm,
node we first need
or nodes in theto analyze
next
which
layernode needs to
are related, be calculated
because the nodeand which
affects the node or nodes
final output in the
result next the
through layer are related,
neurons
because the node
connected affects
with the nodethe final
in the output
next layer, result
which through
also needstheto neurons connected be-
save the relationship with the
node in the
tween next
each layer,
layer nodewhich
and the also needs
node in thetoprevious
save thelayer.
relationship between each layer node
and the node in the previous layer.
(
Mn f f
F = [ An ⊗ wn+1 ](i, j) = ∑m n n +1
=1 ∑ x =1 ∑y=1 [ Am ( t ∗ i + x, t ∗ j + y ) wm ( x, y )]
n + 1
(1)
A (i, j) = F (i, j) + d

Nn + 2P − f
Nn+1 = +1 (2)
t
Sensors 2021, 21, 418 4 of 20

In the formula, i, j are pixel indexes, d is the bias in the calculation process, W is the
weight matrix, and An , An+1 represent the input and output of the n+1 layer, Nn+1 is the
size of A, M is the number of convolution channels, t is the step size, and p and f are
the padding and convolution kernel size. The activation function is usually used after the
convolution kernel, in order to make the model fit the training data better and accelerate
the convergence, avoid the problem of gradient vanishing, in this article, ReLU is selected
as the activation function, as follows:

f ( x ) = max (0, x ) (3)

x is the input value of the upper neural network. the convolutional layer performs
feature extraction, and the obtained information is used as the input of the pooling layer.
The pooling layer can further filter the information, which not only reduces the dimension
of the feature, but also prevents the risk of overfitting. Pool layer generally has average
pool and maximum pool. The expression for the pooling layer is as follows [26]:
1
f f n
(i, j) = [∑ x=1 ∑y=1 Z (t ∗ i + x, t ∗ j + y)s ]
n s
Zm (4)
m

where t is the step size, pixel (i, j) are the same as convolutional layer, s is a specified pa-
rameter. When s = 1, the expression is average pooling, and when s → ∞ , the expression
is maximum pooling. m is the number of channels in the feature map, Z is the output
of pooling layer, and the value of s determines whether the output is average pooling or
maximum pooling. The other variables have the same meaning as convolution.
After feature extraction and dimensionality reduction of convolutional layer and pool-
ing layer, the fully connected layer maps these features to the labeled sample space. After
smoothing, the fully connected layer transforms the feature matrix into one-dimensional
feature vector. The full connection layer also contains the weight matrix and parameters.
The expression of the full connection layer is as follows:

Y = σ (W I + b ) (5)

where I is the input of the fully connected layer, Y is the output, W is the weight matrix,
and b is the bias. σ() is a general term for activation function, which can be softmax,
sigmoid, and ReLU, and the common ones are the multi-class softmax function and the
two-class sigmoid function.

2.2. Fully Convolutional Neural Network

In 2015, Long et al. [27] proposed the concept of a fully convolutional neural net-
work, realized the improved segmentation of the PASPA VOC data set (PASCAL Visual
Object Classes), and confirmed the feasibility of converting a fully connected layer into a
convolutional layer. The difference between fully connected layer and the convolutional
layer is that the connections of the neurons in the convolutional layer are only related
to a local input region, and the neurons in the convolutional layer can share parameters.
This greatly reduces the parameters in the network, improves the calculation efficiency of
neural network, and reduces the storage cost. The structure of the full convolution
Sensors 2021, 21, x FOR PEER REVIEW 5 of 21 model
is shown in Figure 2.

Figure 2. 2.The
Figure Thestructure of full
structure of fullconvolution
convolution model.
model.

2.3. LSTM Neural Network

LSTM (long short term memory) is an improved recurrent neural network (RNN)
[28]. LSTM is usually used to deal with time series data. The emergence of LSTM solves
the problem of gradient disappearance and gradient explosion in the long-term training
process. The output structure of the traditional RNN is composed of bias, activation
Sensors 2021, 21, 418 5 of 20

Figure 2. The structure of full convolution model.

2.3. LSTM Neural Network
2.3. LSTM Neural Network
LSTM (long short term memory) is an improved recurrent neural network (RNN) [28].
LSTM (long short term memory) is an improved recurrent neural network (RNN)
LSTM is usually used to deal with time series data. The emergence of LSTM solves the
[28]. LSTM is usually used to deal with time series data. The emergence of LSTM solves
problem of gradient disappearance and gradient explosion in the long-term training process.
the problem of gradient disappearance and gradient explosion in the long-term training
The output structure
process. The of the traditional
output structure ofRNN is composed
the traditional RNNof bias, activation
is composed function,
of bias, and
activation
weight, and the parameters of each time segment are the same. However, LSTM
function, and weight, and the parameters of each time segment are the same. However, introduces
a "gate" mechanism to control
LSTM introduces themechanism
a "gate" circulationtoand loss the
control of features.
circulation The
andwhole
loss ofLSTM canThe
features. be
regarded whole
as a memory
LSTM can block, in which
be regarded as athere are three
memory block,“Gates”
in which(input gate,
there are forget
three gate(input
“Gates” and
output gate)
gate,and a memory
forget unit. Ingate)
gate and output LSTM, and the order of
a memory importance
unit. In LSTM, the of gate
orderisofforget gate,
importance
of and
input gate, gate output
is forgetgate.
gate,The
input gate,structure
chain and output gate. The
of LSTM unitchain structure
is shown of LSTM
in Figure 3. unit is
shown in Figure 3.

Figure 3. The cell3.structure

Figure of long of
The cell structure short-term memory
long short-term (LSTM).
memory (LSTM).

(1) (1) Forget gate:

Forget gate:

f =  (W [a , x ] + d ) (6)
f t = tσ (W f · [fat−1 t,−x1 t ] +
t
d f )f (6)
where 𝑓𝑡 is the forget gate, which means that some features of 𝐶𝑡−1 are used in the cal-
where f t is the forget gate, which means that some features of Ct−1 are used in the calcula-
culation of 𝐶𝑡 . The value range of elements in 𝑓𝑡 is between [0, 1], while the activation
tion of Ct . The value range of elements in f is between [0, 1], while the activation function
function is generally sigmoid, 𝑊𝑓 is tthe weight matrix of forgetting gate, 𝑑𝑓 is the bias,
is generally sigmoid, W f is the weight matrix of forgetting gate, d f is the bias, and ⊗ is the
and ⊗ is the gate mechanism, which represents the relational operation of bit multipli-
gate mechanism,
cation. which represents the relational operation of bit multiplication.
(2) (2) and
Input gate Input gate andunit
memory memory unit update
update
ut =𝑢 σ=(Wu ·[ at∙− 1 , xt ] + du ) (7)
𝑡 𝜎(𝑊 𝑢 [𝑎𝑡−1 , 𝑥𝑡 ] + 𝑑𝑢 ) (7)

̃ = 𝑡𝑎𝑛ℎ(𝑊𝐶 ∙ [𝑎𝑡−1 , 𝑥𝑡 ] + 𝑑𝐶 ) (8)

Cet =𝐶𝑡tanh (WC ·[ at−1 , xt ] + dC ) (8)
𝐶𝑡 = 𝑓𝑡 ∘ 𝐶𝑡−1 + 𝑢𝑡 ∘ 𝐶̃𝑡 (9)
Ct = f t ◦ Ct−1 + ut ◦ Cet (9)
Among them, Ct is the current state of the unit, Ct−1 is the last state of the unit, and
Cet represents the update state of the unit, which is obtained from the input data xt and
at−1 through the neural network layer. The activation function of the unit state update
is tanh. Wu is the weight matrix of the input gate and Wc is the weight matrix of the
output gate. du and dc are bias of the input gate and bias of the output gate, respectively.
ut is the input gate, and the element value range is [0, 1], which is also calculated by the
sigmoid function.
(3) Output gate
ot = σ (Wo ·[ at−1 , xt ] + do ) (10)

at = ot ◦ tanh(Ct ) (11)
Sensors 2021, 21, 418 6 of 20

Among them, at is derived from the output gate ot and the cell state Ct , and the
average value of do is initialized to 1.

2.4. Evaluating Metrics

In order to better evaluate the prediction effect of the model, this article adopts two
Sensors 2021, 21, x FOR PEER REVIEWcurrently
popular metrics for evaluating the RUL prediction of turbofan engines:7 ofroot 21

mean square error (RMSE) and scoring function (SF), as shown in Figure 4.

Figure 4. The curve of root mean square error (RMSE) and scoring function (SF).
Figure 4. The curve of root mean square error (RMSE) and scoring function (SF).
RMSE: it is used to measure the deviation between the observed value and the actual
3.value.
1-FCLCNN-LSTM
It is a commonPrediction
evaluationModeindex for error prediction; RMSE has the same penalty
3.1.
for Overall Frameworks
early prediction and late prediction in terms of prediction. The calculation formula is
as follows:
The life prediction model 1-FCLCNN-LSTM s proposed in this paper adopts the idea
of classification and parallel processing; 1-FCLCNN 1 X and LSTM network extract spa-
RMSE =
tio-temporal features separately, and the two X types
n
∑i=ofn1 feature
Yi 2
are fused and then input to
(12)

the one-dimensional convolutional neural network and fully connected layer. Specifi-
where Xn is the total number of test samples of turbofan engine; Yi refers to the difference
cally, first, by preprocessing the C-MAPSS data set, the data are standardized and di-
between the predicted value pv of RUL of the i-th turbofan engine and the actual value av
vided
of RUL.into two input parts: INP1 and INP2. These two parts are input to the 1-FCLCNN
and theSF:LSTM
SF is neural network.
structurally Among them,
an asymmetric the 1-FCLCNN
function, which isisdifferent
used to extract the spatial
from RMSE. It is
feature of the data set. At the same time, the LSTM is adopted to extract
more inclined to early prediction (in this stage, RUL prediction value pv is less than the time series
the
feature of the av)
actual value datarather
set. After
than the
late feature extraction,
prediction to avoidthe algorithm
serious 1 is applied
consequences duetotofuse the
delayed
two types ofThe
prediction. feature
lowerandthe then
valueinputs
of RMSEthem andintoSF one-dimensional
score, the better theconvolutional
effect of the neural
model.
network. Finally, the data through
The calculation formula is as follows: the pooling layer and fully connected layer, is delt
with algorithm 2, and the predicted ( RUL result is obtained. The overall framework of the
x Yi
model is shown in Figure 5. ∑i=n 1 e( 13 ) − 1, Y i < 0;
score = x Yi (13)
∑i=n 1 e( 10 ) − 1, Y i ≥ 0;
In the formula, score represents the score. When RMSE, score, and Yi are as small as
possible, the effect of the model will be better.

3. 1-FCLCNN-LSTM Prediction Mode

3.1. Overall Frameworks
The life prediction model 1-FCLCNN-LSTM proposed in this paper adopts the idea
of classification and parallel processing; 1-FCLCNN and LSTM network extract spatio-
temporal features separately, and the two types of feature are fused and then input to the
one-dimensional convolutional neural network and fully connected layer. Specifically, first,
by preprocessing the C-MAPSS data set, the data are standardized and divided into two
input parts: INP1 and INP2. These two parts are input to the 1-FCLCNN and the LSTM
neural network. Among them, the 1-FCLCNN is used to extract the spatial feature of the

Figure 5. The overall framework of the model.

The feature fusion method, which only splices the feature information without
 The life prediction model 1-FCLCNN-LSTM proposed in this paper adopts the idea
of classification and parallel processing; 1-FCLCNN and LSTM network extract spa-
tio-temporal features separately, and the two types of feature are fused and then input to
the one-dimensional convolutional neural network and fully connected layer. Specifi-
Sensors 2021, 21, 418
cally, first, by preprocessing the C-MAPSS data set, the data are standardized and di-
7 of 20
vided into two input parts: INP1 and INP2. These two parts are input to the 1-FCLCNN
and the LSTM neural network. Among them, the 1-FCLCNN is used to extract the spatial
feature of the data set. At the same time, the LSTM is adopted to extract the time series
data set.ofAtthe
feature thedata
same time,
set. thethe
After LSTM is adopted
feature to extract
extraction, the time series
the algorithm feature to
1 is applied of fuse
the data
the
set. After
two typestheof feature
feature extraction, the algorithm
and then inputs 1 is applied
them into to fuse theconvolutional
one-dimensional two types of feature
neural
and then Finally,
network. inputs them into through
the data one-dimensional convolutional
the pooling layer and fullyneural network.layer,
connected Finally, the
is delt
with algorithm 2, and the predicted RUL result is obtained. The overall framework of the
data through the pooling layer and fully connected layer, is delt with algorithm 2, and
predicted
model RUL result
is shown is obtained.
in Figure 5. The overall framework of the model is shown in Figure 5.

Figure 5.
Figure The overall
5. The overall framework
framework of
of the
the model.
model.

The feature fusion method, which only splices the feature information without chang-
The feature fusion method, which only splices the feature information without
ing the content, preserves the integrity of the feature information, and can be used for
changing the content, preserves the integrity of the feature information, and can be used
multi-feature information fusion. Specifically, in the feature dimension splicing, the number
for multi-feature information fusion. Specifically, in the feature dimension splicing, the
of channels (features) increases after the splicing, but the information under each feature
number of channels (features) increases after the splicing, but the information under each
does not increase, and each channel of the splicing maps the corresponding convolution
kernel, the expression is as follows:

A = { Ai |i = 1, 2, 3, · · · , channel } (14)

B = { Bi |i = 1, 2, 3, · · · , channel } (15)

channel channel
Dsin gle = ∑ i =1 A i ∗ Ki + ∑ i =1 Bi ∗ Ki+channel (16)
In the formula, the two input channels are A and B respectively, the single output
channel to be spliced is Dsingle , ∗ means convolution, and K is the convolution kernel.
The algorithm of the spatio-temporal feature fusion and RUL prediction in this paper are
as follows.
Algorithm 1 Spatio-temporal Feature Fusion Algorithm
Input: INP1, INP2
Output: Spatio-temporal fusion feature
1. Conduct regular processing.
2. Keep the chronological order of the entire sequence, and use one-dimensional convolutional
layer to extract local spatial features to obtain spatial information.
3. Extract local spatial extreme values by one-dimensional pooling layer to obtain a
multi-dimensional spatiotemporal feature map
4. Learn the characteristics of the data sequence over time through LSTM.
5. Splice the above two features to get the final spatiotemporal fusion feature.
Sensors 2021, 21, 418 8 of 20

Algorithm 2 RUL Prediction Algorithm

Input: fusion feature
Output:RUL
1. Input the fusion feature into one-dimensional convolutional layer for further feature
extraction. The convolution kernel contains m filters of size n, each filter learns a single
feature, and each column of the output matrix contains a filter weight.
2. Set the maximum pooling layer size to reduce the complexity of output and prevent data
from overfitting.
3. Make the multi-dimensional input into one dimension for the transition from the
convolutional pooling layer to the fully connected layer.
4. Accelerate the convergence by ReLU function, at this time, each neuron in the first layer is
fully connected with the upper node.
5. Reduce the joint adaptability of neurons and improve the generalization ability by the
method of the parameter regularization Dropout.
6. Map the learned features to the labeled sample space by the non-linear mapping ability of
the fully connected layer network, and the spatio-temporal features are matched with the
actual life value to obtain the RUL prediction value.

3.2. Model Settings

Compared with multivariate data points composed of single time step sampling,
time series data provides more information. In the paper, the sliding window strategy
and multivariable time information are adopted to meet the data input requirements
Sensors 2021, 21, x FOR PEER REVIEW 9 of 21 of
the two network paths. The entire model structure is composed of 1-FCLCNN network,
LSTM network, and fully connected layer network. Next are the specific settings of
eachLSTM
component.
network, and fully connected layer network. Next are the specific settings of each
component.
3.2.1. 1-FCLCNN Network
3.2.1. 1-FCLCNN
The input of theNetwork
1-FCLCNN path is INP1. There are three 1D-CNN layers, which are
The input
used to extract theofspatial
the 1-FCLCNN
features.path
Theisstacked
INP1. There
CNN are[29]
three 1D-CNN
layers layers, which
are parsed are max
by three
usedlayers.
pooling to extract
Thethe spatial
first 1D-CNN features. Theconsist
layers stackedofCNN [29] layers
128 filters are parsed
( f ilters = 128 by three
× 1). Themax
second
pooling
consist of 64 layers.
filters (The first= 1D-CNN
f ilters layers
64 × 1), and theconsist of 128offilters
third consist (𝑓𝑖𝑙𝑡𝑒𝑟𝑠
32 filters = 128
( f ilters =×321).The
× 1). The
second consist of 64 filters (𝑓𝑖𝑙𝑡𝑒𝑟𝑠 = 64 × 1), and the third consist of 32 filters (𝑓𝑖𝑙𝑡𝑒𝑟𝑠 =
convolution kernel size of the three 1D-CNN layers is the same (kernel size = 3). “ReLU”
32 × 1 ).The convolution kernel size of the three 1D-CNN layers is the same
activation function(formula 3) is used for the 1D-CNN layers. The Settings of the max
(𝑘𝑒𝑟𝑛𝑒𝑙 𝑠𝑖𝑧𝑒 = 3). “ReLU” activation function(formula 3) is used for the 1D-CNN layers.
pooling layers: pool_size = 2, padding = ”same”, strides = 2. The batch normalization
The Settings of the max pooling layers: 𝑝𝑜𝑜𝑙_𝑠𝑖𝑧𝑒 = 2, 𝑝𝑎𝑑𝑑𝑖𝑛𝑔 = "𝑠𝑎𝑚𝑒", 𝑠𝑡𝑟𝑖𝑑𝑒𝑠 = 2.
speeds convergence and controls
The batch normalization speedsoverfitting
convergenceafter each
and max pooling
controls layer.
overfitting The
after 1-FCLCNN
each max
pathpooling
can be layer.
used Thewith1-FCLCNN
or without path can be used with or without dropout, makingtothe
dropout, making the network less sensitive initial
weights. The detailed architecture is shown in Figure 6.
network less sensitive to initial weights. The detailed architecture is shown in Figure 6.

Figure 6. The
Figure detailed
6. The architecture
detailed architectureof
ofthe
the 1-FCLCNN path.
1-FCLCNN path.

3.2.2. LSTM Network

3.2.2. LSTM
ThisNetwork
part is composed of three LSTM layers, which are used to extract time series
This part
features is composed
of turbofan of The
engines. three LSTM
input layers,
of the LSTMwhich
path isare used
INP2. Thetofirst
extract
LSTMtime series
is de-
fined by 128 cell structures, while the second and the third LSTM consist
features of turbofan engines. The input of the LSTM path is INP2. The first LSTM of 64 and 32 cell is
structures
defined by 128 separately. In the LSTM
cell structures, whilepath,
the the resultand
second of each
the hidden layer isconsist
third LSTM the input
of of
64the
and 32
Meanwhile, 𝑟𝑒𝑡𝑢𝑟𝑛_𝑠𝑒𝑞𝑢𝑒𝑛𝑐𝑒𝑠 = 𝑇𝑟𝑢𝑒
cell structures separately. In the LSTM path, the result of each hidden layer is theofinput
next layer. indicates that the hidden state
output
of the nextcontains the results ofreturn_sequences
layer. Meanwhile, total time steps. = True indicates that the hidden state of
output contains the results of total time steps.
3.2.3. Fully Connected Layer
The output of the 1-FCLCNN path is concatenated with the output features of the
LSTM path. The resultant vector will be applied as input to the fully connected path. The
fully connected path comprises of a convolutional layer, a pooling layer, a flatten layer,
and three full connection layers. The convolutional layer consists of 256 filters𝑓𝑖𝑙𝑡𝑒𝑟𝑠 =
256 × 1, the parameter setting of the max pooling layer is consistent with 1-FCLCNN
Sensors 2021, 21, 418 9 of 20

3.2.3. Fully Connected Layer

The output of the 1-FCLCNN path is concatenated with the output features of the
LSTM path. The resultant vector will be applied as input to the fully connected path. The
fully connected path comprises of a convolutional layer, a pooling layer, a flatten layer, and
three full connection layers. The convolutional layer consists of 256 filters f ilters = 256 × 1,
the parameter setting of the max pooling layer is consistent with 1-FCLCNN network.
Then the data are flattened by flatten layer (transform multidimensional data into one-
dimensional data, which are commonly used in the transition from convolutional layer
to full connected layer). The first two full connected layers consist of 128 and 32 neurons
separately and “ReLU” activation function is utilized. The third layer has 1 neuron as the
output to estimate the RUL.

3.3. The Process of Model Training

In this paper, the FD001 and FD003 data sets are used. The difference of sensor and
engine will cause different physical characteristic state result, therefore, in order to improve
the accuracy of the model and the convergence speed of the model, the original monitoring
status data are processed by the normalization method. The model limits the data size to
between [0, 1].

Xm,n − Xnmin
X ∗ m,n = , ∀m, n (17)
Xnmax − Xnmin
∗ represents a value of the mth data point of the nth feature after
In the formula, Xm,n
normalization processing. Xm,n represents the initial data before processing. Xnmax , Xnmin
are the maximum and minimum values of the features respectively.
In the model training section, the purpose of training is to minimize the cost function
and the loss, and to obtain the best parameters. The cost function as RMSE is defined by
the model (Formula (12)). In the meantime, Adam algorithm [30] and Early stopping [31]
are adopted to optimize the training process. The Early stopping can not only verify the
effect of the model in the training process, but also avoid overfitting. During the training,
the normalized data are segmented by sliding window. The input data INP1 and INP2 are
in the form of two-dimensional tensors with the size of ssw × n f , which are processed in
parallel paths separately, they are 1D convolutional layer and LSTM network. In order to
make the gradient larger and reduce the gradient disappearance problem [32], normalized
operation is used after each max pooling layer. In addition, the normalized operation
normalizes the activation values of each layer of the neural network to maintain the same
distribution. In the meantime, a larger gradient means an increase in convergence and
training speed.
The 1-FCLCNN-LSTM training algorithm is as follows:

Algorithm 3 1-FCL CNN -LSTM training algorithm

Input: C-MAPSS dataset(FD001, FD003)
Output: 1-FCLCNN -LSTM model based on weight determination.
1. Feature selection of data and standardization of data.
2. Prepare training set and test set; normalization processing of training set and test set
3. Calculate the test set label.
4. Extract input data of the network; Use a sliding window to split the data (Define the sliding
window function to extract features).
5. Use sliding window to extract training set label and test set label.
6. Use Adam optimization algorithm to update weights. When the number of training periods
of 1-FCLCNN -LSTM model is less than the set value or does not reach Early Stopping
condition, the extracted features will be input into 1-FCLCNN -LSTM for forward
propagation and the predicted output will be obtained. Calculate the error between the
predicted value and the actual value.
7. Obtain the trained 1-FCLCNN-LSTM model.
 Sensors 2021, 21, 418 10 of 20

ors 2021, 21, x FOR PEER REVIEW 11 of 21

The overall flow of the proposed method in this paper is shown in Figure 7:

Figure
Figure 7. The flow chart The flow method.
7.proposed
of the chart of the proposed method.

4. Experiments and Analysis

4. Experiments and Analysis
First, the C-MAPSS data set is introduced in detail, second, preprocesses the data,
First, the C-MAPSS data set is introduced in detail, second, preprocesses the data,
test and verify the proposed prediction model on the data set. Then parameter settings
test and verify the proposed prediction model on the data set. Then parameter settings
are adjusted through the training model. Finally, compares the experimental results with
are adjusted through the training model. Finally, compares the experimental results with
other methods.
other methods.
4.1. C-MAPSS Data Set
In this paper, NASA C-MAPSS turbofan engine degradation data set
(https://ti.arc.nasa.gov/tech/dash/groups/pcoe/prognostic-data-repository/) [33] was
used, which was derived from C-MAPSS database created by NASA Army Research Lab.
The main control system consists of fan controller, regulator, and limiter. Fan controls the
Sensors 2021, 21, 418 11 of 20

4.1. C-MAPSS Data Set

In this paper, NASA C-MAPSS turbofan engine degradation data set (https://ti.arc.
Sensors 2021, 21, x FOR PEER REVIEW 12 of 21
nasa.gov/tech/dash/groups/pcoe/prognostic-data-repository/) [33] was used, which
was derived from C-MAPSS database created by NASA Army Research Lab. The main
control system consists of fan controller, regulator, and limiter. Fan controls the normal
normal operation
operation of the
of the flight flight conditions,
conditions, sendingsending air into
air into the the
inner inner
and and
outer outer culverts,
culverts, as shown as
in Figure 8. A low pressure compressor (LPC) and high pressure
shown in Figure 8. A low pressure compressor (LPC) and high pressure compressor compressor (HPC)
supply compressed
(HPC) supply high temperature,
compressed high pressure
high temperature, gases to the
high pressure combustor.
gases Low pressure
to the combustor. Low
turbine
pressure(LPT) can (LPT)
turbine decelerate and pressurize
can decelerate air to improve
and pressurize theimprove
air to chemical energy
the conversion
chemical energy
efficiency
conversion ofefficiency
aviation kerosene. High
of aviation pressureHigh
kerosene. turbines (HPT)turbines
pressure generate(HPT)
mechanical
generateenergy
me-
by using energy
chanical high temperature
by using highandtemperature
high pressure
andgas strike
high turbine
pressure gasblades. Low-pressure
strike turbine blades.
rotor (N1), high-pressure
Low-pressure rotor (N2), and
rotor (N1), high-pressure nozzle
rotor (N2),guarantee
and nozzle theguarantee
combustion theefficiency
combustion of
the engine.
efficiency of the engine.

Figure 8. C-MAPSS turbofan engine diagram [34].

Figure 8. C-MAPSS turbofan engine diagram [34].
The C-MAPSS database contains four subsets of data (FD001-FD004) generated from
The C-MAPSS database contains four subsets of data (FD001-FD004) generated from
different time series and including cumulative spatial complexity. Each data subset includes
different time series and including cumulative spatial complexity. Each data subset in-
a test data set and a training data set and the number of engines varies in each data subset.
cludes a test data set and a training data set and the number of engines varies in each data
Each engine has varying degrees of initial wear-and-tear, and this kind of wear-and-tear is
subset. Each engine has varying degrees of initial wear-and-tear, and this kind of
considered normal. There are three operating settings that have a significant impact on
wear-and-tear
engine is considered
performance. The enginenormal.
worksThere are three
normally at theoperating
beginningsettings
of eachthat
timehave
seriesa and
sig-
nificant impact on engine performance. The engine works normally at the beginning
fails at the end of the time series. In the training set, the fault increases until the system fails of
each time series and fails at the end of the time series. In the training set,
and in the test set, the time series ends at some time before the system fails [35]. In each the fault in-
creases
time until21the
series, system
sensor fails and in
parameters andthe3 test
otherset, the time series
parameters showends at some state
the running time before
of the
the system fails [35]. In each time series, 21 sensor parameters and 3 other
turbofan engine. The data set is provided as a compressed text file. Each row is a snapshot parameters
show
of the the
datarunning state ofa single
taken during the turbofan engine.
operation The
cycle, anddata
eachsetcolumn
is provided as a compressed
is a different variable.
text file. Each row is a snapshot of the
The specific contents are shown in Tables 1 and 2.data taken during a single operation cycle, and
each column is a different variable. The specific contents are shown in Tables 1 and 2.
Table 1. Column contents of data set file.
Table 1. Column contents of data set file.
Serial NUMBER Variable Name
Serial NUMBER Variable Name
1 unit number
1 unit number
2 time, in cycles
32 time, in cycles
operational setting 1
43 operational
operational setting
setting 21
54 operational setting 32
operational setting
6 sensor measurement 1
5 operational setting 3
7 sensor measurement 2
86 sensor measurement
... 1
97 sensor
sensormeasurement
measurement262
8 …
9 sensor measurement26
Sensors 2021, 21, 418 12 of 20

Table 2. Data description of turbofan engine sensor.

Sensor Number Sensor Description Units

1 (Fan inlet temperature) (◦ R)
2 (LPC outlet temperature) (◦ R)
3 (HPC outlet temperature) (◦ R)
4 (LPT outlet temperature) (◦ R)
5 (Fan inlet Pressure) (psia)
6 (bypass-duct pressure) (psia)
7 (HPC outlet pressure) (psia)
8 (Physical fan speed) (rpm)
9 (Physical core speed) (rpm)
(Engine pressure ratio
10 ——
(P50/P2)
11 (HPC outlet Static pressure) (psia)
12 (Ratio of fuel flow to Ps30) (pps/psia)
13 (Corrected fan speed) (rpm)
14 (Corrected core speed) (rpm)
15 (Bypass Ratio) ——
16 (Burner fuel-air ratio) ——
17 (Bleed Enthalpy) ——
18 (Required fan speed) (rpm)
(Required fan conversion
19 (rpm)
speed)
(High-pressure turbines Cool
20 (lb/s)
air flow)
(Low-pressure turbines Cool
21 (lb/s)
air flow)

According to the needs of experiment, this paper adopts the data set FD001 and FD003
for model verification, and the specific description of the data set is shown in Table 3:

Table 3. Data sets FD001 and FD003 are detailed.

Type of Operating
Data Set Training Set Test Set Operating Conditions Fault Mode Number of Sensors
Parameters
FD001 100 100 1 1 21 3
FD003 100 100 1 2 21 3

In this table, the training set in the data set includes the data of the entire engine life
cycle, while the data trajectory of the test set terminates at some point before failure. FD001
and FD003 were simulated under the same (sea level) condition. FD001 was only tested in
the case of HPC degradation, and FD003 was simulated in two fault modes: HPC and fan
degradation. The number of sensors and the type of operation parameters are consistent
for the four data subsets (FD001-FD004). The data subsets FD001 and FD003 contain
actual RUL values, so that the effect of the model can be seen according to the comparison
between the actual value and the predicted value. The result of the experiment is to predict
the number of remaining running cycles before the failure of the test set, namely RUL.

4.2. Data Preprocessing

In the training stage of the model, the original turbofan engine data should be pre-
processed, and the pre-processed data can be put into the model to obtain the parameters
required by the model. The pre-processing process includes feature selection, data stan-
dardization and normalization, setting the size of sliding window, and RUL label setting of
training set and test set. The FD001 dataset contains 21 sensor features and 3 operating
parameters (flight altitude, Mach number, and throttling parser Angle). The number of
running cycles is also one of the features, so with a total of 25 features. In order to ensure
Sensors 2021, 21, x FOR PEER REVIEW 14 of 21

Sensors 2021, 21, 418 13 of 20

3 operating parameters (flight altitude, Mach number, and throttling parser Angle). The
number of running cycles is also one of the features, so with a total of 25 features. In order
to ensure the consistency of the input and output of the model and the comparison effect
the consistency of the input and output of the model and the comparison effect of different
of different data sets, the feature selection of FD003 is consistent with FD001. Because
data sets, the feature selection of FD003 is consistent with FD001. Because multiple sensors
multiple sensors will produce multiple features, in order to eliminate the influence of
will produce multiple features, in order to eliminate the influence of different dimensions
different dimensions on the prediction results, the normalization method of formula (14)
on the prediction results, the normalization method of formula (14) is adopted. The input
is adopted. The input data are a 2D matrix containing 𝑠𝑠𝑤 (as the size of the sliding
data are a 2D matrix containing ssw (as the size of the sliding window) with n f (as the
window) with 𝑛𝑓 (as the number of the selected features). In order to keep the size of
number of the selected features). In order to keep the size of input and output of FD001
input and output of FD001 and FD003 data subsets constant and the data processing
and FD003 data subsets constant and the data processing process is accelerated. This paper
process is accelerated. This paper uses a larger window to get more detailed features. The
uses a larger window to get more detailed features. The sliding window and the number
sliding window and the number of features is set to 50 and 25 respectively.
of features is set to 50 and 25 respectively.
In the neural network model, we need to get the corresponding output according to
In the neural network model, we need to get the corresponding output according to
the
the inputdata.
input data.TheThestate
stateofofthe
thesystem
systematateacheachtime
timestep
stepandandthe thespecific
specificinformation
informationof of
the
the target RUL are based on the physical model and are difficult to determine.To
target RUL are based on the physical model and are difficult to determine. Tosolve
solve
this
thisproblem,
problem,different
differentsolutions
solutionshave havebeen
beenproposed.
proposed.One Onesolution
solutionisistotosimply
simplyallocate
allocate
the
therequired
requiredoutput
outputas asthe
theactual
actualtimetimeremaining
remainingbefore
beforeaa functional
functionalfailure,
failure,but butat atthe
the
same time the state of the system will decline linearly [36]. Another
same time the state of the system will decline linearly [36]. Another option is to obtainoption is to obtain the
desired output
the desired valuevalue
output basedbased
on the onappropriate
the appropriatedegradation
degradation model. Referring
model. to thetocur-
Referring the
rent literature, this paper adopts piece-wise linear degradation
current literature, this paper adopts piece-wise linear degradation model to determinemodel to determine the
target RUL RUL
the target [37–40]. Piece-wise
[37–40]. linearlinear
Piece-wise regression modelmodel
regression can prevent the algorithm
can prevent from
the algorithm
overestimating
from overestimatingRUL. ForRUL. anFor
engine, equipment
an engine, can becan
equipment considered
be consideredhealthy during
healthy its ini-
during its
tial period.
initial TheThe
period. degradation
degradationprocess willwill
process be obvious
be obviousbefore the whole
before the wholeequipment
equipment runsruns
for
afor
period of time
a period or isorused
of time by by
is used a certain
a certainextent,
extent,that
thatis,is,
near
nearthetheend
endofofthe thelife
lifeofofthe
the
equipment. Set the normal working state of the device to a constant
equipment. Set the normal working state of the device to a constant value, and the RUL value, and the RUL of
the device
of the willwill
device drop linearly
drop withwith
linearly timetime
afterafter
a certain period
a certain of time.
period ThisThis
of time. model limits
model the
limits
maximum
the maximum value of RUL,
value which
of RUL, whichis determined
is determined bybythetheobserved
observeddata dataininthe
thedegradation
degradation
stage.
stage.TheThemaximum
maximumRUL RULvalue
valueofofthethedata
dataset
setobserved
observedfrom fromthe thedegradation
degradationphase phaseof of
the
the experiment
experiment is setsettoto125,
125,andandthethe part
part exceeding
exceeding 125 125 is uniformly
is uniformly specified
specified as 125.asWhen
125.
When the critical
the critical period period is reached,
is reached, RUL decreases
RUL decreases linearly as linearly as theperiod
the running running period The
increases. in-
creases.
Piece-wiseTheLinear
Piece-wise
RUL Linear
Target RUL Target
Function Function
is shown is shown
in Figure 9. in Figure 9.

Figure 9. Piece-wise linear remaining useful life (RUL) target function.

Figure 9. Piece-wise linear remaining useful life (RUL) target function.
4.3. Parameter Settings
4.3. Parameter
Because Settings
the model needs to adjust parameters in the training process, it is important
Because
to select the the model needs
appropriate to adjust for
parameters parameters
the whole in the training process,
experiment. For twoit data
is important
subsets,
FD001 and FD003, each data subset is divided into training set accounting
to select the appropriate parameters for the whole experiment. For two data subsets, for 85% and
verification
FD001 set accounting
and FD003, forsubset
each data 15%. All data setsinto
is divided are trained
traininginset
theaccounting
mini-batchfor
method [35].
85% and
For the main
verification setparameters
accountinginvolved
for 15%. in
Allthe data
data setare
sets in this experiment,
trained this papermethod
in the mini-batch adopts
a one-by-one optimization method according to the most commonly used values of the
parameters. The selection of parameters included epoch (value: 40, 60, 80,100), batch size
2021, 21, x FOR PEER REVIEW 15 of 21
2021, 21, x FOR PEER REVIEW 15 of 21

Sensors 2021, 21, 418 14 of 20

[35]. For
[35]. For the
the main
main parameters
parameters involved
involved inin the
the data
data set
set in
in this
this experiment,
experiment, this this paper
paper
adopts a one-by-one optimization method according to the most
adopts a one-by-one optimization method according to the most commonly used values commonly used values
of the
of the parameters.
parameters. The
The selection
selection ofof parameters
parameters included
included epochepoch (value:
(value: 40,40, 60,
60, 80,100),
80,100),
batch size
batch size (value:
(value: 64,128,256,512),
(value: 64,128,256,512),
64,128,256,512), dropout raterate
dropout
dropout rate (value: 0.1,0.1,
(value:
(value: 0.1, 0.2,0.2,
0.2, 0.3,
0.3, 0.4).
0.3, InInthe
0.4).
0.4). In the course
thecourse ofof experimental
courseof
experimental training,
training, ififthe
thetraining
training error
error and
and verification
verification error
error do
do not
experimental training, if the training error and verification error do not decrease in thenot decrease
decrease ininthe
the five training
five training periods, the training shall be stopped in advance. The parameter
five training periods, the training shall be stopped in advance. The parameter results of of FD001 are
periods, the training shall be stopped in advance. The parameter results of
results
FD001 are
FD001 are shown
shown in Figure
shown
in Figure 10. The
in Figure
10. The parameter
10. The results
parameter
parameter of FD003
results
results of FD003
of FD003 are shown
are shown
are shownin Figure
in Figure 11. 11.
in Figure
11.

(a) BATCH SIZE and RMSE (b) EOPCH and RMSE

(a) BATCH SIZE and RMSE (b) EOPCH and RMSE

(c) DROPOUT and RMSE.

(c) DROPOUT and RMSE.
Figure 10. Experimental results
Figure 10. of different
Experimental parameters
results of FD001.
of different parameters of FD001.
Figure 10. Experimental results of different parameters of FD001.

(a) BATCH SIZE and RMSE. (b) EOPCH and RMSE

(a) BATCH SIZE and RMSE. (b) EOPCH and RMSE

(c) DROPOUT
(c) DROPOUT and RMSE.
and RMSE.
Figure 11. Experimental results of different parameters of FD003.
Figure 11. Experimental results
Figure 11. of different
Experimental parameters
results of FD003.
of different parameters of FD003.
After model
After model training
training and comparative
and
After model comparative analysis
training andanalysis of experimental
of
comparativeexperimental results, the
results, the param-
analysis of experimental param-
results, the parameter
eter setting of FD001
eter setting of FD001 and
settingand FD003 data
FD003 and
of FD001 subsets
data FD003 with
subsetsdata the best
withsubsets model
the bestwith
model performance
theperformance is finally
best model isperformance
finally is finally
obtained, as
obtained, as shown
shown in Table
obtained,
in Table 4.
as shown
4. in Table 4.
Sensors 2021, 21, 418 15 of 20

Sensors 2021, 21, x FOR PEER REVIEW 16 of 21

Table 4. Model parameter Settings for FD001 and FD003 data subsets.

Data Subset
FD001
Table 4. Model parameter Settings for FD001 and FD003 data subsets. FD003
Parameter
Data Subset
epoch 60 60
FD001 FD003
batchParameter
size 256 512
dropout
epoch 0.2 60 60 0.2
batch size 256 512
4.4. Experimental Resultsdropout
and Comparison 0.2 0.2

In this section, we mainly introduce the prediction results of this model and the
4.4. Experimental Results and Comparison
comparative analysis with the recent popular research methods. With the same data input
In this section, we mainly introduce the prediction results of this model and the
andcomparative
the same pretreatment process, the prediction results of the traditional convolutional
analysis with the recent popular research methods. With the same data in-
neural
put and the same pretreatmentwith
network are compared thethe
process, 1-FCLCNN-LSTM
prediction results of model proposed
the traditional in this paper.
convolu-
Thetional
traditional convolutional
neural network neuralwith
are compared network consists of two
the 1-FCLCNN-LSTM convolutional
model proposed inlayers,
this two
pooling
paper.layers, and a full
The traditional connectedneural
convolutional layer.network
For FD001 and
consists ofFD003 data subset,
two convolutional this paper
layers,
compares the training
two pooling effect
layers, and of convolutional
a full connected layer.neuralFor FD001network and FCLCNN-LSTM
and FD003 data subset, this model
paper
under thecompares
same datatheset
training effect of The
and engine. convolutional neural network
training effects of engines andwith
FCLCNN-LSTM
FD001 and FD003
datamodel under
subsets on the
thesame data set are
two models and shown
engine. inThe training12effects
Figures and 13. of engines with FD001
The training diagrams of
the and
twoFD003
models data
insubsets
a singleondata
the two models
subset canarebe shown
obtained in Figures 12 and
as follows: 13. The
RUL training
began to decrease
diagrams of the two models in a single data subset can be obtained as follows: RUL began
with the increase of time step, and finally failed. From the process of RUL reduction, it can
to decrease with the increase of time step, and finally failed. From the process of RUL
be observed that with the increase of time, the higher the prediction accuracy, the closer the
reduction, it can be observed that with the increase of time, the higher the prediction
predicted value and the
accuracy, the closer the actual thevalue
predicted values andare,
thewhich means
actual the that
values the
are, smaller
which RUL
means thatis closer
to the
the potential
smaller RUL fault. In this
is closer paper,
to the RMSE
potential is In
fault. used
this to express
paper, RMSE the training
is used effectthe
to express of FD001
andtraining
FD003effect
training sets, and
of FD001 as shown in Formula
FD003 training sets, (12). Thein
as shown comparison
Formula (12). results are shown in
The compar-
Table
ison5.results are shown in Table 5.

(a) Training diagram of CNN model (b) Training diagram of 1-FCLCNN-LSTM model
Sensors 2021, 21, x FOR PEER REVIEW 17 of 21
Figure
Figure 12. 12. Training
Training diagramof
diagram ofthe
the same
same FD001
FD001engine
engineunder twotwo
under models.
models.

(a) Training diagram of CNN model. (b) Training diagram of 1-FCLCNN-LSTM model.
Figure
Figure 13. Training
13. Training diagramofofthe
diagram thesame
same FD003
FD003engine
engineunder
undertwotwo
models.
models.
Table 5. RMSE training values of FD001 and FD003 under the two models.

FD001 FD003
CNN 8.25 14.00
1-FCLCNN-LSTM 4.87 7.56
Sensors 2021, 21, 418 16 of 20
(a) Training diagram of CNN model. (b) Training diagram of 1-FCLCNN-LSTM model.
Figure 13. Training diagram of the same FD003 engine under two models.
Table 5. RMSE training values of FD001 and FD003 under the two models.
Table 5. RMSE training values of FD001 and FD003 under the two models.
FD001 FD003
FD001 FD003
CNN 8.25 14.00
CNN 8.25 14.00
1-FCLCNN-LSTM 4.87 7.56
1-FCLCNN-LSTM 4.87 7.56

From
From Table
Table 55 and
and the
the training
training diagrams
diagrams of the two
of the two models
models onon different
different data
data sets,
sets, it
it
can
can be
be concluded
concluded that
that the 1-FCLCNN-LSTM proposed
the 1-FCLCNN-LSTM proposed in this paper
in this paper performs
performs better
better in
in
the
the training
training process than the
process than the traditional
traditional single
single CNN
CNN neural
neural network.
network. Among
Among them, the
them, the
RMSE of of 1-FCLCNN-LSTM
1-FCLCNN-LSTMmodel modelonon FD001 training set was 41% lower than
FD001 training set was 41% lower than that of CNN that of
CNN model, and the RMSE of 1-FCLCNN-LSTM model on FD003
model, and the RMSE of 1-FCLCNN-LSTM model on FD003 training set was 46% lower training set was 46%
lower thanofthat
than that CNN of model.
CNN model.
FD003FD003
has twohas twomodes
fault fault modes while FD001
while FD001 has onlyhasone,
only one,
which
which indicates
indicates that
that the the multi-neural
multi-neural networknetwork
modelmodel has certain
has certain advantages
advantages in dealing
in dealing with
with complex
complex fault fault problems.
problems.
The test
testsets
sets of FD001
of FD001 and FD003
and FD003 wereinto
were input input into the
the trained CNN trained CNN and
and 1-FCLCNN-
1-FCLCNN-LSTM
LSTM models to obtain modelsthe
to obtain the prediction
prediction results,
results, which arewhich
shownareinshown
Figuresin 14
Figures 14
and 15,
respectively.
and 15, respectively.

Sensors 2021, 21,(a) Prediction

x FOR results of CNN model.
PEER REVIEW (b) Prediction results of 1-FCLCNN-LSTM model. 18 of 21
Figure
Figure 14.
14. RUL
RUL prediction
prediction results
results of
of FD001 with different
FD001 with different models.
models.

(a) Prediction results of CNN model. (b) Prediction results of 1-FCLCNN-LSTM model.
Figure 15.
Figure 15. RUL
RUL prediction
prediction results
results of
of FD003
FD003 with
with different
differentmodels.
models.

In this paper,
paper, the
the RMSE
RMSE is is used
used to
to express
express the
the effects
effects of
of FD001
FD001 and
and FD003
FD003 test
test sets,
sets, as
as
shown in Formula (12).
(12). See Table 6 for details.
Table 6 for details.

Table 6. RMSE predicted by FD001 and FD003 under the two models.

FD001 FD003
CNN 17.22 15.50
FCLCNN-LSTM 11.17 9.99
Sensors 2021, 21, 418 17 of 20

Table 6. RMSE predicted by FD001 and FD003 under the two models.

FD001 FD003
CNN 17.22 15.50
FCLCNN-LSTM 11.17 9.99

It can be seen from Table 6 that the training effect of the model directly affects the
performance of the test set of the model. As shown in the above table, the RMSE of 1-
FCLCNN-LSTM model on FD001 test set is 35% lower than that of CNN model and the
RMSE of 1-FCLCNN-LSTM model on FD003 is 35.5% lower than that of CNN model.
In order to measure the prediction performance of the model more comprehensively,
this paper selects the latest advanced RUL prediction method, and compares the deviations
of various methods under the same data set. The evaluation indicators are RMSE and the
score function, both of which are as low as possible. The comparison results of FD001 data
set are shown in Table 7, and the comparison results of FD003 data set are shown in Table 8.

Table 7. Comparison results of various models on the FD001 dataset.

Models RMSE Score

1-FCLCNN-LSTM 11.17 204
HDNN [29] 13.02 245
DCNN [20] 12.61 274
LSTM-FNN [41] 16.14 338
LSTMBS [42] 13.27 216
Autoencoder-BLSTM [23] 13.63 261
VAE-D2GAN [40] 11.60 221
D-LSTM [43] 16.14 338
RF [44] 17.91 480
BiRNN-ED [45] 14.74 273

Table 8. Comparison results of various models on the FD003 dataset.

Models RMSE Score

1-FCLCNN-LSTM 9.99 234
HDNN [29] 12.22 288
DCNN [20] 12.64 284
LSTMBS [42] 16.00 317
D-LSTM [43] 16.18 852
RF [44] 20.27 711
SVM [44] 46.32 22542
Rulclipper [45] 16.00 317
FADCNN [46] 19.82 1596
GB [44] 16.84 577

The comparison results with multiple models show that the model proposed in
this paper has the lowest score and RMSE values on both FD001 and FD003 data sets.
The RMSE of 1-FCLCNN-LSTM model on FD001 was 11.4–36.6% lower than that of RF,
DCNN, D-LSTM, and other traditional methods, and the RMSE of 1-FCLCNN-LSTM
model on FD003 was 37.5–78% lower than that of GB, SVM, LSTMBS, and other traditional
methods. The above results are attributed to the multi-neural network structure and
parallel processing of feature information in this model, which can effectively extract
RUL information. Compared with the current popular multi-model Autoencoder-BLSTM,
VAE-D2GAN, HDNN, and other methods, the RMSE of FD001 was decreased by 4–18%,
and the RMSE of FD003 was decreased by 18–37.5% compared with that on HDNN,
DCNN, Rulclipper, and other methods. The above results are attributed to the same multi-
model structure and multi-network structure, the 1-FCLCNN-LSTM model has advantages
Sensors 2021, 21, 418 18 of 20

in feature processing in the1-FCLCNN path, and the fused data are processed by the 1D
full-convolutional layer to obtain more accurate prediction results. The score of 1-FCLCNN-
LSTM model in FD001 was 5% lower than the optimal LSTMBS in the previous model. The
score of 1-FCLCNN-LSTM model in FD003 was 17.6% lower than the optimal DNN in the
previous mode. This indicates that the prediction accuracy of this model in C-MAPSS data
set is improved, and no expert knowledge or physical knowledge is required, which can
help maintain predictive turbofan engines, as a research direction of mechanical equipment
health management.

5. Conclusions
This paper has presented a method for RUL prognosis of spatiotemporal feature
fusion modeled with 1-FCLCNN and LSTM network. From the current data sets which are
issued from some location and state sensors, the proposed method extracts spatiotemporal
feature and estimates the trend of the future remaining useful life of the turbofan engine. In
addition, for different data sets and prognosis horizon, it is shown that the RUL prognosis
error is superior to other methods. Future research will improve the use of the approach for
online applications. The main challenge is to incrementally update the prognosis results.
The use of the approach for other real applications and large machinery systems will also
be considered.

Author Contributions: Conceptualization, C.P. and Z.T.; methodology, W.G. and C.P.; software, Y.C.;
validation, Y.C., Q.C.; formal analysis, C.P. and L.L.; investigation, C.P.; resources, W.G.; writing—
original draft preparation, C.P. and Y.C.; writing—review and editing, C.P.; visualization, Y.C.;
supervision, C.P.; project administration, C.P.; funding acquisition, Z.T. and W.G. All authors have
read and agreed to the published version of the manuscript.
Funding: This work is supported by Natural Science Foundation of China (No. 61871432, No.
61771492), the Natural Science Foundation of Hunan Province (No.2020JJ4275, No.2019JJ6008, and
No.2019JJ60054).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Wei, J.; Bai, P.; Qin, D.T.; Lim, T.C.; Yang, P.W.; Zhang, H. Study on vibration characteristics of fan shaft of geared turbofan engine
with sudden imbalance caused by blade off. J. Vib. Acoust. 2018, 140, 1–14. [CrossRef]
2. Tuzcu, H.; Hret, Y.; Caliskan, H. Energy, environment and enviroeconomic analyses and assessments of the turbofan engine used
in aviation industry. Environ. Prog. Sustain. Energy 2020, 3, e13547. [CrossRef]
3. You, Y.Q.; Sun, J.B.; Ge, B.; Zhao, D.; Jiang, J. A data-driven M2 approach for evidential network structure learning. Knowl. Based
Syst. 2020, 187, 104800–104810. [CrossRef]
4. De Oliveira da Costa, P.R.; Akcay, A.; Zhang, Y.Q.; Kaymak, U. Attention and long short-term memory network for remaining
useful lifetime predictions of turbofan engine degradation. Int. J. Progn. Health Manag. 2020, 10, 34.
5. Ghorbani, S.; Salahshoor, K. Estimating remaining useful life of turbofan engine using data-level fusion and feature-level fusion.
J. Fail. Anal. Prev. 2020, 20, 323–332. [CrossRef]
6. Sun, H.; Guo, Y.Q.; Zhao, W.L. Fault detection for aircraft turbofan engine using a modified moving window KPCA. IEEE Access
2020, 8, 166541–166552. [CrossRef]
7. Ahmadzadeh, F.; Lundberg, J. Remaining useful life estimation: Review. Int. J. Syst. Assur. Eng. Manag. 2014, 5, 461–474.
[CrossRef]
8. Kok, C.; Jahmunah, V.; Shu, L.O.; Acharya, U.R. Automated prediction of sepsis using temporal convolutional network. Comput.
Biol. Med. 2020, 127, 103957. [CrossRef]
9. Cheong, K.H.; Poeschmann, S.; Lai, J.; Koh, J.M. Practical automated video analytics for crowd monitoring and counting. IEEE
Access 2019, 7, 83252–183261. [CrossRef]
10. Saravanakumar, R.; Krishnaraj, N.; Venkatraman, S.; Sivakumar, B.; Prasanna, S.; Shankar, K. Hierarchical symbolic analysis and
particle swarm optimization based fault diagnosis model for rotating machineries with deep neural networks. Measurement 2021,
171, 108771. [CrossRef]
Sensors 2021, 21, 418 19 of 20

11. Du, X.L.; Chen, Z.G.; Zhang, N.; Xu, X. Bearing fault diagnosis based on Synchronous Extrusion S transformation and deep
learning. Modul. Mach. Tool Autom. Process. Technol. 2019, 5, 90–93, 97.
12. Peng, C.; Tang, Z.H.; Gui, W.H.; He, J. A bidirectional weighted boundary distance algorithm for time series similarity computation
based on optimized sliding window size. J. Ind. Manag. Optim. 2019, 13, 1–16. [CrossRef]
13. Peng, C.; Tang, Z.H.; Gui, W.H.; Chen, Q.; Zhang, L.X.; Yuan, X.P.; Deng, X.J. Review of key technologies and progress in
industrial equipment health management. IEEE Access 2020, 8, 151764–151776. [CrossRef]
14. Yang, Y.K.; Fan, W.B.; Peng, D.X. Driving behavior recognition based on one-dimensional convolutional neural network and
noise reduction autoencoder. Comput. Appl. Softw. 2020, 37, 171–176.
15. Peng, C.; Chen, Q.; Zhou, X.H.; Tang, Z.H. Wind turbine blades icing failure prognosis based on balanced data and improved
entropy. Int. J. Sens. Netw. 2020, 34, 126–135. [CrossRef]
16. Peng, C.; Liu, M.; Yuan, X.P.; Zhang, L.X. A new method for abnormal behavior propagation in networked software. J. Internet
Technol. 2018, 19, 489–497.
17. Zhang, J.D.; Zou, Y.S.; Deng, J.L.; Zhang, X.L. Bearing remaining life prediction based on full convolutional layer neural networks.
China Mech. Eng. 2019, 30, 2231–2235.
18. Yang, B.Y.; Liu, R.N.; Zio, E. Remaining useful life prediction based on a double-convolutional neural network architecture. IEEE
Trans. Ind. Electron. 2019, 66, 9521–9530. [CrossRef]
19. Hsu, H.Y.; Srivastava, G.; Wu, H.T.; Chen, M.Y. Remaining useful life prediction based on state assessment using edge computing
on deep learning. Comput. Commun. 2020, 160, 91–100. [CrossRef]
20. Li, X.; Ding, Q.; Sun, J.Q. Remaining useful life estimation in prognostics using deep convolutional neural networks. Reliab. Eng.
Syst. Saf. 2018, 172, 1–11. [CrossRef]
21. Zhang, J.J.; Wang, P.; Yuan, R.Q.; Gao, R.X. Long short-term memory for machine remaining life prediction. J. Manuf. Syst. 2018,
48, 78–86. [CrossRef]
22. Kong, Z.; Cui, Y.; Xia, Z.; Lv, H. Convolution and long short-term memory hybrid deep neural networks for remaining useful life
prognostics. Appl. Sci. 2019, 9, 4156. [CrossRef]
23. Song, Y.; Xia, T.B.; Zheng, Y.; Zhuo, P.C.; Pan, E.S. Residual life prediction of turbofan Engines based on Autoencoder-BLSTM.
Comput. Integr. Manuf. Syst. 2019, 25, 1611–1619.
24. Yan, C.M.; Wang, W. Development and application of a convolutional neural network model. Comput. Sci. Explor. 2020, 18, 1–22.
25. Goodfellow, I.; Bengio, Y.; Courville, A. Deep Learning; MIT Press: Cambridge, MA, USA, 2016; pp. 326–366.
26. Estrach, J.B.; Szlam, A.; LeCun, Y. Signal recovery from pooling representations. In Proceedings of the 31st International
Conference on Machine Learning (ICML), Beijing, China, 21–26 June 2014; pp. 307–315.
27. Jonathan, L.; Shelhamer, E.; Darrell, T.; Berkeley, U.C. Fully convolutional networks for semantic segmentation. IEEE Trans.
Pattern Anal. Mach. Intell. 2017, 39, 640–651.
28. Zhang, Y.; Kong, W.; Dong, Z.Y.; Jia, Y.; Hill, D.J.; Xu, Y. Short-term residential load forecasting based on LSTM recurrent neural
network. IEEE Trans. Smart Grid 2019, 12, 312–325.
29. Al-Dulaimi, A.; Zabihi, S.; Asif, A.; Mohammadi, A. A multimodal and hybrid deep neural network model for Remaining Useful
Life estimation. Comput. Ind. 2019, 108, 186–196. [CrossRef]
30. Kingma, D.; Ba, J. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning
Representations (ICLR), Banff, AB, Canada, 14–16 April 2014.
31. Famouri, M.; Azimifar, Z.; Taheri, M. Fast linear SVM validation based on early stopping in iterative learning. Int. J. Pattern
Recognit. Artif. Intell. 2015, 29, 1551013. [CrossRef]
32. Loffe, S.; Szegedy, C. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv 2015,
arXiv:1502.03167.
33. Ramasso, E.; Gouriveau, R. Prognostics in switching systems: Evidential Markovian classification of real-time neuro-fuzzy
predictions. In Proceedings of the Prognostics and Health Management Conference IEEE PHM, Portland, OR, USA, 10–16
October 2010.
34. Frederick, D.; de Castro, J.; Litt, J. User’s Guide for the Commercial Modular Aero-Propulsion System Simulation (C-MAPSS);
NASA/ARL: Hanover, MD, USA, 2007.
35. Saxena, A.; Goebel, K.; Simon, D.; Eklund, N. Damage propagation modeling for aircraft engine run-to-failure simulation.
In Proceedings of the 1st International Conference on Prognostics and Health Management (PHM08), Denver, CO, USA, 6–9
October 2008.
36. Peel, L. Data driven prognostics using a Kalman filter ensemble of neural network models. In Proceedings of the 2008 International
Conference on Prognostics and Health Management IEEE, Denver, CO, USA, 6–9 October 2008.
37. Li, N.; Lei, Y.; Gebraeel, N.; Wang, Z.; Cai, X.; Xu, P.; Wang, B. Multi-sensor data-driven remaining useful life prediction of
semi-observable systems. IEEE Trans. Ind. Electron. 2020, 1. [CrossRef]
38. Gou, B.; Xu, Y.; Feng, X. State-of-health estimation and remaining-useful-life prediction for lithium-ion battery using a hybrid
data-driven method. IEEE Trans. Veh. Technol. 2020, 69, 10854–10867. [CrossRef]
39. Hsu, C.; Jiang, J. Remaining useful life estimation using long short-term memory deep learning. In Proceedings of the 2018 IEEE
International Conference on Applied System Innovation (ICASI), Tokyo, Japan, 13–17 April 2018; pp. 58–61.
Sensors 2021, 21, 418 20 of 20

40. Xu, S.; Hou, G.S. Prediction of remaining service life of turbofan engine based on VAE-D2GAN. Comput. Integr. Manuf. Syst. 2020,
23, 1–17.
41. Khelif, R.; Chebel-Morello, B.; Malinowski, S.; Laajili, E.; Fnaiech, F.; Zerhouni, N. Direct remaining useful life estimation based
on support vector regression. IEEE Trans. Ind. Electron. 2016, 64, 2276–2285. [CrossRef]
42. Liao, Y.; Zhang, L.; Liu, C. Uncertainty Prediction of Remaining Useful Life Using Long Short-Term Memory Network Based
on Bootstrap Method. In Proceedings of the IEEE International Conference on Prognostics and Health Management (ICPHM),
Seattle, WA, USA, 11–13 June 2018; pp. 1–8.
43. Zheng, S.; Ristovski, K.; Farahat, A.; Gupta, C. Long short-term memory network for remaining useful life estimation. In
Proceedings of the 2017 IEEE International Conference on Prognostics and Health Management (ICPHM), Dallas, TX, USA, 19–21
June 2017; pp. 88–95.
44. Zhang, C.; Lim, P.; Qin, A.; Tan, K. Multiobjective deep belief networks ensemble for remaining useful life estimation in
prognostics. IEEE Trans. Neural Netw. Learn Syst. 2017, 28, 2306–2318. [CrossRef] [PubMed]
45. Yu, W.N.; Kim, I.Y.; Mechefske, C. Remaining useful life estimation using a bidirectional recurrent neural network based
autoencoder scheme. Mech. Syst. Signal Process. 2019, 129, 764–780. [CrossRef]
46. Babu, G.S.; Zhao, P.; Li, X. Deep convolutional neural network based regression approach for estimation of remaining useful life.
In Proceedings of the International Conference on Database Systems for Advanced Applications, Dallas, TX, USA, 16–19 April
2016; pp. 214–228.