Wear 376-377 (2017) 1227–1233

Contents lists available at ScienceDirect

Wear
journal homepage: www.elsevier.com/locate/wear

Parameter estimation and remaining useful life prediction of

lubricating oil with HMM
Ying Du a,b, Tonghai Wu a,n, Viliam Makis b
a
Key Laboratory of Education Ministry for Modern Design and Rotor-Bearing System, Xi'an Jiaotong University, Xi'an, Shaanxi 710049, China
b
Department of Mechanical and Industrial Engineering, University of Toronto, Toronto, Ontario, Canada, M5S 3G8

art ic l e i nf o a b s t r a c t

Article history: Lubricating oil plays a vital role in the full life-span performance of the machine. Lubricating oil dete-
Received 2 September 2016 rioration, which leads to the attenuation of oil performance and severe wear afterwards, is a slow de-
Received in revised form grading process, which can be observed by condition monitoring, but the actual degree of the oil de-
23 November 2016
gradation is often very difficult to examine. The main purpose of lubricating oil degradation prediction is
Accepted 25 November 2016
to estimate the failure time when the oil no longer fulfills its functions. We suppose that the state process
evolution of lubricating oil degradation can be modeled using a hidden Markov model (HMM) with three
Keywords: states: healthy state, unhealthy state, and failure state. Only the failure state is observable. While the
Lubricating oil deterioration lubricating oil is in service, vector data that are stochastically related to the deterioration state are ob-
Hidden Markov model
tained through on-line condition monitoring by an OLVF (On-line Visual Ferrograph) sensor at regular
Parameter estimation
sampling epochs. A method of Time Series Analysis (TSA) is applied to the healthy portions of the oil data
Remaining useful life
histories to get the residuals as the observable process containing partial information to fit the hidden
Markov model. The unknown parameters of the fitted hidden Markov model are estimated by the Ex-
pectation-Maximization (EM) algorithm. The remaining useful life (RUL) of lubricating oil can be eval-
uated through explicit formulas of the characteristics such as the conditional reliability function (CRF)
and mean residual life (MRL) function in terms of the posterior probability.
& 2017 Elsevier B.V. All rights reserved.

1. Introduction Wear debris level in lubricating oil has been proved to be one of
the most common degradation features to evaluate lubricating oil
Lubricating oil is used to reduce wear and friction from the degradation [2]. Relative wear debris concentration has been ob-
mobile components, eliminate contamination, remove heat from tained from an image captured by an on-line sensor (OLVF) at a
friction surfaces, and avoid machine failure and reduce the cost for sampling epoch [4]. By this, the wear debris presenting in a lu-
unscheduled maintenance afterwards. Therefore, lubricating oil bricating oil sample can be categorized as a large and small group
condition should be monitored and the oil should be replaced according to their sizes by controlling the oil flow rate and magnet
regularly to extend the period when the machine is in good state field intensity [4]. When the machine is in operation, wear debris
[1]. Recently, condition monitoring (CM) of lubricating oil has at- accumulate in the lubricating oil and the concentration increases,
tracted a considerably attention in research and it plays a vital role which leads to the lubricating oil degradation [5].
in industries [2]. The oil data obtained from CM have been used to Although analysis on wear debris in lubricating oil has been
assess the actual condition of the operating machine in [3], but to utilized in practice for many years to estimate machine condition,
little work has been done using statistical approaches to model
our knowledge, HM models for lubricating oil deterioration and
and analyze oil data for the purpose of assessing the lubricating oil
replacement when the machine is in the healthy state have not
degradation and predicting its remaining useful life. The predic-
been developed in the literature. Taking into account that the time
tion with the CM data can be obtained by the conditional relia-
period when a machine is in the healthy state is usually con-
bility function (CRF) and mean residual life (MRL) function [3],
siderably longer than the length of time between oil replacements,
which indicates the failure time when the oil cannot fulfill its
it is assumed in this paper that the machine condition is stable and
functions anymore, and should be changed. In condition mon-
will not affect considerably the speed of oil deterioration. itoring area, RUL prediction was applied for particle contaminated
lubricating oil by applying physical models using a particle filter-
n
Corresponding author. ing technique [5], as well as the application on rotational bearings
E-mail address: [email protected] (T. Wu). with two-phase threshold model using Bayesian methods [6].

http://dx.doi.org/10.1016/j.wear.2016.11.047
0043-1648/& 2017 Elsevier B.V. All rights reserved.
1228 Y. Du et al. / Wear 376-377 (2017) 1227–1233

To our knowledge, no statistical models have been developed in

the literature which could be applied to the RUL prediction of
lubricating oil.
In this paper, we present a statistical approach to predict the
RUL of lubricating oil subject to deterioration using oil data ob-
tained by an OLVF sensor. Statistical methods of deteriorating
systems, which are utilized in industry to model the degradation
process for early fault detection using CM information, are cate-
gorized into one of three main approaches: the proportional ha-
zard modeling (PHM) [7], stochastic recursive filtering [8], and
hidden Markov modeling (HMM) [9]. HMM, which has been
proved to be efficient in gradual degradation system modeling
[10,11] and the RUL prediction [11], and applied in many areas
such as speech recognition, econometrics and condition-based
maintenance [12], is employed in this paper with two types of
stochastic processes, namely hidden state process and observation
vector process in order to model the degradation process and
predict the RUL of lubricating oil. The state and observation
parameter estimations of the fitted HMM under partial observa-
tions of lubricating oil deterioration can be obtained by the Ex-
pectation-Maximization (EM) algorithm [11].
The residuals are obtained using a vector autoregressive (VAR)
model as the observation process in the hidden Markov frame-
work. Once the parameters of the HMM are estimated, we use the
explicit formulas for the conditional reliability function (CRF) and
the mean residual life (MRL) function in terms of the posterior
probability [13], which can be used for RUL prediction of lu-
bricating oil. The CRF indicates the probability that the oil can
survive during a period of time and has not failed yet, and the MRL
can be calculated by using the posterior probability [14]. It is very
new in the Tribology area to apply the EM algorithm and HMM to
model the degradation process and estimate the RUL of lubricating Fig. 1. HMM-based procedure for RUL prediction.
oil focusing on the lubricating oil condition.
The HMM-based procedure is shown in Fig. 1. The rest of the
paper is organized as follows. In Section 2, a VAR model is fitted to
the real 2-dimensional oil data from an OLVF sensor collected at
regular time epochs, and the residuals for both the healthy and
unhealthy portions of the oil data histories are obtained. In Section
3, the state and residual process are modeled as an HMM, and the
estimation procedure of the unknown parameters is developed
using the EM algorithm. In Section 4, the formulas for the condi-
tional RF and MRL have been applied, which can be used to predict
the RUL of lubricating oil. Finally, the conclusions and future re-
search are summarized in Section 5.

2. Vector autoregressive modeling and computation of

residuals

The real condition monitoring data were obtained for the de-
tection of lubricating oil deterioration from a four-ball test rig [15]
Fig. 2. Wear debris concentration of small and large particles.
in order to predict the RUL of the lubricating oil afterwards. During
the operational life of the tribo-pairs, oil data were collected every
▵ = 4 minutes by an OLVF sensor, which is an on-line ferrographic Table 1
Working conditions for the test.
sensor based on Image Technology, and it provided wear debris
concentrations that came from the direct wear during these Test no. Load/N Rotated rate/rpm Time/min Downtime duration/min
4 minutes. The total number of data histories recorded is 27, which
consist of N ¼ 11 failure histories and M¼ 16 suspension histories. 1 1500 1000 360 0
2 1500 1000 360 240
The failure history is defined as the history that ends with ob-
3 2000 1000 210 720
servable failure, which indicates that the lubricating oil is out of 4 2000 1000 240 60
use at that moment, and the suspension history is defined as the 5 2000 2000 60 480
history that ends when the lubricating oil is still in operation and
has not lost its functions.
To avoid over-parameterization, we use the 2-dimensional history is given in Fig. 2, and the working conditions are listed in
monitoring data consisting of small wear particles and large wear Table 1. IPCA shows three stages including run-in, normal, and
particles obtained from the OLVF sensor for analysis. A typical data severe stages, which agrees with the typical “Bathtub Curve”.
 Y. Du et al. / Wear 376-377 (2017) 1227–1233 1229

At the beginning of the sampling, the tribo-pairs are in the run-in ^ Sp

stage, which should be removed from the data set in the modeling C=
T − (2p + 1) (8)
process. After the run-in period, the tribo-pairs are operating un-
N+M
der normal conditions, which are regarded as the healthy portion where T= (ti − p) is the total number of the available
∑i = 1
of the data history. Severe wear dramatically appears at 190th ^ ^
healthy portions of the data histories, and Sp = (W − VA)′(W − VA)
sampling epoch after 760 operational minutes, which is regarded is the residual sum of squares matrix. The estimation of model
as the beginning of the unhealthy portion of the data history. For order p ∈ N is obtained by testing H0: Φp = 0 against Ha: Φp ≠ 0
our experiment, the four-ball test rig was stopped and restarted using the likelihood ratio statistic given by
when working conditions changed, which explains the peak points
at the beginning of each section in Fig. 2. Besides, this particular oil ⎛ 1 ⎞ ⎛ det (Sp ) ⎞
Mp = − ⎜ T − 2p − 1 − ⎟ ln ⎜ )⎟
sample failed at 290th sampling epoch after 1156 operational ⎝ 2 ⎠ ⎝ det (Sp − 1 ⎠ (9)
minutes.
As described by Kim and Makis [11], one should first fit a model For our real 2-dimensional oil data, we obtained M4 = 54.33
using only the healthy portions of the oil data histories that ac- and M5 = − 25.64 . From the chi-square distribution with 2 de-
2 2
count for cross and autocorrelation in the data. Since there is no grees of freedom and α = 0.05, χ2,0.05 = 5.99. Since M4 > χ2,0.05 and
agreed upon criterion for selecting the ‘optimal’ segmentation, in 2
M5 < χ2,0.05 , we reject H0: Φ4 = 0 and fail to reject H0: Φ5 = 0 .
order to partition a non-stationary time series data, we simply
Therefore, p^ = 4 is an adequate model order, and the VAR model
divide the data histories into two portions (healthy and unhealthy)
parameter estimates are given by
via graphical examination. The purpose of segmentation is to
identify the healthy portions of the data histories so that a sta- ⎛ 0.0974 ⎞ ^ ⎛ 3.4290e − 3 1.2452e − 4 ⎞
p^ = 4, μ^ = ⎜ ⎟, C = ⎜ ⎟,
tionary time series model can be fitted and the residuals can be ⎝ 0.0236 ⎠ ⎝ 1.2452e − 4 1.8464e − 3⎠
computed using the fitted model. For each of the N + M = 27 data ⎛ 0.2492 − 0.0675⎞ ^ ⎛ 0.1887 − 0.1083⎞
histories, the healthy portions of the data histories are denoted as Φ^1 = ⎜ ⎟, Φ 2 = ⎜ ⎟,
⎝ 0.2318 − 0.1982⎠ ⎝ 0.4168 − 0.0157⎠
{ z1i, z2i , … , ztii , i = 1, 2, … , N + M . } ⎛ 0.0011 0.0155⎞ ^ ⎛ 0.0143 − 0.0042⎞
The healthy portions of the data histories are assumed to follow Φ^ 3 = ⎜ ⎟, Φ4 = ⎜ ⎟
⎝ − 0.0045 0.0161⎠ ⎝ 0.0129 0.0026 ⎠
a common stationary vector auto-regressive (VAR) process [16]
given by ^
Using estimates γ^ = (μ^ , p^ , Φ^1, Φ^ 2, Φ^ 3, Φ^4, C), we define the re-
p sidual process (Yn ▵ ) by
(Z n − μ 0 ) − ∑ Φr (Zn− r − μ0 ) = εn, n ∈ Z →
r=1 (1) YnΔ :=Z n − E γ^ (Z n | Z n − 1) (10)
where εn are i.i.d. N2 (0, C), the model order p ∈ N , the auto- →
where Z n − 1 = (Z1, Z2, … , Zn − 1).
correlation matrices Φr ∈ R2 × 2, and the mean and covariance
p^
model parameters μ0 ∈ R2 and C ∈ R2 × 2. All the model parameters For n > p^ , Yn ▵ = Zn − (μ^ + ∑r = 1 Φr Zn − r ), and for n < p^ , we re-
are unknown and need to be estimated. cursively compute Yn ▵ using the Kalman filter, applying the pro-
In order to get the standard form of VAR Model, set cedure used in reference [17]. Therefore, the residuals can then be
p
μ = μ0 − ∑r = 1 Φr μ0 . Then, Eq. (1) can be written as computed for both the healthy and unhealthy portions of all
p monitoring oil data histories, provided graphically in a 2-dimen-
Zn = μ + ∑ Φr Zn− r + εn, n ∈ Z sional scatter plot shown in Fig. 3. The crosses in blue are residuals
r=1 (2) computed from the healthy portions of the oil data histories, and
Therefore, the regression representation W = VA + E is ob- the circles in red are residuals computed from the unhealthy
tained for the observed healthy portions of the data histories portions of the oil data histories.
The statistical test of the normality assumption was performed
{ }
z1i, z2i , … , ztii , i = 1, 2, … , N + M , where
using the Henze-Zirkler Multivariate Normality Test [18] with a given
W = ⎡⎣ ztNN++ M
M
, …, z pN++1M , …, zt11 , …, z p1 + 1⎤⎦ ′ (3) significance level α = 0.05, and the obtained results are shown in

⎡ 1 z N+M M ⎤
… ztNN++ M
⎢ tN + M − 1 − p⎥
⎢1 ⋮ ⋮ ⋮ ⎥
⎢ N+M
⎥
⎢ 1 z p … z1N + M ⎥
V = ⎢1 ⋮ ⋮ ⋮ ⎥
⎥
⎢
⎢ 1 zt1 − 1 … zt1− p ⎥
1
1
⎢ ⎥
⎢1 ⋮ ⋮ ⋮ ⎥
⎢1 z 1p z11 ⎥⎦
⎣ … (4)

A = ⎡⎣ μ , Φ1, …, Φp ⎤⎦ ′ (5)

E = ⎡⎣ εtNN ++ MM , …, εpN++1M , …, εt11 , …, εp1+ 1⎤⎦ ′ (6)

Using the method of least squares [16], the estimates for A and
C are given by
Fig. 3. Scatter plot for all the residuals. (For interpretation of the references to color
^
A = (V′V)−1V′W (7) in this figure, the reader is referred to the web version of this article.)
1230 Y. Du et al. / Wear 376-377 (2017) 1227–1233

Table 2 oil data histories as { F1, F2, …FN }, with the form of
P-value of the residual normality test.
( )
Yi = y1i , y2i , …yTi , (i = 1, 2, …N ) for each failure history Fi. The
i
Test Healthy data set Unhealthy data set observable failure time ξi for each history is
ξi = ti, (Ti ▵ < ti ≤ (Ti + 1)▵). The sampling history Yi consists of the
Normality(Henze-Zirkler) 0.8232 0.2983
residuals yti ∈ R2, t ≤ Ti , calculated from the collection of oil data
until the lubricating oil fails at time ti.
Table 2, which show that the calculated residuals follow the multi- Similarly, we denote the M ¼16 suspension oil data histories as
variate normal distribution and satisfy the assumption of multi-
variate normality for both healthy and unhealthy data sets [19].
{ S1, S2, …SM }, with the form of Yj = ( y , y , …y ), (j = 1, 2, …M )
1
j
2
j j
Tj

for each suspension history Sj. The failure time ξj is unobservable,

In the next section, we will apply the HMM framework using
where ξj = t j > Tj ▵ . The sampling history Yj consists of the re-
the obtained residuals as the observation process, and estimate the
state and observation parameters by applying the EM algorithm. siduals ytj ∈ R2, t ≤ Tj , calculated from the collection of oil data but
the lubricating oil is still in operation at time tj, and has not failed
yet.
Let O = { F1, F2, …FN , S1, S2, …SM } represents a vector of all the
3. Hidden Markov modeling and parameter estimation using
observable lubricating oil data, and L (λ , θ|O ) be the associated
EM algorithm
likelihood function, where λ = (λ 01, λ 02, λ12 ) are the unknown state
parameters and θ = (μ0 , μ1, Σ 0, Σ1) are the unknown observation
Lubricating oil deterioration is a slow degrading process, and it
parameters. Expectation-Maximization (EM) algorithm is adopted
is difficult to be examined. In order to investigate the RUL of the
by iteratively maximizing the so-called pseudo likelihood function
lubricating oil, a degradation model of the process should be
to estimate the unknown parameters.
considered. We assume that the deterioration state of lubricating
EM algorithm consists of two steps: E-step and M-step, where
oil characterized by the two monitoring indexes follows a con-
E-step computes the expectation of the associated likelihood
tinuous time homogeneous Markov process { Xt : (t ∈ R +) }, with
function and M-step obtains the maximization of the unknown
the state space S = { 0, 1} ∪ { 2}. In general, states 0 and 1 are
states and observation parameters. The E-steps and M-steps are
always unobservable, representing the healthy and unhealthy
operational states of lubricating oil, respectively. Only the failure repeated until the Euclidean distance (λ*, θ*) − λ^ , θ^ < ε , where ( )
state (state 2) is observable. It is assumed that the lubricating oil λ^ , θ^ are the estimates of the unknown parameters from the former
starts in a “good as new” state, i.e. X0 = 0, and the transition rate M-step, λ*, θ* are the new parameters estimated from the current
matrix is given by M-step, and ε is the selected stopping criterion value. More spe-
⎛ − (λ 01 + λ 02 ) λ 01 λ 02 ⎞ cifically, the EM algorithm works as follows.
⎜ ⎟ E-step: the pseudo likelihood function is defined by
Λ=⎜ 0 − λ12 λ12 ⎟
⎜ ⎟
⎝ 0 ⎠
0 0 (11)
( )
Q λ , θ|λ^, θ^ :=E λ^ , θ^ (ln L (λ , θ|C )|O) (13)
where λ 01, λ 02, λ12, ∈ (0, + ∞) are unknown model parameters.
Let ξ = inf { t ∈ R +: Xt = 2} be the observable failure time when { }
where C = F¯1, …F¯N , S¯1, …S¯M represents the complete lubricating
the lubricating oil no longer fulfills its functions. The residual oil data histories.
process (Yn ▵: n ∈ N) defined in Eq. (10) is assumed to be con- M-step: compute λ*, θ* by
ditionally independent given the state of the lubricating oil, and
λ*, θ * ∈ argmaxQ (λ , θ|λ^, θ^)
for each n ∈ N , we assume that Yn ▵ , conditional on Xn ▵ = x , has λ, θ (14)
bivariate normal distribution N2 (μx , Σ x ), where x ¼0,1, with the
density given by Explicit formulas for the computation of the pseudo likelihood
function defined in Eq. (13) and the unique maximums of the
1 ⎛ 1 ⎞ unknown state and observation parameters can be found in
fYnΔ | XnΔ (y|x) = exp ⎜ − (y − μ x )′Σ x−1 (y − μ x ) ⎟
(2π )2det (Σ x ) ⎝ 2 ⎠ (12) [11,17].

where μx ∈ R2 and Σ x ∈ R2 × 2, (x = 0, 1) are unknown parameters of

With the stopping criterion (λ*, θ*) − λ^ , θ^ < 10−4 , we ob- ( )
the observation process. tained the results shown in Table 3.
When the oil fails, P (Yn ▵ = η|Xn ▵ = 2) = 1, η ∉ R2 is the failure Thus, the deterioration state of lubricating oil is modeled as a
signal obtained by the OLVF sensor. We denote the N ¼11 failure continuous time homogeneous Markov chain (Xt : t ∈ R +) with

Table 3
Iterations of the EM algorithm.

Parameters Initial values First iteration Second iteration Final iteration

λ01 0.01 0.0118 0.0132 0.0189

λ02 0.0005 0.00056 0.00059 0.000047
λ12 0.5 0.4884 0.4572 0.1812
μ0
()0
0 ( −0.0011
4.31e −4 ) ( −9.64e −4
3.99e −4 ) ( 1.25e −4
−8.16e −4 )
μ1
( )
0.05
0.02 ( 0.0036
−0.0020 ) ( 0.0030
−0.0067 ) ( 0.0197
−0.0131 )
Σ0
( )
0.02 0.02
0.02 0.05 ( 0.003088 0.000266
0.000266 0.001828 ) ( 0.003068 0.000264
0.000264 0.001829 ) ( 0.002767 0.000288
0.000288 0.001680 )
Σ1
( )
0.03 0.02
0.02 0.05 ( 0.003893 0.000823
0.000823 0.001858 ) ( 0.003604 0.001222
0.001222 0.001984 ) ( 0.002966 0.002096
0.02096 0.002077 )
Q  81.3594  168.3331  169.7593  329.8362
Time/s 0.25 0.4063 0.5313 2.3281
 Y. Du et al. / Wear 376-377 (2017) 1227–1233 1231

three states { 0, 1, 2}. The transition rate matrix is given by

⎛ − 0.01894 0.018893 0.000047⎞
⎜ ⎟
Λ=⎜ 0 − 0.18116 0.181160⎟
⎝ 0 0 0 ⎠ (15)

and the observations Yn ▵ conditional on Xn ▵ = x, x = 0, 1, have

bivariate normal distribution N2 (μ0 , Σ 0 ) for healthy portions of the
oil data histories and N2 (μ1, Σ1) for unhealthy portions, where
⎛ 0.000125 ⎞ ⎛ 0.002767 0.000288⎞
μ0 = ⎜ ⎟, Σ 0 = ⎜ ⎟
⎝ − 0.000816⎠ ⎝ 0.000288 0.001680 ⎠
⎛ 0.0197 ⎞ ⎛ 0.002966 0.002096⎞
μ1 = ⎜ ⎟ , Σ1 = ⎜ ⎟
⎝ − 0.0131⎠ ⎝ 0.002096 0.002077⎠ (16)

In the next section, we use the estimated parameters of the

model to compute the CRF and MRL functions for RUL prediction of
lubricating oil based on the posterior probability that the oil is in
state 1 (warning state), which is not observable. Fig. 4. The residual observation process of real failure history. (For interpretation of
the references to color in this figure, the reader is referred to the web version of this
article.)

4. CRF and MRL functions for RUL prediction

Explicit formulas of the CRF and the MRL functions of the g (yn|XnΔ = 1, ξ > nΔ, yn − 1) × P (XnΔ = i|ξ > nΔ, yn − 1)
Πn =
model proposed in Section 3 for the RUL prediction of lubricating ∑j g (yn|XnΔ = j, ξ > nΔ, yn − 1) × P (XnΔ = j|ξ > nΔ, yn − 1) (19)
oil will be considered in this section, and as it is shown in [3], both
formulas are functions of the posterior probability of the lu- where
bricating oil being in the warning state (state 1). ⎧ f (y |μ , Σ 0 ), i = 0
⎪
n 0
We assume that the deteriorating state process of lubricating g (yn|XnΔ = i, ξ > nΔ, yn − 1) = ⎨ ⎪
oil is described by a continuous-time homogeneous Markov chain ⎩ f (yn|μ1, Σ1), i = 1 (20)
(Xt : t ∈ R +), with state space Z = { 0, 1, 2}, where state 0 denotes
Therefore, we have
the state that the lubricating oil is working in a healthy condition,
state 1 denotes the state that the lubricating oil is operating in a (P01 (Δ)(1 − Πn − 1 ) + P11 (Δ) Πn − 1 )
Πn =
warning condition, and state 2 represents the failure or absorbing f (yn |μ0 , Σ 0 )
P00 (Δ)(1 − Πn − 1 ) + (P01 (Δ)(1 − Πn − 1 ) + P11 (Δ) Πn − 1 )
state of lubricating oil, which indicates that it requires an im- f (yn |μ1, Σ1 ) (21)
mediate replacement. The transition rate matrix of the state pro-
where Pij are given in Eq. (17), and the ratio of normal densities has
cess is given by Eq. (15). Moreover, lubricating oil starts in state
0 and runs on a continuous basis. While the lubricating oil is in the following representation
service, the residual observation process (Yn ▵: n ∈ N), which is ⎛ 1 ⎞
obtained from Eq. (10) using the collected data through condition exp ⎜ − (yn − μ 0 )′Σ 0−1 (yn − μ 0 ) ⎟
f (yn|μ 0 , Σ 0 ) Σ1 ⎝ 2 ⎠
monitoring at equidistant sampling times ▵, 2▵, …n▵ , for = ·
f (yn|μ1, Σ1) Σ0 ⎛ 1 ⎞
▵ = 40 min , has a state-dependent multivariate normal distribu- exp ⎜ − (yn − μ1)′Σ1 (yn − μ1) ⎟
− 1
⎝ 2 ⎠ (22)
tion defined in Eqs. (15,16). By solving the Kolmogorov backward
differential equations [20], the transition probability matrix for the For more details, see reference [17].
state process Xt has the following form: Suppose that at decision epoch n, the lubricating oil has not
failed, i.e. ξ > n▵ . For any t ∈ [0, ▵], the conditional reliability

⎡ λ (e −λ12 t − e −(λ 01+ λ 02 ) t ) λ (e −λ12 t − e −(λ 01+ λ 02 ) t ) ⎤

⎢ e −(λ 01+ λ 02 ) t 01 1 − e −(λ 01+ λ 02 ) t − 01 ⎥
P (t ) = Pij (t ) = ⎢ ⎥
λ01 + λ02 − λ12 λ01 + λ02 − λ12
⎢ 0 e −λ12 t 1 − e −λ12 t ⎥
⎢ ⎥
⎣ 0 0 1 ⎦ (17)

where the transition probabilities Pij (t ) = P (Xt = j|X0 = i ), i, j ∈ { 0, 1}. function (CRF), which denotes the probability that the lubricating
We will show that the CRF and MRL functions can be expressed oil will not fail by n▵ + t , is defined as
in terms of the posterior probability statistic Πn, denoting the
posterior probability that the lubricating oil is in warning state R (t|Πn ) = P (ξ > n▵ + t|ξ > n▵, Y▵, …, Yn ▵, Πn )
(state 1) given all available information until time n▵ , which is
= P00 (t ) + P01 (t ) + (P11 (t ) − P00 (t ) − P01 (t )) Πn (23)
defined as
Πn = P (XnΔ = 1|ξ > nΔ, YΔ, Y2Δ, …, YnΔ ) (18) Using the CRF in Eq. (23), the MRL function at the nth epoch can
where Π0 = P (X0 = 1) = 0. Using Bayes’ rule for n ≥ 1, we have be obtained by the following formula:
1232 Y. Du et al. / Wear 376-377 (2017) 1227–1233

changed, and the full inspection and a subsequent oil replacement

action should be taken.

5. Conclusions and future researches

In this paper, we have considered the situation where the lu-

bricating oil degradation is driven by a continuous time homo-
geneous Markov chain and the observation process is represented
by a 2-dimensional on-line monitoring oil data obtained from a
four-ball test rig, where the actual oil states are unobservable
except the failure state. A vector autoregressive model has been
fitted to the healthy portions of all oil data histories classified as
failure histories and suspension histories. The residuals calculated
from the VAR model are then used as the observation process in
the HMM framework, and the state and the observation process
Fig. 5. Posterior probability of the typical failure oil history. parameter estimates of the lubricating oil have been obtained
using the EM algorithm. With the unknown parameters estimated,
RUL prediction of lubricating oil has been developed by deriving
the explicit formulas of the CRF and MRL functions expressed in
terms of the posterior probability for the purpose of future deci-
sion-making.
The HMM considered in this paper assumes that the sojourn
times in both the healthy and warning states are exponentially
distributed, which may not be realistic in some situations. Also,
this paper studies only the degradation process of lubricating oil
when the machine is in healthy condition, and the RUL from the
‘good as new’ state to the oil failure state. In future research, more
general distributions of the sojourn times will be considered, such
as Erlang of phase-type distributions, which will cover more real
situations and provide better estimates of RUL when exponentially
distributed sojourn times are not appropriate.

Acknowledgment
Fig. 6. The conditional reliability function of real failure history.
The financial support for the present research was provided by
→ ∞
the National Science Foundation of China (Grant Nos. 51675403,
μ nΔ = E (ξ − nΔ|ξ > nΔ, Πn ) = ∫0 R (t|Πn ) dt
51275381), which is very much appreciated. Special thanks are also
λ12 + λ 01 + Πn (λ 02 − λ12 ) due to the China Scholarship Council (Grant No. 201506280150) for
=
λ12 (λ 01 + λ 02 ) (24) their financial support during the term of this project.
The CRF and MRL functions have been widely used for RUL
prediction [3]. (See e.g., [3] and the references in that paper for
References
more details).
We apply the above formulas to one of our failure histories
[1] Y. Du, T. Wu, J. Cheng, Age detection of lubricating oil with on-line sensors, in:
with bivariate residual observations, which is shown in Fig. 4. This SENSORS, 2015 IEEE, IEEE, 2015a, pp. 1–4.
failure history ended between 25th and 26th sampling epoch. The [2] J. Zhu, J. Yoon, D. He, B. Qiu, E. Bechhoefer, Online condition monitoring and
remaining useful life prediction of particle contaminated lubrication oil, in:
black solid line denotes the residual observations of small parti- Proceedings of the Prognostics and Health Management (PHM), 2013 IEEE
cles, and the red solid line denotes the residual observations of Conference on, IEEE, 2013, pp. 1–14.
large particles. [3] A. Khaleghei, V. Makis, Model parameter estimation and residual life predic-
tion for a partially observable failing system, Nav. Res. Logist. (NRL) 62 (3)
Fig. 5 plots the posterior probability of lubricating oil data (2015) 190–205.
history presented in Fig. 4. The posterior probability was calcu- [4] T. Wu, J. Mao, J. Wang, J. Wu, Y. Xie, A new on-line visual ferrograph, Tribol.
Trans. 52 (5) (2009) 623–631.
lated using Eqs. (21,22). The plot shows that the lubricating oil [5] J. Zhu, J.M. Yoon, D. He, E. Bechhoefer, Online particle-contaminated lubrica-
started in the healthy state (state 0) at time 0, ran under the tion oil condition monitoring and remaining useful life prediction for wind
healthy condition till sometime between 1st and 15th sampling turbines, Wind Energy 18 (6) (2015) 1131–1149.
[6] N. Chen, K.L. Tsui, Condition monitoring and remaining useful life prediction
epoch, and after 16th sampling epoch, the lubricating oil was using degradation signals: revisited, IIE Trans. 45 (9) (2013) 939–952.
working in the warning state (state 1). [7] M.-Y. You, L. Li, G. Meng, J. Ni, Two-zone proportional hazard model for
equipment remaining useful life prediction, J. Manuf. Sci. Eng. 132 (4) (2010)
Fig. 6 provides the estimated conditional reliability at each 041008.
sampling epoch for this failure oil data history shown in Fig. 4. It [8] X.-S. Si, W. Wang, C.-H. Hu, M.-Y. Chen, D.-H. Zhou, A Wiener-process-based
can be seen that the value of the reliability dropped at the 15th degradation model with a recursive filter algorithm for remaining useful life
estimation, Mech. Syst. Signal Process. 35 (1) (2013) 219–237.
sampling epoch when the state changed to the warning state [9] D. Lin, V. Makis, Filters and parameter estimation for a partially observable
(state 1), and after 25th sampling epoch, the lubricating oil shifted system subject to random failure with continuous-range observations, Adv.
Appl. Probab. (2004) 1212–1230.
to the failure state (state 2). It is shown that there is a high [10] R. Jiang, System availability maximization and residual life prediction under
probability at that time that the state of the lubricating oil has partial observations, (Ph.D. thesis), University of Toronto, 2011.
 Y. Du et al. / Wear 376-377 (2017) 1227–1233 1233

[11] M.J. Kim, V. Makis, R. Jiang, Parameter estimation for partially observable Tribology Conference 2015, Malaysian Tribology Society, vol. 2015, 2015b, pp.
systems subject to random failure, Appl. Stoch. Model. Bus. Ind. 29 (3) (2013) 185–186.
279–294. [16] G.C. Reinsel, Elements of Multivariate Time Series Analysis, Springer Science &
[12] Z.-J. Zhou, C.-H. Hu, D.-L. Xu, M.-Y. Chen, D.-H. Zhou, A model for real-time Business Media, Springer, New York, 2003.
failure prognosis based on hidden Markov model and belief rule base, Eur. J. [17] M.J. Kim, Optimal control and estimation of stochastic systems with costly
Oper. Res. 207 (1) (2010) 269–283. partial information, (Ph.D. thesis), University of Toronto, 2012.
[13] V. Makis, Multivariate Bayesian control chart, Oper. Res. 56 (2) (2008) [18] N. Henze, B. Zirkler, A class of invariant consistent tests for multivariate nor-
487–496. mality, Commun. Stat.-Theory Methods 19 (10) (1990) 3595–3617.
[14] A. Ghasemi, S. Yacout, M.-S. Ouali, Evaluating the reliability function and the [19] J. Yang, V. Make, Dynamic response of residuals to external deviations in a
mean residual life for equipment with unobservable states, IEEE Trans. Reliab. controlled production process, Technometrics 42 (3) (2000) 290–299.
59 (1) (2010) 45–54. [20] G. Grimmett, D. Stirzaker, Probability and Random Processes, Oxford Uni-
[15] Y. Du, T. Wu, J. Cheng, R. Gong, Lubricating oil deterioration on a four-ball test versity Press, Oxford, 2001.
rig via on-line monitoring, in: Proceedings of Malaysian International