Knowledge-Based Systems xxx (xxxx) xxx

Contents lists available at ScienceDirect

Knowledge-Based Systems
journal homepage: www.elsevier.com/locate/knosys

A personalized diagnosis method to detect faults in gears using

numerical simulation and extreme learning machine✩
∗
Xiaoyang Liu a , Haizhou Huang b , Jiawei Xiang a ,
a
College of Mechanical and Electrical Engineering, Wenzhou University, Wenzhou, 325035, PR China
b
Huadian Electric Power Research Institute, Hangzhou, 310030, PR China

article info a b s t r a c t

Article history: Fault classification methods are a long-term research focus in both science society and engineering
Received 5 December 2019 application. Generally, every real-world running mechanical system is exhibit personalized vibration
Received in revised form 30 December 2019 behaviors and the corresponding fault samples of such systems are difficult to be obtained. Therefore,
Accepted 10 February 2020
extreme learning machine (ELM), a typical fault classification method is failed to attain agreeable fault
Available online xxxx
detection results. In this paper, a personalized fault diagnosis method using finite element method
Keywords: (FEM) simulation and ELM is proposed to detect faults in gears. The method includes three steps.
Personalized fault diagnosis Firstly, The FEM model of gears with faults is constructed to obtain fault samples (simulation signals).
Gears Secondly, to achieve ELM training process, the meshing frequency components of each simulation
Finite element method signal is separated into sub-signals and the corresponding time and time–frequency domains indicators
Numerical simulation
are served as training samples. Finally, the measured vibration signals of gear transmission systems
Extreme learning machine
are employed as testing samples of trained ELM to recognize its fault types. The classification accuracy
ratios of gear states in a cracked teeth of driving gear, a peeled teeth of driving gear, a broken teeth of
driving gear, a peeled teeth of driving gear and a broken teeth of driven gear, a broken teeth of driving
gear and a broken teeth of driven gear are 85%, 90%, 92.5%, 90% and 85% respectively. It is expect that
the proposed personalized fault diagnosis method can set up a bridge between fault classification
methods and real-world applications.
© 2020 Elsevier B.V. All rights reserved.

1. Introduction intelligent diagnosis methods are proposed to detect the faults in

gearboxes [8,9]. Wei et al. [10] proposed a fault diagnosis method
With the rapid development of artificial intelligence (AI) tech- for gearbox based on local mean decomposition (LMD), permuta-
nology, classification of faults in mechanical components using tion entropy (PE) and extreme learning machine (ELM). Through
machine learning is a hot topic in the field of science and engi- training based on different types of measured fault samples,
neering. As the important component of mechanical transmission the fault detection of gears can be realized by using intelligent
system, gear faults will make the system exhibit personalized methods such as ELM, support vector machine (SVM), and neural
vibration behaviors, which directly reflects the reliability and networks (NNs). However, in real world, it is difficult to obtain
stability of the entire mechanical equipment. Therefore, in order suitable training samples to represent various types of gear faults
to detect gear faults by using vibration signals, many fault di- that may occur in actual mechanical systems, and the lack of
agnosis methods are proposed, such as fault feature extraction fault training samples greatly limits the application of various
methods, AI model-based methods, etc. [1–6]. To detect faults intelligence diagnosis methods in engineering applications.
in gear transmission systems, Wang et al. [7] proposed a three- To in-depth understand the fault mechanism and obtain an
stage intelligent fault diagnosis approach to identify different agreeable fault detection results, numerical simulations are used
gears faults with less prior knowledge and diagnostic expertise. for fault diagnosis. Sun et al. [11] used failure dynamics simula-
With small training samples, two kinds of deep transfer-based tion to select the most sensitive indicators and the margin factors,
as the prior information reflecting the local gear fault and extract-
ing the fault characteristics from vibration signals. In Ref. [12],
✩ No author associated with this paper has disclosed any potential or
the influence of tooth crack on the vibration response of spur
pertinent conflicts which may be perceived to have impending conflict with
gear primary gearbox was investigated by dynamic simulations.
this work. For full disclosure statements refer to https://doi.org/10.1016/j.knosys.
2020.105653. To detect the abnormal conditions of shafts, two kinds of wavelet-
∗ Corresponding author. based numerical simulation models are carried out to calculate
E-mail address: [email protected] (J. Xiang). the dynamic responses of a shaft [13,14]. Thereafter, based on

https://doi.org/10.1016/j.knosys.2020.105653
0950-7051/© 2020 Elsevier B.V. All rights reserved.

Please cite this article as: X. Liu, H. Huang and J. Xiang, A personalized diagnosis method to detect faults in gears using numerical simulation and extreme learning machine,
Knowledge-Based Systems (2020) 105653, https://doi.org/10.1016/j.knosys.2020.105653.
2 X. Liu, H. Huang and J. Xiang / Knowledge-Based Systems xxx (xxxx) xxx

the fault mechanism, Xiang and Zhong [15] further proposed a artificial intelligence models will be activated for the diagnosis of
novel fault diagnosis scheme using finite element method (FEM), mechanical faults under different working conditions. Focus on
wavelet packet transform (WPT) and SVM. To detect faults in re- gear transmission systems, the main contributions of the present
ciprocating machinery, To detect faults in bearings of axial piston method is that we proposed a specific personalized diagnosis
pump under natural impulse working conditions, Wang et al. [16] method based on FEM simulation and ELM. To solve no fault sam-
proposed a combination method using the minimum entropy ple problem for detecting faults in gear transmission systems, the
deconvolution (MED) and band-pass filter which the center fre- FEM simulation and cosine similarity-based FEM model updating
quency calculated by FEM simulation. In general, as a powerful technique are employed to construct effective fault FEM model
tool, FEM simulation can obtain a large number of numerical to calculate fault training samples. The paper is organized as
experimental data/signals at a lower experimental cost, especially follows. In Section 2, the basic principle of personalized diagnosis
those which are difficult to obtain through physical experiments. method is introduced. In Section 3, FEM model of a single stage
In addition, the experimental data are pure without interference, gear transmission system is constructed and further updated
and the corresponding working conditions of mechanical system experimental investigations using the cosine similarity calculated
are deterministic and controllable. by numerical simulation and experimental signals. Then, using
FEM models are usually based on highly idealized engineering measured signal as testing samples and simulation signals as
blueprints and designs, and the results of analysis and predic- training samples. Finally, the ELM classification results of five
tion are often different from the actual situation. In order to types of faults located in gears are given. The conclusion of this
reduce the difference from the actual condition and get more study is given in Section 4.
accurate results, it is necessary to update the FEM model. Ba-
sically, as an optimization problem, the FEM model updating is 2. Basic idea of the present method
achieved by correcting the design parameters constantly, such
as boundary conditions and materials [17]. According to the dy- 2.1. Cosine similarity-based FEM model updating technique
namic theory of mechanical systems, the dynamic information
Cosine similarity-based FEM model updating technique is a
in different assembly conditions and running states is carried in
novel FEM model updating method based on time-domain vi-
the corresponding vibration signals, such as displacement and
bration response, which can update the FEM model without
acceleration [18]. Therefore, for dynamic analysis of mechanical
modal experiment. As we know, vibration of mechanical sys-
systems, various model parameters can be identified effectively
tem is the external reflection of its intrinsic dynamic charac-
by comparing the measured signals with the standard signals.
teristics, thus vibration signals can be used to determine the
Cosine similarity-based FEM model updating method is the com-
corresponding main parameters, such as materials and bound-
monly used tool [19]. The cosine similarity between measured
ary conditions [18]. Many researchers have done researches on
and simulated signals is used as the optimized objective func-
model parameter identification and fault diagnosis based on
tion in time-domain to update the uncertain model parameters
time-domain response [24–26]. Sarin et al. [27] made a com-
effectively.
parative study of time-domain signals, analyzed several methods
The construction of feature vectors is an important step in
of residual errors between time-domain signals, and provided
signal preprocessing, which directly affects the diagnosis results.
a theoretical basis for model updating based on time-domain
Generally, feature vectors are generated by calculating the fea-
response. Zapico-Valle et al. [28] defined the residual between
ture indicators of vibration signals in time-domain or frequency the time-domain response of FEM simulation and that of exper-
domain. For gear faults, Chen et al. [20] proposed sixteen time- iment, and applied several optimization methods to search the
domain feature parameters (mean, root mean square, standard minimum residual, so as to achieve the purpose of FEM model
deviation, skewness, kurtosis and shape indicator) and thirteen updating. Here, we directly use the cosine similarity between the
frequency-domain feature indicators (such as average frequency, time-domain response signals of simulation and measurement to
standard deviation frequency, etc.) to construct feature sets for update FEM model by adjusting the selected parameters. The co-
training samples. Liu et al. [21] used a hybrid time–frequency sine similarity between the measured signals and the simulation
analysis method to get the feature information of gear faults. signals is used as the objective function:
Based on the personalized vibration behaviors in different work- ∑n
ing states of real-world running mechanical systems, the vibra- (xi × yi ) X ·Y
cos (θ) = √ i=1
√∑ = (1)
tion signals can be automatically identified and classified by these ∑n n ∥X ∥ × ∥Y ∥
i=1 (xi ) × i=1 (yi )
2 2
feature parameters.
ELM [22,23] is firstly proposed by Professor Huang to form where X and Y represent the measured signal and the simulation
a new learning scheme of feed forward neural networks. ELM signal, respectively. When the closer the value of cos(θ ) is to 1,
randomly selects all the hidden node parameters, after which the the more similar the two signals are. Generally, cos (θ) > 0.6 will
network can be represented as a linear system and the output lead to a satisfactory result [29].
weights can be computed analytically. Due to the ELM tends
to obtain the smallest training error and the smallest norm of 2.2. A brief review of ELM
weights, the advantages of ELM are the universal approxima-
tion capability, fast learning speed and simple learning structure. ELM is a learning algorithm that is developed for generalized
Therefore, ELM is widely used in fault classification. However, due single hidden layer feed forward networks (SLFNs) with wide
to the lack of fault training samples, the recognition of faults types variety of hidden nodes. Fig. 1 shows the network structure of
by ELM is severely limited in engineering applications. the ELM.
Recently, motivated by personalized medicine, Xiang [15] de- Consider N arbitrary distinct samples (Xi , ti ) ∈ Rn × Rm . If
veloped a personalized diagnosis idea for mechanical systems by a SLFN with L hidden nodes can approximate these N samples
the combination of numerical simulation and machine learning with zero error, it then implies that there exist βi , αi and bi such
methods or artificial intelligence models. If the updated FEM that [30]:
model can effectively represent the real-world ones, the lack of L
∑
fault training samples will be inexpensive represented by sim- βi G αi · Xj + bi , = tj , j = 1, . . . , N .
( ) ( )
f L Xj = (2)
ulation data/signals. Thereafter, machine learning methods or i=1

Please cite this article as: X. Liu, H. Huang and J. Xiang, A personalized diagnosis method to detect faults in gears using numerical simulation and extreme learning machine,
Knowledge-Based Systems (2020) 105653, https://doi.org/10.1016/j.knosys.2020.105653.
 X. Liu, H. Huang and J. Xiang / Knowledge-Based Systems xxx (xxxx) xxx 3

Table 1
Indicators in time and time–frequency domains.

Indicator Equation
√∑
N
Standard deviation xstd xstd = i=1 (x (i) − xm )2 /N
Peak xP xP = max |x (i)|
∑N
i=1 (x (i) − xm ) /N
3
Skewness xske xske =
∑N
i=1 (x (i) − xm ) /Nxstd
4 4
Kurtosis xkur xkur =
√∑
N
i=1 (xi ) /N
2
Root mean square xrms xrms =

Peak factor PF PF = xP /xrms

(∑
N√ )2
Square root amplitude xsra xsra = i=1|x (i)|/N
(∑ )
N
Shape factor SF SF = xrms / i=1 |x (i)| /N
(∑ )
N
Impulse factor IF IF = xP / i=1 |x (i)| /N
∑N ∑2j −1 ∑N
n=1 |xi (n)| /
2 2
Wavelet packet decomposition (WPD) energy WPDi = k=0 n=1 |xk (n)|
∑N NI N
n=1 |IMFi (n)| /
2 2
∑ ∑
Ensemble empirical mode decomposition(EEMD energy EEMDi = k= 1 n=1 |IMFk (n)| )

Note: x is the data; N is the number of data points; xm is the mean value of x; σs is the standard deviation. xk (n) is the
decomposition coefficient sequence of the kth (k = 0, 1, . . ., 2j − 1, j is the WPD decomposition level) frequency band
using WPD; IMF k (n) is kth data after EEMD, and NI is the decomposition level using EEMD. If WPDi (i = 1, 2, . . ., 8)
and EEMDi (i = 1, 2, . . ., 6) are used, we have 23 indicators in time and time–frequency domains.

is estimated as:
β = H ∗T (6)
where H* is the Moore–Penrose generalized inverse [31] of the
hidden layer output matrix H. The ELM algorithm which only
consists of three steps can then be summarized as:
ELM Algorithm: Given a training set {(Xi , ti )|Xi ∈ Rn , ti ∈ Rm ,
i = 1, . . . , N }, activation function g(x), and hidden node number
L,
(1) Assign random hidden nodes by randomly generating pa-
rameters (αi , bi ), i = 1, . . . , L.
(2) Calculate the hidden layer output matrix H.
(3) Calculate the output weight β = H ∗ T .

2.3. The personalized diagnosis method using numerical simulation

and ELM
Fig. 1. The network structure of ELM.
For different gear transmission systems, the vibration response
signals exhibit personalized characteristics. A generalized fault
where αi and bi are the learning parameters of the hidden
( nodes,) samples of gears definitely not available for all the gear trans-
βi is the output weight, tj is the test point, and G αi , bi , Xj mission systems. Therefore, a new idea for personalized fault di-
denotes the output of the ith hidden node with respect to the agnosis is developed using FEM simulation to activate ELM. Fig. 2
input. shows the flowchart of the present method and the procedures
Eq. (2) can be written compactly as: are:

Hβ = T (3) (1) Construct and update the FEM model of the mechanical sys-
tem
where the hidden layer output matrix of the network H is: FEM model is constructed using the commercial FEM software
H (α1 , . . . , αL , b1 , . . . , bL , X1 , . . . , XN ) ANSYS [32]. Firstly, in order to keep the FEM model effectively
represents the corresponding physical mechanical systems, the
G (α1 , b1 , X1 ) G (αL , bL , X1 )
⎡ ⎤
··· cosine similarity cos(θ ) is calculated between FEM simulation and
.. .. ..
=⎣ . (4) measurement time-domain vibration signals of the health me-
⎢ ⎥
. . ⎦
G (α1 , b1 , XN ) ··· G (αL , bL , XN ) N ×L chanical system. Secondly, iteratively changed the key sensitive
model parameters to meet the criterion cos(θ ) > 0.6, and the
in which the output weight vector and testing sample vector β FEM model with agreeable key sensitive model parameters can
and T is: well represent its physical mechanical systems. Finally, the FEM
β1T models with faults will be obtained by inserted faults into the
⎡ ⎤ ⎡ ⎤
T1T
β = ⎣ ... ⎦ ,
⎢.⎥ updated FEM model.
T = ⎣ .. ⎦ (5)
⎢ ⎥
(2) Obtain the fault training samples
βT
L TNT
In this part, firstly, the FEM models of the mechanical sys-
According to ELM theories, all the hidden nodes (αi , bi ) can be tem with different faults are calculated to obtain the simulation
randomly assigned instead of being tuned. The solution of Eq. (3) vibration signals, and the length of each one is the same in

Please cite this article as: X. Liu, H. Huang and J. Xiang, A personalized diagnosis method to detect faults in gears using numerical simulation and extreme learning machine,
Knowledge-Based Systems (2020) 105653, https://doi.org/10.1016/j.knosys.2020.105653.
4 X. Liu, H. Huang and J. Xiang / Knowledge-Based Systems xxx (xxxx) xxx

Fig. 2. The flowchart of the proposed personalized fault diagnosis method.

time domain. Then, each simulation signal is divided into sub- Table 2
The material and geometric parameters of gears.
signals with the same length, and the corresponding time and
time–frequency domains indicators (as shown in Table 1) of Density 7850 kg/m3 Pressure angle 20◦

each sub-signal are further calculated. Therefore, the number of Elastic modulus E 2.06×10 11
Pa Tooth number of driving gear 55
Poisson’s ratio ν 0.3 Tooth number of driven gear 75
training samples corresponding to each fault is the same, which
Tooth width 20 mm Modulus 2
ensure the balance of samples.
(3) Fault classification using ELM Table 3
The fault training samples formed by the time and time– The material and geometric parameters of bearings.
frequency domains indicators are employed as inputs to train Density 7860 kg/m3 Outer race diameter 45 mm
ELM. Same as the processing of simulation signals, each measured Elastic modulus E 2.06×10 11
Pa Inner race diameter 30 mm
signal (the fault type is unknown) is used to generate the testing Poisson’s ratio ν 0.3 Roller diameter 7.5 mm
samples of trained ELM, and finally fault patterns are detected. Roller contact length 12 mm Number of rollers 15

3. Experimental investigation
3.1. FEM model construction and updating for gears
In this section, an experimental investigation of a single stage
gear transmission system is carried out, which proves that the According to the experimental gearbox, the three-dimensional
(3D) FEM model of gear transmission system is constructed by
personalized diagnosis method is feasible in gear faults diagnosis
using commercial FEM software ANSYS. In order to reduce the
using FEM numerical simulation driving ELM. The experimental
calculation time, the gearbox housing simplified into four bearing
gear transmission system consists of a couple of spur gears and
seats, the stiffness and damping characteristics of bearings are
shafts, and four cylindrical roller bearings of type N205, as shown
represented by using contact pairs, shown in Fig. 6.
in Fig. 3. The experimental device includes gearbox, accelera- To construct FEM model, 3D solid element (SOLID164) is em-
tion sensors, motor, speed controller, data acquisition instrument, ployed to mesh the geometry of the gear transmission system.
data processing software, etc. as shown in Fig. 4. In order to apply rotating loadings, shell element (SHELL163) is
The main geometric parameters and material parameters of necessary to mesh the surface of driving shaft. Meshing the whole
gears and bearings are listed in Tables 2 and 3, respectively. The geometric model by sweeping, it can ensure the convergence
rotation speed is 1474 rpm, and the reverse torque of 5 N m is of the computing process and reduce to computing time, avoid
applied to the driven gear as the resistance load. The sampling singular values in simulation signals. Gear meshing is a complex
frequency is 5120 Hz. Fig. 5 shows the geometrical dimension of contact process. In order to reduce the computing time and
the test rig. improve the computing accuracy, the element size is changed

Please cite this article as: X. Liu, H. Huang and J. Xiang, A personalized diagnosis method to detect faults in gears using numerical simulation and extreme learning machine,
Knowledge-Based Systems (2020) 105653, https://doi.org/10.1016/j.knosys.2020.105653.
 X. Liu, H. Huang and J. Xiang / Knowledge-Based Systems xxx (xxxx) xxx 5

according to the position on the gear. The element size of teeth

is all 1.57 mm, 3.5 mm of driving gear body and driving shaft, 5
mm of driven gear body and driven shaft, and the element size of
bearing seats (gearbox housing) is 5 mm. Finally, the FEM model
of gear transmission system is constructed, as shown in Fig. 6.
Based on the actual working condition of gears, five contact
pairs are set up in the FEM model, as shown in Fig. 6, including
four contact pairs between shaft and bearing seats (denoted by
C1 , C2 , C3 and C4 ), one contact pair between driving gear and
driven gear (denoted by C5 ), and the whole constraints of all
nodes are conducted in the outer surface of four bearing seats.
In the construction of FEM model, six model parameters are
involved, including the friction coefficient, stiffness and damping
coefficient at each contact pair.
Due to the meshing of gears is in the state of mixed lubrica-
Fig. 3. The single stage gear transmission system.
tion, the corresponding contact friction coefficient is calculated
according to the Ref. [33] and the main working condition, and it
is finally set as the maximum value of 0.08 in the whole process
of gears meshing to avoid slipping. The function for calculating
the friction coefficient is as follows:
45.94 · w
( )
µf = 0.0127 lg (7)
η0 · VS · Vr2
where µf is the friction coefficient of tooth contact, w is the
load on unit width of gear teeth, the empirical value of absolute
viscosity of lubricants η0 = 0.02 Pa s, the sliding velocity VS =
V2 − V1 , the sliding velocity Vr = V1 − V2 , in which V1 represents
the tangential velocity of driving gear at meshing point, and V2
is the tangential velocity of driven gear at meshing point. The
friction coefficient of the selected cylindrical roller bearing is very
small, which has little effect on the vibration response of system.
Therefore, the friction of corresponding contact pairs (C1 , C2 , C3
Fig. 4. The experimental setup of a single stage gear transmission system.
and C4 ) is not considered in the FEM model.

Fig. 5. The geometrical dimension of gear transmission system.

Fig. 6. The FEM model of gear transmission system.

Please cite this article as: X. Liu, H. Huang and J. Xiang, A personalized diagnosis method to detect faults in gears using numerical simulation and extreme learning machine,
Knowledge-Based Systems (2020) 105653, https://doi.org/10.1016/j.knosys.2020.105653.
6 X. Liu, H. Huang and J. Xiang / Knowledge-Based Systems xxx (xxxx) xxx

Fig. 7. The measured and simulation signals: (a) the measured signal in time-domain, (b) the simulation signal in time-domain, (c) the spectrum of measured signal,
(d) the spectrum of simulation signal, (e) the spectrum of measured signals from 1200 Hz to 1500 Hz, (f) the spectrum of simulation signal from 1200 Hz to 1500
Hz.

Fig. 8. The measured and simulation signals after filtering: (a) the measured signal, (b) the simulation signal.

Generally, the damping coefficient of gear transmission system known, the characteristics of contact damping can be realized
in running state is variable, but the damping ratio is constant by setting the damping ratio ξ in ANSYS. Therefore, based on
and can be used to indirectly determine the contact damping the rated load and geometric parameter of the bearing, carry
coefficient when the stiffness is known [34,35]. Therefore, when out the calculation according to Refs. [36–39], the ξ is calculated
the stiffness coefficient of cylindrical roller bearings and gears is out and set to the constant value 0.09 as the comprehensive

Please cite this article as: X. Liu, H. Huang and J. Xiang, A personalized diagnosis method to detect faults in gears using numerical simulation and extreme learning machine,
Knowledge-Based Systems (2020) 105653, https://doi.org/10.1016/j.knosys.2020.105653.
 X. Liu, H. Huang and J. Xiang / Knowledge-Based Systems xxx (xxxx) xxx 7

27.4R1x1.5 R1x2.5 l0.805 Z 0.805 Fr0.19

C = 1.05
(8)
α 0.81 0.05 n1.05 r 0.645
η (R1 + r ) E ′−0.045 (1 + γ )1.695 (1 − γ )1.05 R1x1.5 + (1 − γ )1.695 (1 + γ )1.05 R1x2.5 ·
[ ]
0 i
1
K = (9)
0.13 · A3 · Fr−1.13 + A2 · InFr + A1 + A2
8.16 8.16R2 r
[ ]
1
· In 0.39E ′ ZI (R1 + r ) + 1.15 − 0.5In
( )
A1 = · (10)
E′ · Z · I π E ′ (R2 − r ) ZI
1
A2 = −6.68 · (11)
E ′ ZI
0.7
A3 = 0.28 · α 0.54 (η0 ni ) r 0.43 (R1 + r ) E ′−0.03 Z −0.03 I 0.13 ·
(12)
(1 − γ )1.13 (1 + γ )0.7 + (1 + γ )1.13 (1 − γ )0.7
[ ]

C
ξ= √ (13)
2 K ·m

Box I.

contact damping ratio of bearings. The calculation is given in

Box I where K and C respectively represent the comprehensive
contact stiffness and damping coefficient of the bearing under
rated load, the RX1 and RX2 are equivalent radius of curvature
at the contact of roller with inner raceway and outer raceway
respectively, 1/Rx1 = 1/R1 + 1/r , 1/Rx2 = 1/R2 + 1/r, R1 and
R2 represent the radius of inner raceway and outer raceway; I
denotes the effective contact length between roller and raceway;
Z is the number of cylindrical rollers; Fr represents the radial
load; the viscous-pressure coefficient of lubricating oil is α =
2.3 × 108 Pa−1 ; ni is the rotation speed of inner ring, r represents
the radius of roller; E’ is the equivalent elastic modulus, E ′ =
E /(1 − v2 ), γ = r /(R1 + r).
For setting the contact stiffness to represent the stiffness of
bearings, the normal contact stiffness factor FKN is employed in
ANSYS. FKN is used to estimate the contact stiffness based on the
material properties and the elemental deformations, and it is with
the interval [0.1, 1] to avoid penetration generally [40]. Due to
the contact stiffness of bearings is very large, the corresponding Fig. 9. The change of cosine similarity values through iteratively adjusting of
parameters.
FKN is empirically set to the value of 0.12 for ensuring the
convergence of calculation.
However, the meshing stiffness and meshing damping of gears
are unknown, and they are the main factors affecting the vi- time-domain, and the model parameters are updated using the
bration response of the whole system. Therefore, the meshing cos(θ ) model updating technique.
stiffness and meshing damping ratio are updated in updating the The FKN corresponding to gears meshing stiffness is initially
FEM model of the gear transmission system. Based on the maxi- set to be 0.1, and meshing damping ratio is 0. By calculating the
mum value of FKN while ensuring the convergence of calculation, initial FEM model of the gears under healthy state, the measured
the meshing stiffness FKN is updated in the range of 0.1–0.16. and simulation signals for 1s are shown in Fig. 7.
Refer to Ref. [41], the update range of meshing damping ratio is It can be seen that both the time-domain and frequency-
set 0–0.05. domain signals of the measured and simulation ones are different.
The vibration of mechanical system is the external reflection However, from Fig. 7(e) and (f), it is clearly that the two spectrum
of its dynamic characteristics. Therefore the dynamic parameters near the meshing frequency is similar. After filtering by bandpass
can be effectively identified to update the FEM model by com- filter with bandwidth [1200 Hz 1500 Hz], the simulation and
paring the simulation signal with the measured signal [18]. The
measured signals are more similar in time-domain, as shown in
meshing stiffness and meshing damping are the most important
Fig. 8(a) and (b).
factors affecting the vibration response of a gear transmission
Through iteratively adjusting the FKN of the meshing stiffness
system, and their changes will obviously change the frequency
and meshing damping ratio according to the flowchart shown in
spectrum of vibration signal, which is reflected in the modulated
side bands on both sides of the meshing frequency [42–44]. After Fig. 2, we finally obtain the cosine similarity cos(θ ) = 0.742. The
calculation, the meshing frequency of the gear is 1352 Hz, and changing trend of cosine similarity values are shown in Fig. 9
the rotational frequencies of the gear pair are 24.5 Hz and 18 and Table 4. Fig. 10 shows the comparison between measured
Hz, respectively. Therefore, in the whole process of updating FEM and simulation signals when cos(θ ) = 0.742. The signals are
model, the simulation and measured signals are filtered firstly by matched well at a certain degree, which proves that the updated
band-pass filter (the central frequency is 1352 Hz and the optimal FEM model constructed by using the input parameters of ANSYS
bandwidth is 200 Hz), then the filtered signal is normalized in the as shown in Table 5 is effective.

Please cite this article as: X. Liu, H. Huang and J. Xiang, A personalized diagnosis method to detect faults in gears using numerical simulation and extreme learning machine,
Knowledge-Based Systems (2020) 105653, https://doi.org/10.1016/j.knosys.2020.105653.
8 X. Liu, H. Huang and J. Xiang / Knowledge-Based Systems xxx (xxxx) xxx

Fig. 10. The comparison between measured and simulation signals when cos(θ ) = 0.742: (a) the measured signal, (b) the simulation signal.

Fig. 11. The FEM models of gears in different state: (a) a cracked teeth in the driving gear, (b) a peeled teeth in the driving gear, (c) a broken teeth in the driving
gear, (d) a combination of peeled teeth in the driving gear and broken teeth in the driven gear, (e) a combination of broken teeth in the driving gear and broken
teeth in the driven gear.

3.2. Generate simulation fault training samples by T2 ), (3) a broken teeth in the driving gear (denoted by T3 ),
(4) a combination of peeled teeth in the driving gear and broken
According to the commonly occurred faults of a gear transmis- teeth in the driven gear (denoted by T4 ), (5) a combination of
sion systems, five types of faults are considered in fault samples broken teeth in the driving gear and broken teeth in the driven
as shown in Fig. 11: (1) a cracked teeth in the driving gear gear (denoted by T5 ). Using the updated FEM model parameters
(denoted by T1 ), (2) a peeled teeth in the driving gear(denoted (shown in Table 5), the FEM models with the above five types

Please cite this article as: X. Liu, H. Huang and J. Xiang, A personalized diagnosis method to detect faults in gears using numerical simulation and extreme learning machine,
Knowledge-Based Systems (2020) 105653, https://doi.org/10.1016/j.knosys.2020.105653.
 X. Liu, H. Huang and J. Xiang / Knowledge-Based Systems xxx (xxxx) xxx 9

Fig. 12. The filtered simulation signals in time-domain and frequency domain: (a) fault type T1 (b) fault type T2 , (c) fault type T3 , (d) fault type T4 , (e) fault type T5 .

Table 4 Table 5
The cosine similarity values through iteratively adjusting of parameters. The input parameters of updated FEM model in ANSYS.
Cosine similarity Meshing damping ratio Friction coefficient 0.08
0 0.01 0.02 0.03 0.04 0.05 Synthetic damping ratio of bearings 0.09
0.1 0.417 0.401 0.389 0.421 0.382 0.395 Contact damping ratio of gears 0.005
0.11 0.385 0.491 0.466 0.511 0.436 0.379 Meshing stiffness FKN of gears 0.15
0.12 0.451 0.501 0.533 0.612 0.621 0.509 Comprehensive contact stiffness FKN of bearings 0.16
Meshing stiffness FKN 0.13 0.496 0.606 0.569 0.663 0.671 0.532
0.14 0.503 0.583 0.675 0.711 0.655 0.506
0.15 0.473 0.594 0.722 0.742 0.691 0.585
0.16 0.432 0.505 0.606 0.562 0.593 0.511 300 data points as the length), then calculate the time and time–
frequency domains indicators of each sub-signal. For each faulty
state in the gears, 9 × 40 = 360 features are forming the feature
vector, and finally the fault training samples of five types of faults
of faults are constructed to generate the simulation signals, and is a 360 × 5 matrix.
the length of each simulation signal is 12,000 data points. Fig. 12
shows the corresponding filtered signals (using bandpass filter 3.3. Experimental investigations using the proposed personalized
with bandwidth [1200 Hz 1500 Hz]) in time-domain (normalized) diagnosis method
and frequency-domain, respectively.
In the section, the measured dataset corresponding to FEM
To generate the fault training samples, each filtered simulation simulation is used to conduct the experimental investigations
signal is divided into 40 sub-signals (12,000 data points in total, to verify the effectiveness of the proposed personalized fault

Please cite this article as: X. Liu, H. Huang and J. Xiang, A personalized diagnosis method to detect faults in gears using numerical simulation and extreme learning machine,
Knowledge-Based Systems (2020) 105653, https://doi.org/10.1016/j.knosys.2020.105653.
10 X. Liu, H. Huang and J. Xiang / Knowledge-Based Systems xxx (xxxx) xxx

Fig. 13. The faults of gears in the experimental operation: (a) a cracked teeth in the driving gear, (b) a peeled teeth in the driving gear, (c) a broken teeth in the
driving gear, (d) a broken teeth in the driven gear.

diagnosis method for gears. Fig. 13 shows the fault gears in the classification accuracy ratios of 9 time-domain indicators
the experimental operation, include a cracked teeth in driving and 14 time–frequency domain indicators for all fault type are
gear, a peeled teeth in the driving gear, a broken teeth in the obviously lower than those of the usage of 23 indicators.
driving gear, and a broken teeth in the driven gear. The four fault In order to illustrate the generalized ability of the present
gears and two health gears generate the five fault types of gear personalized diagnosis method, SVM is used as classifier with the
transmission system (T1 , T2 , T3 , T4 , and T5 , as shown in Fig. 11). same fault samples. In the process of classification, the adjustable
The measured vibration signals are collected from each fault type, parameters of SVM are selected as P1 = 1.05 and C = 10 [45]. Us-
and the corresponding filtered signals are shown in Fig. 14. To ing the SVM toolkit trained by Franc et al. [46], the classification
keep the same length with FEM simulations, the length of each results are listed in Table 8. As shown in Table 8, the classification
measured signal is the same 12,000 data points. Similar to the accuracy ratios are small lower than those using ELM (as shown
fault training samples generation in Section 3.2, the testing sam- in Table 6). However, the corresponding relative errors are vary-
ples (Unknown fault type of physical gear transmission system) ing from 5.8% to 11.1%, nearly same to ELM. The reason is the
will be constructed. precision of SVM model is somewhat lower than ELM to classify
Five types of faults T1 , T2 , T3 , T4 , and T5 are labeled from 1 fault types of gear faults. The comparison investigations testify
to 5, respectively. In the process of classification, the adjustable that the FEM simulation can activate many AI methods to judge
parameters of ELM are selected as L = 25. Based on the method the fault types of gear transmission systems.
of ELM, the five types of faults in gears are classified and the
final classification results are shown in Table 6. The classification 4. Conclusion
accuracy ratios of fault type T1 , T2 , T3 , T4 , and T5 are 85%, 90%,
92.5%, 90% and 85% respectively. We can see that the classification Fault samples of gears are difficult to be acquired from real-
accuracy ratios are not agreeable. To make a fair comparison, world running gear transmission systems. Numerical simulations
the classification accuracy using the measured signal alone (the using FEM provide a possible way to obtain the fault samples for
training and testing samples are selected from the same mea- every type of gear faults under any running conditions. With the
sured signals) is also given using the same ELM parameters L = development of FEM and the corresponding modeling updating
25. From Table 6, the relative errors are varying from 2.6% to technique, it is obviously possible to build up agreeable numer-
10.5%. The comparison investigations testify that the proposed ical simulation models to simulate the gear faults of running
personalized fault diagnosis method is feasible for judge the fault gear transmission systems in real-world. To obtain enough fault
types of gear transmission systems. samples, the concept of the FEM simulation-based personalized
Moreover, to illustrate the effectiveness of using the 23 in- diagnosis is proposed to classify faults in gears using ELM. Nu-
dicators to generate fault training samples, the 9 time-domain merical simulations and experimental investigations show that
indicators (xstd , xp , xske , xkur , xrms , PF , xsra , SF , IF , as shown in the fault types of gears can be distinguished effectively. More
Table 2) and 14 time–frequency domain indicators (WPDi , i = specifically, the classification accuracy ratios of gear states in a
1, 2, . . . , 8 and EEMDi , i = 1, 2, . . ., 6, as shown in Table 2) are cracked teeth of driving gear, a peeled teeth of driving gear, a
further employed to make comparisons, and the corresponding broken teeth of driving gear, a peeled teeth of driving gear and
classification results are listed in Table 7. Compared with Table 6, a broken teeth of driven gear, a broken teeth of driving gear

Please cite this article as: X. Liu, H. Huang and J. Xiang, A personalized diagnosis method to detect faults in gears using numerical simulation and extreme learning machine,
Knowledge-Based Systems (2020) 105653, https://doi.org/10.1016/j.knosys.2020.105653.
 X. Liu, H. Huang and J. Xiang / Knowledge-Based Systems xxx (xxxx) xxx 11

Fig. 14. The filtered measured signals in time-domain: (a) fault type T1 (b) fault type T2 , (c) fault type T3 , (d) fault type T4 , (e) fault type T5 .

Table 6
The classification results using the proposed method (23 indicators) and measured fault signals alone.
Fault Training Testing Faults Classification accuracies Classification accuracies Relative error/%
type samples samples labels using the present using the measured
method/% signal alone/%
T1 40 40 1 85 90 5.6
T2 40 40 2 90 92.5 2.7
T3 40 40 3 92.5 95 2.6
T4 40 40 4 90 95 5.3
T5 40 40 5 85 95 10.5

Table 7
The classification results using the 9 time-domain indicators and 14 time–frequency domain
indicators.
Fault Training Testing Faults Classification accuracies Classification accuracies
type samples samples labels using 9 time-domain using 14 time–frequency
indicators/% domain indicators/%
T1 40 40 1 65 50
T2 40 40 2 50 80
T3 40 40 3 80 85
T4 40 40 4 75 65
T5 40 40 5 55 50

Please cite this article as: X. Liu, H. Huang and J. Xiang, A personalized diagnosis method to detect faults in gears using numerical simulation and extreme learning machine,
Knowledge-Based Systems (2020) 105653, https://doi.org/10.1016/j.knosys.2020.105653.
12 X. Liu, H. Huang and J. Xiang / Knowledge-Based Systems xxx (xxxx) xxx

Table 8
The classification results using the SVM.
Fault Training Testing Faults Classification accuracies Classification accuracies Relative error/%
type samples samples labels using the present using the measured signal
method/% alone/%
T1 40 40 1 80 85 5.8
T2 40 40 2 87.5 92.5 5.4
T3 40 40 3 80 90 11.1
T4 40 40 4 90 95 5.2
T5 40 40 5 85 95 10.5

and a broken teeth of driven gear are 85%, 90%, 92.5%, 90% and [11] R.B. Sun, Z.B. Yang, W. Luo, B.J. Qiao, X.F. Chen, Weighted sparse repre-
85%, respectively. It confirms that the proposed personalized fault sentation based on failure dynamics simulation for planetary gearbox fault
diagnosis, Meas. Sci. Technol. 30 (4) (2019).
diagnosis method open a new way to activate the machine learn-
[12] S.Y. Wu, M.J. Zuo, A. Parey, Simulation of spur gear dynamics and
ing methods by using only the FEM simulations. However, this estimation of fault growth, J. Sound Vib. 317 (3–5) (2008) 608–624.
method is still some limitations, such as the different between the [13] J.W. Xiang, Z.S. Jiang, X.F. Chen, A class of wavelet-based rayleigh-euler
FEM simulations and experimental measurement of mechanical beam element for analyzing rotating shafts, Shock Vib. 18 (3) (2011)
systems with uncertain boundary conditions and fitting surfaces, 447–458.
[14] J.W. Xiang, J.Q. Long, Z.S. Jiang, A numerical study using hermitian cubic
the time-consuming of complex machine using FEM simulations, spline wavelets for the analysis of shafts, Proc. Inst. Mech. Eng. 224 (9)
etc. Therefore, it is important to find feasible way to short FEM (2010) 1843–1851.
simulation time and eliminate the influence of uncertainty. Be- [15] J.W. Xiang, Y.T. Zhong, A novel personalized diagnosis methodology using
sides, this method can be applied in other complex mechanical numerical simulation and an intelligent method to detect faults in a shaft,
Appl. Sci.-Basel 6 (12) (2016) 2–19.
systems, such as bearings, rotors, etc. It would be an interesting
[16] S.H. Wang, J.W. Xiang, H.S. Tang, Minium entropy deconvolution based on
research topic needs further in-deep investigations. simulation determined band pass fiter to detect faults in bearings of axial
piston pumps, ISA Trans. 88 (5) (2019) 186–198.
CRediT authorship contribution statement [17] W.X. Ren, H.B. Chen, Finite element model updating in structural dy-
namics by using the response surface method, Eng. Struct. 32 (8) (2010)
2455–2465.
Xiaoyang Liu: Data curation, Visualization, Investigation, Writ- [18] B. Liu, S.F. Ling, On the selection of informative wavelets for machinery
ing - original draft. Haizhou Huang: Validation. Jiawei Xiang: diagnosis, Mech. Syst. Signal Process. 13 (1) (1999) 145–162.
Supervision, Conceptualization, Methodology, Writing - review & [19] S.K. Biswas, P. Milanfar, One shot detection with laplacian object and fast
editing. matrix cosine similarity, IEEE Trans. Pattern Anal. 38 (3) (2016) 546–562.
[20] F.F. Chen, B.P. Tang, T. Song, Multi-fault diagnosis study on roller bearing
based on multi-kernel support vector machine with chaotic particle swarm
Acknowledgments optimization, Measurement 47 (2014) 576–590.
[21] W.Y. Liu, J.G. Han, X.N. Lu, A new gear fault feature extraction method
This work is supported by the National Natural Science Foun- based on hybrid time-frequency analysis, Neural Comput. Appl. 25 (2)
(2014) 387–392.
dation of China (Nos. U1909217, U1709208), Zhejiang Special
[22] G.B. Huang, Q.Y. Zhu, C.K. Siew, Extreme learning machine: a new learning
Support Program for High-level Personnel Recruitment of China scheme of feed forward neural networks, in: International Joint Conference
(No. 2018R52034) and Wenzhou Major Science and Technology on Neural Networks, Vol. 2, IJCNN, IEEE, 2004, pp. 985–990.
Innovation Project of China (2018ZG023, ZG2019018). [23] G.B. Huang, Q.Y. Zhu, C.K. Siew, Extreme learning machine: theory and
applications, Neurocomputing 70 (1–3) (2006) 489–501.
[24] F.M. Hemez, S.W. Doebling, Test-analysis correlation and finite element
References model updating for nonlinear transient dynamics, in: Proceedings of the
17th International Modal Analysis Conference, Vol. 2, IMAC, 1999, pp.
[1] H.D. Shao, H.K. Jiang, F.A. Wang, An enhancement deep feature fusion 1501–1510.
method for rotating machinery fault diagnosis, Knowl.-Based Syst. 119 [25] M. Link, M. Weiland, Damage identification by multi-Model updating in the
(2017) 200–220. modal and in the time-domain, Mech. Syst. Signal Process. 23 (6) (2009)
[2] D.Y. Dou, J.G. Yang. Wang, J.T. Liu, A rule-based intelligent method for fault 1734–1746.
diagnosis of rotating machinery, Knowl.-Based Syst. 36 (2012) 1–8. [26] T.K. Hasselman, M.C. Anderson, W.S. Gan, Principal components analysis
[3] Z.X. Wei, Y.X. Wang, S.L. He, J.D. Bao, A novel intelligent method for bearing for nonlinear model correlation, updating and uncertainty evaluation, in:
fault diagnosis based on affinity propagation clustering and adaptive Proceedings of the 16th International Modal Analysis Conference, Vol. 1,
feature selection, Knowl.-Based Syst. 116 (2017) 1–12. IMAC, 1998, pp. 644–651.
[4] Y.X. Wang, Z.J. He, J.W. Xiang, Y.Y. Zi, Application of local mean decom- [27] H. Sarin, M. Kokkolaras, G. Hulbert, P. Papalambros, S. Barbat, R.J. Yang,
position to the surveillance and diagnostics of low-speed helical gearbox, Comparing time histories for validation of simulation models: error
Mech. Mach. Theory 47 (2012) 62–73. measures and metrics, Trans. ASME, J. Dyn. Syst. 132 (6) (2010) 173–179.
[5] X.L. Zhang, B.J. Wang, X.F. Chen, Intelligent fault diagnosis of roller bearings [28] J.L. Zapico-Valle, R. Alonso-Camblor, A new method for finite element
with multivariable ensemble-based incremental support vector machine, model updating in structural dynamics, Mech. Syst. Signal Process. 24 (7)
Knowl.-Based Syst. 89 (2015) 56–85. (2010) 2137–2159.
[6] J. Seshadrinath, B. Singh, B.K. Panigrahi, Vibration analysis based interturn [29] B. Yang, J. Shim, Practical datasets for similarity measures and their
fault diagnosis in induction machines, IEEE Trans. Ind. Inform. 10 (1) (2014) threshold values, J. Soc. e-Bus. Stud. 18 (2013) 97–105.
340–350. [30] G.B. Huang, Learning capability and storage capacity of two-hidden-layer
[7] Y.X. Wang, Z.X. Wei, J.W. Yang, Feature trend extraction and adaptive feed forward networks, IEEE Trans. Neural Netw. 14 (2) (2003) 274–281.
density peaks search for intelligent fault diagnosis of machines, IEEE Trans. [31] D. Serre, Matrices: Theory and Applications, Springer-Verlag, New York,
Ind. Inform. 15 (1) (2019) 105–115. 2002.
[8] Z.Y. He, H.D. Shao, X.Y. Zhang, J.S. Cheng, Y. Yang, Improved deep [32] S. Moaveni, Finite Element Analysis Theory and Application with ANSYS,
transfer auto-encoder for fault diagnosis of gearbox under variable Prentice Hall, Upper Saddle River, New Jersey, 1999.
working conditions with small training samples, IEEE Access 7 (2019) [33] G.H. Benedict, B.W. Kelley, Instantaneous coefficients of gear tooth friction,
115368–115377. ASLE Trans. 4 (1) (1961) 59–70.
[9] Z.Y. He, H.D. Shao, P. Wang, J. Lin, J.S. Cheng, Y. Yang, Deep transfer multi- [34] J. Wei, P. Gao, X.L. Hu, W. Sun, J. Zeng, Effects of dynamic transmission
wavelet auto-encoder for intelligent fault diagnosis of gearbox with few errors and vibration stability in helical gears, J. Mech. Sci. Technol. 28 (6)
target training samples, Knowl.-Based Syst. (2019). (2014) 2253–2262.
[10] Y. Wei, M.Q. Xu, Y.B. Li, W.H. Huang, Gearbox fault diagnosis based on local [35] A. Al-Shyyab, A. Kahraman, Non-linear dynamic analysis of a multi-
mean decomposition, permutation entropy and extreme learning machine, mesh gear train using multi-term harmonic balance method: period-one
J. Vibroeng. 18 (3) (2016) 1459–1473. motions, J. Sound Vib. 279 (1–2) (2005) 417–451.

Please cite this article as: X. Liu, H. Huang and J. Xiang, A personalized diagnosis method to detect faults in gears using numerical simulation and extreme learning machine,
Knowledge-Based Systems (2020) 105653, https://doi.org/10.1016/j.knosys.2020.105653.
 X. Liu, H. Huang and J. Xiang / Knowledge-Based Systems xxx (xxxx) xxx 13

[36] C.H. Venner, Y.H. Wijnant, Validation of EHL contact predictions under time [42] R. Ma, Y. Chen, Q. Cao, Research on dynamics and fault mechanism of spur
varying load, Proc. Inst. Mech. Eng. J 219 (4) (2005) 1–13. gear pair with spalling defect, J. Sound Vib. 331 (9) (2012) 2097–2109.
[37] P. Dietl, J. Wensing, G.C. van Nijen, Rolling bearing damping for dynamic [43] S. Jia, I. Howard, Comparison of localized spalling and crack damage from
analysis of multi-body systems-experimental and theoretical results, Proc. dynamic modelling of spur gear vibrations, Mech. Syst. Signal Process. 20
Inst. Mech. Eng. K 214 (1) (2000) 33–43. (2) (2006) 332–349.
[38] T.A. Harris, M.N. Kotzalas, Advanced Concepts of Bearing Technol- [44] Z. Wan, H. Cao, Y. Zi, W.P. He, Z.J. He, An improved time-varying mesh
ogy:Rolling Bearing Analysis, fifth ed., Taylor & Francis Group-CRC Press, stiffness algorithm and dynamic modeling of gear-rotor system with tooth
New York, 2006. root crack, Eng. Fail. Anal. 42 (5) (2014) 157–177.
[39] D. Dowson, G.R. Higginson, Elastic-Hydrodynamic Lubrication: The Funda- [45] J.W. Xiang, M. Liang, Y.M. He, Experimental investigation of frequency-
mentals of Roller and Gear Lubrication, Pergamon Press, London, 1966, p. based multi-damage detection for beams using support vector regression,
235. Eng. Fract. Mec. 131 (2014) 257–268.
[40] S.Y. Chen, C. Kung, J.C. Hsu, Dynamic analysis of a rotary hollow shaft with [46] V. Franc, V. Hlavác, Statistical Pattern Recognition Toolbox for Matlab,
hot-fit part using contact elements with friction, Trans. Can. Soc. Mech. Center for Machine Perception, Czech Technical University, Prague, Czech
Eng. 35 (3) (2011) 461–474. Republic, 2004.
[41] Q. Chen, H. Li, Y.G. Yin, Y.L. Chen, Y.M. Wu, Normal contact damping model
of tooth surface based on fractal theory, CJAM 35 (6) (2018) 1353–1360.

Please cite this article as: X. Liu, H. Huang and J. Xiang, A personalized diagnosis method to detect faults in gears using numerical simulation and extreme learning machine,
Knowledge-Based Systems (2020) 105653, https://doi.org/10.1016/j.knosys.2020.105653.