Professional Documents
Culture Documents
Measurement
journal homepage: www.elsevier.com/locate/measurement
A R T I C L E I N F O A B S T R A C T
Keywords: This paper aims at the development of a method for monitoring oil debris as well as oil viscosity, using the optical
Oil condition monitoring measurement. An on-line optical oil debris sensor is designed firstly, then a multi-target tracking algorithm based
Optical measurement on Gaussian Mixture Model is custom-developed to compute the properties (velocity, diameter) of wear debris in
Response surface methodology (RSM)
micro-flow channel through images captured by sensor. A mathematical relationship between debris velocity and
Wear debris velocity
Oil viscosity
debris diameter, debris density, oil flow, oil viscosity, oil temperature is statistically established using a three-
level-five-factor box behnken design approach based on response surface methodology analysis. Then the
analysis of variance is used to confirm the significance of the variables and their interactions. Finally, to validate
the mathematical model, an experiment is carried out to compare its performance to that of results obtained
using a pin disc test rig. The results demonstrate the measurement error of debris density and oil viscosity is
6.07% and 7.97%, respectively. The proposed approach has excellent suitability to monitor the oil condition.
* Corresponding authors.
E-mail addresses: jiayinly@nuaa.edu.cn (Y. Liu), rms@nuaa.edu.cn (H. Zuo).
https://doi.org/10.1016/j.measurement.2022.111152
Received 10 January 2022; Received in revised form 23 March 2022; Accepted 1 April 2022
Available online 4 April 2022
0263-2241/© 2022 Elsevier Ltd. All rights reserved.
Z. Liu et al. Measurement 195 (2022) 111152
performance than others, such as, the color, diameter, texture of the F-test of the ANOVA. Finally, an on-line experiment on the pin disc test
wear debris and the color of the oil can be seen directly, as a result, rig is conducted to validate the monitoring effect of the mathematical
substantial research have been dedicated to the wear debris detector model.
[25–27], restoration of wear debris image [28,29], wear debris classi This paper is further organized as follows: Section 2 is the description
fication [30]. Currently, ferrography has been accepted as the most of the on-line optical oil debris sensor and the debris features extraction
common technique for obtaining the morphology information from algorithm; the RSM modeling and the variables effects are given in
debris images of WDA. However, wear debris attracted by the magnetic Section 3; the verification of the proposed method is presented in Sec
force always form into agglomerated chains. Although efforts have been tion 4, followed by discussions in Section 5; and conclusions are
made to separate debris into chains [31], it is still difficult to extract depicted in Section 6.
features of individual particles. Meanwhile, it is very time-consuming
and relies on the experience of the observers or accuracy of the classi 2. Optical sensor and algorithm
fication algorithm, and it cannot monitor the viscosity of the oil, as the
other optical sensors based on image analysis of wear debris. 2.1. On-line optical oil debris sensor
Accordingly, this paper proposes a new methodical measurement
approach for monitoring wear debris properties and oil viscosity A novel method for monitoring the wear debris properties and oil
simultaneously based on WDA using only one optical sensor, by inte viscosity simultaneously based on WDA from the perspective of dynamic
grating dynamic image acquisition and velocities for moving wear image acquisition and velocities for moving wear debris using optical
debris. In this paper, an on-line optical oil debris sensor is designed and a measurement is proposed in this work. Firstly, an on-line optical oil
multi-target tracking algorithm for wear debris velocity measurement is debris sensor is designed to capture the moving wear debris images.
developed. Then, considering the velocity of wear debris in the micro Fig. 1(a) illustrates the main components of the sensor through the
fluidic chip is affected by the size and density of the debris, as well as the stereogram, and Fig. 1(b) provides a detailed view of the schematic di
viscosity and velocity of oil [26]. Meanwhile, the viscosity is seriously agram. Then we can see that it consists of an image acquisition sub
affected by temperature. Hence, a series of tests are designed to develop system to capture the images of wear debris, and a data transmission
a mathematical model and evaluate the statistically significant effect of subsystem to transfer the images. The image acquisition subsystem has
the debris diameter, debris density, oil flow rate, oil viscosity, and several essential components: a pump, a micro-flow channel, a lens, an
applied temperature on the debris velocity using the RSM method. image sensor, a light source. The reflecting light source is coaxial with
Subsequently, the significance of the model equation is assessed by the the optical lens, which helps to see the surface of debris because of the
Fig. 1. (a) Stereogram of on-line optical oil debris sensor. (b) Schematic diagram of on-line optical oil debris sensor.
2
Z. Liu et al. Measurement 195 (2022) 111152
uniform illuminating ray. The pump control oil feed rate in flow channel When the pixel matches the i-th Gaussian component in the corre
from 1 ml/h to 1000 ml/h. So, the wear debris are suspended in the oil sponding background model, set the corresponding Boolean val
and flow with it, and the oil is passed through a specifically engineered ueMi,t = 1 , otherwiseMi,t = 0. And the weight of the Gaussian
flow channel with a low thickness value that enables the picture of the component Wi,t is updated by.
debris to be captured using a camera. The resolution of the camera is
Mi,t = (1 − α)Wi,t− 1 + αMi,t (2)
2048 × 1536 pixel, and its physical size represented by each pixel is
3.45 μm. The flow channel is arranged vertically just beneath the
where α is the learning rate, and α satisfies0 < α < 1, whose size controls
portable micro-optical lens in order to prevent debris from being
the update speed of the background model. According to Equation (2),
deposited in it. Then the captured real-time images of the oil with wear
the greater the matching number of Gaussian distribution, the higher the
debris can be observed in the operating environment via the data
weight. On the other hand, the Gaussian distribution will reduce the
transmission system. The computer is used for image processing in near
weight value if it is not matched. After updating the weight, it is also
real-time through the intelligent algorithm, which calculates the fea
necessary to update theUi,t ,σ2i,t , and the update formulas are presented in
tures of the wear debris in oil. The algorithm is described in the
following subsection of this paper. Equations (3), (4), (5). Where λi,t is determined by the learning rate and
Furthermore, we can also see that there are two other main com the probability density functionρ:
ponents in Fig. 1(a), the fan and the lens holder. The fan is designed to ( )
Ui,t = 1 − λi,t Ui,t− 1 + λi,t Xt (3)
speed up the air circulation inside the sensor, as the heat generated by a
computer working in an enclosed space is enormous, and the fan pre ( ) ( )T ( )
σ 2i,t = 1 − λi,t σ 2i,t− 1 + λi,t Xt − Ui,t Xt − Ui,t (4)
vents the internal temperature of the sensor from becoming too high and
affecting the normal operation of other components. The lens holder is
λi,t = αρ(Xt |Uk , σ k ) (5)
used to maintain the stability of the lens (positioned horizontally) in
operation. It is equipped with precision knobs for up and down, left and
right, forward and backward movement control of the lens. Because the (b) Pixels do not match GMM
lens is a precision device, a slight movement in position can produce a
huge deviation in the imaging effect of the wear debris. Suppose the pixel point does not match all the Gaussian components
in the corresponding background model. In that case, the Gaussian
2.2. Debris features extraction algorithm component with the smallest weight is discarded, and a new component
is added, with the current point’s pixel value serving as the new com
To obtain the movement of wear debris in the oil, it is necessary to ponent’s mean value. The initial variance and weight are set to the
first extract the target debris and their coordinate positions at different default values.
moments from the videos acquired from the optical sensor, then the
speeds of debris can be calculated combining the frame rate of 3 Background model estimation
recording. In this research, the GMM and blob are used to construct
algorithms for wear debris detection, tracking, and velocity calculation. The weight of the Gaussian distribution is normalized once the
parameter is updated, and all Gaussian distributions in the background
2.2.1. Wear debris detection algorithm based on GMM model corresponding to each pixel are organized in descending order of
Because the camera is fixed in the optical sensor and the background priorityWi,t |σ i,t . Background threshold T is then set to indicate a previous
varies slowly, the GMM algorithm is utilized to extract the binary image global probability, which is the percentage of background component in
by separating all the moving debris from the oil background. After an image. Then according to the sorting of Wi,t |σ i,t and threshold T, the
initializing the mean and weight parameters of the first Gaussian dis first B Gaussian components are selected by.
tribution corresponding to each pixel of the first frame of an image, it is ( )
∑ s
divided into five steps: template matching, parameter update, back B = argmint wk > T ; 0.5 < T < 1 (6)
ground model estimation, foreground detection, morphological noise k=1
removal.
where arg is the abbreviation of argument.
1 Template Matching
4 Foreground detection
When a fresh image is reading at time t, each pixel in the background
model is matched with K Gaussian components, as shown in Equation The current image’s real background model has been obtained in the
(1). The pixel point Xt matches the background model when it fits the i- preceding three phases, and the foreground target will be retrieved next.
th component of the Gaussian models. The current image should be compared to the model in the background,
⃒ ⃒ if it matches, the point is the background area. Otherwise, it belongs to
⃒Xt − Ui,t− 1 ⃒⩽Dσ 2
i,t− 1 (1) the foreground target.
whereUi,t− 1 , σ i,t− 1 represent the mean, variance of the i-th Gaussian 5 Morphological noise removal
distribution at time t-1, respectively. D is a constant, generally set to 2.5,
and K is usually set to 3–5. The foreground target retrieved by the GMM method maintains the
shape and edge information of the debris. Still, there are mistaken holes
2 Parameter Update and small debris within, and the constant opening and closing proced
ures greatly improve the situation.
The pixel does not belong to the background model if it does not
satisfy the Equation (1), there are two possibilities for parameter 2.2.2. Wear debris tracking algorithm based on blob
updates: The tracking method based on Blob is employed to monitor and
quantify the speed of numerous moving wear debris. It includes four
(a) Pixels match the GMM steps: Blob extraction, Kalman tracking, data association, and Kalman
correction.
3
Z. Liu et al. Measurement 195 (2022) 111152
3. Methodology
Fig. 2. Tracking effect of multi-target tracking algorithm: (a) nylon debris, (b)
iron debris. In this section, a wear debris velocity model is established through
conducting a set of tests, which can measure the properties of wear
(1) The extracted linked domain is a Blob, and pixels fulfilling eight debris and oil viscosity. DoE method is a set of efficient mathematical
neighborhoods have the same value. The linked domain area tools that can be applied to statistical modeling and systematic problem
threshold in the method is set to 10 μm, and wear debris less than analysis. Among which, RSM is an excellent method for building an
10 μm are discarded as noise [1]. Through the extraction of the empirical model, its primary benefit over other methods is the decreased
linked domain, the number and area characteristics of wear number of experimental runs required and a more systematic exami
debris can be retrieved. Then the size of debris d can be obtained nation of the input variables, including their simultaneous interaction
through: [32].
√̅̅̅̅̅̅̅̅̅̅
d = 4S/π (7)
3.1. RSM methodology modeling
where S is the debris area.
Based on mathematical models (linear, quadratic polynomial, etc.),
(2) Kalman filter is employed to predict the movement of the debris RSM is a collection of techniques that define the connection between
since the target is moving at a constant small pace in the video numerous independent variables and one or more response variables
sequence. [33], and then the model is verified using mathematical statistics. The
(3) The Munkres algorithm is devoted to establishing an incidence link model may then be utilized for optimization, prediction, or inter
matrix to achieve data matching between several tracking debris pretation. Fig. 3 depicts a full flowchart of the suggested modeling
and observation debris. technique. The modeling technique can be separated into five modules
based on the flowchart: a) define, b) design, c) run experiment, d) fit and
4
Z. Liu et al. Measurement 195 (2022) 111152
Table 2
Kinematic Viscosity (mm2/s) of three types of lubricants at different
temperatures.
Temperature Great wall oil Mobil Jet Oil Aeroshell Turbine Oil
(◦ C) 4010 II w80
− 40 2150 11,000 /
40 13.47 27.6 118
100 3.31 5.1 14.5
failure is easy to occur inside the aeroengine and other rotating ma
chinery, not only due to its complex structure, but also owing to the fact
that their main internal support components, bearings and accessory
Fig. 4. Frame diagram of a conventional gas turbine engine. drive system, are exposed to high speeds and high temperature for long
periods. Moreover, the performance of the main bearing directly affects
the life and reliability of engine. Therefore, the oil serves as the engine’s
Table 1
blood, carrying a wealth of wear data about the bearing. Components of
The commonly used bearing materials.
bearings include: outer ring, inner ring, roller, retainer, sealing ring, etc.
Materials Density (kg/m3) Location Table 1 lists the commonly used materials for each part [34]. Accord
Steel 7800 outer ring, inner ring, roller, bearing retainer ingly, the density of several types debris in the table can be selected as
Aluminium 2700 bearing bridge the input variables A, and in our experiment, the corresponding material
Copper 8900 oil retainer, bearing retainer
particle can be obtained by purchasing from Yingtai Company.
Silver 1050 surface plating layer
Lead 1134 Main bearing, connecting rod bearings The link between debris formation and wear process is depicted in
Silicon 2320 sealing ring Fig. 5[35–37]. Under the early operating circumstances of a new engine,
Zinc 7140 rubber seals we can see that most debris are typically less than 10 μm in size, and the
Titanate 4500 bearing chock debris production rate is approximately consistent even though there are
chromium 7190 sealing ring, bearing retainer
Nickel 8900 roller
also having a few debris larger than 10 μm. When the wear between
Ceramics 2400 roller engine components becomes severe, the rate of formation and concen
Nylon 1200 bearing retainer tration of bigger wear debris between 10 μm and 100 μm steadily in
creases until the engine fails. Then the debris size exceeds 100 μm, the
number of debris produced increases dramatically, especially if the en
gine is experiencing a malfunction or is in critical condition. Hence,
debris larger than 10 μm should be monitored to offer an early warning
of failure for an intelligent engine. So, the range of 10–100 μm can be
utilized to determine the value of the input variable B, and it can be
measured directly by the camera in the optical oil debris sensor.
5
Z. Liu et al. Measurement 195 (2022) 111152
Table 3 number of factors as the CCD [40]. Therefore, BBD is used to determine
Independent variables in actual and coded levels. the sample points for the response surface model, and the five variables
Independent variables Symbol level are analyzed at three distinct levels coded as − 1, 0, and +1. Table 3
summarizes the five variables’ actual and coded values.
− 1 0 1
X1: wear debris density (Kg/m3) A 1200 4500 7800 3.1.3. Experimenting and getting the results
X2: wear debris diameter (μm) B 20 50 80
X3: oil flow rate (mm/s) C 2 7 12
The next crucial step is to arrange a series of experiments in a certain
X4: oil viscosity (40 ◦ C, mm2/s) D 35 70 105 sequence through a computer software. The popular statistical software
X5: temperature (◦ C) E 20 40 60 for modeling are Statistica (Stat Soft), Minitab (Minitab Inc.), MATLAB
Response (Mathworks), and Design Expert (State-Ease Inc.) [41]. Experiments are
R: (mm/s) Y
designed and analyzed using Design-Expert ver. 12 in this paper. Ac
cording to the BBD matrix (Fig. 3), the number of 43 (=2k(k− 1) + C0,
the value should be appropriate for the experimental equipment’s where k is the number of independent factors and C0 is the number of
hardware and software conditions. center points) experiments incorporating 5 factors, and 3 center points
are proposed by the software.
3.1.2. Selection of experimental design strategy In our experiment, a spoonful of particles with specified properties is
Then the adoption of an effective approach to carrying out the tests is put in 10 ml of the corresponding viscosity oil respectively, shaken
necessary after carefully choosing the most appropriate parameters thoroughly to mix them well and placed the oil in an ultrasonic cleaner
affecting the velocity of the wear debris. So, strategies for experimental to remove bubbles. Then the oil is heated to the specified temperature
design should be estimated and determined in module 2. Several design using a heating base, the oil flow rate is controlled by the pump in the
methods can be applied for modeling of wear debris velocity, the optical debris sensor. These 43 trials based on BBD and corresponded
effective, typical, and widely utilized RSM designs are the CCD and the response (wear debris velocity (mm/s)) for each trial are shown in
BBD. The BBD is optimal for two to five variables, and because there is Table 4. Moreover, the run order of experiments is randomized to avoid
no axial point in the BBD, it requires fewer test numbers with the same systematic errors.
Table 4
Experimental design and results of the BBD.
Run order A B C D E Y
wear debris density (Kg/m3) wear debris diameter (μm) oil flow rate (mm/s) oil viscosity (mm2/s) temperature (◦ C) wear debris velocity (mm/s)
1 1200 50 7 35 40 5.92
2 4500 50 2 70 20 2.114
3 4500 80 7 105 40 6.84
4 4500 20 7 105 40 7.26
5 4500 50 7 35 60 5.577
6 4500 50 7 70 40 6.17
7 4500 20 7 35 40 6.73
8 1200 50 7 70 20 9.25
9 4500 50 7 70 40 6.72
10 4500 50 12 35 40 10.57
11 7800 50 7 70 20 7.142
12 1200 50 7 105 40 8.22
13 7800 50 7 35 40 4.1
14 4500 80 7 35 40 5.68
15 4500 80 7 70 20 6.73
16 4500 50 12 105 40 14.18
17 4500 80 12 70 40 11.73
18 4500 80 7 70 60 5.71
19 4500 50 7 105 20 7.4
20 1200 20 7 70 40 7.685
21 4500 50 2 105 40 1.994
22 4500 20 2 70 40 1.54
23 4500 50 2 70 60 1.174
24 7800 50 7 70 60 4.93
25 4500 20 7 70 20 8.22
26 7800 50 12 70 40 10.57
27 4500 50 2 35 40 0.96
28 1200 50 2 70 40 2.167
29 1200 50 12 70 40 14.8
30 7800 50 7 105 40 7.592
31 4500 20 12 70 40 12.57
32 1200 50 7 70 60 6.67
33 7800 20 7 70 40 6.777
34 4500 50 7 105 60 6.6125
35 4500 80 2 70 40 1.22
36 4500 50 7 70 40 6.73
37 4500 20 7 70 60 6.277
38 7800 50 2 70 40 0.658
39 1200 80 7 70 40 7.636
40 7800 80 7 70 40 5.649
41 4500 50 12 70 20 12.33
42 4500 50 7 35 20 5.29
43 4500 50 12 70 60 11.504
6
Z. Liu et al. Measurement 195 (2022) 111152
Table 5 the statistical significance of the above regression model. We can see
Expressions used for model verification. that the F-test result (F = 68.24) with a low probability value (p-value)
Terms Expressions Remarks (<0.0001) and the values of the R2, Adj. R2, Pred. R2 are 0.9841, 0.9697,
0.9372, respectively, indicating that the regression model has a high
R 2 SSRes Close to 1.0 is ideal
R2 =
SST statistical significance and that there is a strong relationship between the
R2adjusted R2adjusted= Close to 1.0 is ideal. experimental data of the wear debris velocity and the five input vari
SSRes /(n − p) ables. In addition, the value of PRESS (32.54) and LOF (0.2234 > 0.05)
1−
SST /(n − 1) also reflects the excellent predictive ability of the model.
PRESS PRESS = The value should be small.
∑n In the factorial procedure, it is typical to run significance tests (t-
̂i )2
i=1 (yi − y
tests) on each model coefficient and remove insignificant variables [32].
R2predicted PRESS No more than 0.2 discrepancy between R2adjusted
R2predicted = 1 −
SST The data listed in Table 6 show that the coefficients of five independent
and R2adjusted should be expected.
variables, two interactive cross-product coefficients of AC and CD with
Lack of SSLOF /(f − p) FLOF should be larger than 0.05
Fit
FLOF =
SSPE /(n − p)
p-value <0.05, significantly affect the wear debris velocity, whereas the
other coefficients have negligible effect and, thus, are eliminated from
SSRes = square sum of error; SST = total sum of squares; f = number of distinctly the description model. As a consequence of the analysis, the following
different factor combinations; p = number of parameters; n = experimental
regression equation can represent the base coded and real relationship
number; yi = observed value; ŷi = predicted value.
between the response Y and the five independent variables:
Table 6
ANOVA of the regression model for wear debris velocity.
Source Sum of Squares Degree of Freedom (DF) Mean Square F value p-value
7
Z. Liu et al. Measurement 195 (2022) 111152
Fig. 6. Diagnostic plots for wear debris velocity (mm/s): (a) normal plot of residuals, (b) residuals versus predicted values plot, (c) predicted versus actual values
plot, (d) residuals versus run number plot.
quadratic model is appropriate and acceptable for the experimental whereas oil flow rate and oil viscosity have a positive coefficient. It is
data. Fig. 6(d) depicts the residual versus order of experimental run for observed that the influence of the five parameters on the wear debris
debris velocity, indicating a random uniform distribution of points. The velocity follows the order: oil flow rate > oil viscosity > wear debris
model is consequently adequate and there is no reason to suspect that density > temperature > wear debris diameter.
the assumption of independent or constant variance is violated in all
runs. 3.2.2. Binary effects of the variables on wear debris velocity
As shown in Fig. 8, various response surface plots are produced to
show the mutual interaction between two targeted factors on wear
3.2. Effects of the variables on wear debris velocity
debris velocity by converting them into a three-dimensional (3D) figure,
and the shape of the contours reflects the magnitude of the interaction
3.2.1. Main effects of the variables on wear debris velocity
effect between the two factors, with an ellipse indicating a stronger
Fig. 7 summarizes the effects of all independent variables on wear
interaction and the opposite for a circle. Hence, the values of fixed
debris velocity at the midpoint of coded levels. Sharp curvature curves in
variables in all contour and response surface plots are set to wear debris
the perturbation plot indicate a high sensitivity of response to the
density = 4500 kg/m3, wear debris diameter = 50 μm, oil flow rate = 7
related variables, while semi-flat curves reveal a lower sensitivity of
mm/s. oil viscosity = 70 mm2/s, and temperature = 40 ◦ C.
debris velocity to relevant factors. With increasing parameter C, namely
Fig. 8(a) depicts the interactive effects of wear debris diameter and
oil flow rate, sharply accelerate of wear debris velocity occurs, which
wear debris density on debris velocity, denoting that there is almost no
indicates the variable C has the strongest effect on the wear debris ve
relation between the density and diameter of debris. The interaction
locity. On the contrary, Fig. 7 also indicates that increasing the
effect of oil flow rate and wear debris density on the response, as
parameter A, E, and B negatively affect the wear debris velocity. Such
schematically shown in Fig. 8(b), demonstrates the vital role of oil flow
findings are consistent with Equation (9) where wear debris density,
rate in contrast to debris density. The pairwise interaction of oil viscosity
wear debris diameter and temperature have a negative coefficient
8
Z. Liu et al. Measurement 195 (2022) 111152
9
Z. Liu et al. Measurement 195 (2022) 111152
Fig. 8. Response surface plots for the mutual effects of: (a) wear debris density and diameter; (b) wear debris and oil flow rate; (c) oil flow rate and oil viscosity; (d)
oil viscosity and temperature.
Fig. 9. On-line optical debris sensor installed on a pin disc test rig.
10
Z. Liu et al. Measurement 195 (2022) 111152
Table 7
Validation of the developed model for monitoring of wear debris density and oil viscosity.
Run Experimental factor Model RSM estimated value Experimental value Relative errorσ
order (%)*
B C E Y A D A D A D
wear debris oil flow Temperature wear debris wear debris oil viscosity wear debris oil viscosity wear oil
diameter rate (◦ C) velocity density (kg/ (mm2/s) density (kg/ (mm2/s) debris viscosity
(μm) (mm/s) (mm/s) m3) m3) density
Fig. 10. Surface images of specimens of the first validation experiment (disc is Brass, and pin is GCr15 Steel). (a) Surface image of the disc specimen; (b) Microscopic
surface image of the disc specimen with wear. (c) Surface image of the pin specimen.
[25,27], the improvements of current work can be summarized as: a) A reliable and comprehensive health information for oil condition moni
coaxial reflection light source is employed in the image acquisition toring, which brings much convenience to the work of rotating equip
system to strengthen the details of the debris surface; b) The proposed ment condition maintenance engineers.
multi-objective tracking algorithm can precisely locate the coordinates Obviously, the measurement of oil condition based on wear debris
of each wear debris separately and quickly calculate the movement velocities using optical sensors can increase the dimensionality of
speed of each debris; c) The most important innovation is the con monitoring indicators. However, there are still some problems to be
struction of measurement models. Previous works for the measurement solved in the monitoring process. Due to the limitation of the fixed
of WDA and oil viscosity have focused on creative research on various illumination angle of a single light source, the magnification of the lens
online, inline, and offline monitoring sensor principles, such as elec and the depth of field, the texture of the debris surface cannot be
trostatic, inductive [13], capacitive, ultrasonic [11], optical [42], fer observed clearly, which leads to need further research on the wear
rography [43], etc. as well as on the improvement of monitoring mechanism of mechanical components and the failure mechanism of the
accuracy in terms of wear debris classification [44], quantity and surface components cannot be located precisely in the dynamic oil. But the
texture [29], without considering the research of uniform measurement surface morphology of the wear debris can be seen by replacing the lens
of multiple indicators in oil. The current work is to establish a multidi with a high magnification, when the oil carrying the debris in the
mensional parameter monitoring model of oil based on the velocities of microchannel is in a static state. Although the proposed method is able
wear debris movement in optical sensors by using response surface to calculate the composition of debris and the oil viscosity with satis
method, which allows real-time monitoring of the debris size in oil, the factory accuracy, the debris concentration cannot be too high, because
debris compositions, the debris basic morphology, and the viscosity of excessive number of debris will affect the calculation speed of the
oil. Compared with traditional measurement methods such as ferrog tracking algorithm, causing time delay and reducing the accuracy of the
raphy, viscometer and mechanical assembly of multiple sensors, the calculation results. This problem can be improved by improving the
advantage of the proposed measurement method is that it can use a robustness of the algorithm.
single sensor based on the optical principle to visually monitor the Real-time health condition monitoring of oil is an important method
health status of the oil in real time with multiple indicators, lower time to reflect the wear status of mechanical rotating parts during operation.
and lower financial costs. For example, the degree of oil contamination In this paper, a mathematical model of condition monitoring is estab
can be initially determined based on the imaging effect of the oil in the lished from the perspective of liquid-solid two-phase flow. Future work
optical sensor, and the metal and non-metal can be preliminarily sepa is prepared to optimize the model, improve the accuracy of monitoring
rated based on the imaging result of the debris color as Fig. 2, then parameters, and explore the correlation between other oil monitoring
combine with the established model to quickly calculate the oil viscosity parameters. In addition, the structure of the flow channel of wear debris
and accurately distinguish the debris composition based on the debris will be optimized to identify the type of wear debris from the velocity
density results. Therefore, the newly proposed method can provide perspective. For example, various flow channel and debris are modeled
11
Z. Liu et al. Measurement 195 (2022) 111152
using the 3D software, and different moving behavior of debris are ob [8] S.B. Glavatskih, A method of temperature monitoring in fluid film bearings, Tribol.
Int. 37 (2) (2004) 143–148, https://doi.org/10.1016/S0301-679X(03)00050-1.
tained in the fluid simulation software, then the type of debris can be
[9] H. Shinde, A. Bewoor, Analyzing the relationship between the deterioration of
identified. engine oil in terms of change in viscosity, conductivity and transmittance, pp.
36–41.
[10] R.A. Potyrailo, I. Tokarev, S. Go, P. Ottikkutti, N. Kuzhiyil, J. Mihok, C. Anzini,
6. Conclusions
S. Shartzer, Multivariable Electrical Resonant Sensors for Independent
Quantitation of Aging and External Contaminants in Lubricating Oils, IEEE Sens. J.
The characteristics of wear debris and oil can be used for under 19 (4) (2019) 1542–1553.
standing the wear mechanisms of mechanical equipment. A new optical [11] C. Xu, P. Zhang, H. Wang, Y. Li, C. Lv, Ultrasonic echo waveshape features
extraction based on QPSO-matching pursuit for online wear debris discrimination,
measurement method for monitoring wear debris and oil viscosity Mech. Syst. Sig. Process. 60-61 (2015) 301–315.
simultaneously is proposed. The main features are that: 1) the designed [12] B.J. Roylance, J.A. Williams, R. Dwyer-Joyce, Wear debris and associated wear
online monitoring optical sensor can collect images of wear debris in phenomena—fundamental research and practice, Proc. Inst. Mech. Eng. Part J: J.
Eng. Tribol. 214 (1) (2005) 79–105, https://doi.org/10.1243/
dynamic oil in real time and accurately, whose equivalent diameter is 1350650001543025.
greater than 10 μm and the shape can be random; 2) the developed [13] H. Xiao, X. Wang, H. Li, J. Luo, S. Feng, An Inductive Debris Sensor for a Large-
multi-objective tracking algorithm can accurately capture the coordi Diameter Lubricating Oil Circuit Based on a High-Gradient Magnetic Field, Appl.
Sci. 9 (8) (2019) 1546.
nate position of each debris and perform real-time tracking updates to [14] M. Appleby, F.K. Choy, L.i. Du, J. Zhe, Oil debris and viscosity monitoring using
calculate its motion speed; 3) 43 sets of tests are designed and completed ultrasonic and capacitance/inductance measurements, Lubr. Sci. 25 (8) (2013)
by the RSM, and the final mathematical model is established to monitor 507–524.
[15] Y. Liu, Z. Liu, Y. Xie, Z. Yao, Research on an on-line wear condition monitoring
several parameters of debris density, debris diameter, and oil viscosity. system for marine diesel engine, Tribol. Int. 33 (12) (2000) 829–835.
Then the significance of the model is examined by ANOVA and the [16] Y. Iwai, T. Honda, T. Miyajima, S. Yoshinaga, M. Higashi, Y. Fuwa, Quantitative
performance is verified from a specific experimental point of view using estimation of wear amounts by real time measurement of wear debris in lubricating
oil, Tribol. Int. 43 (1-2) (2010) 388–394.
a pin-disk tester, a solids densitometer and a viscometer. The results
[17] T. Wu, H. Wu, Y. Du, N. Kwok, Z. Peng, Imaged wear debris separation for on-line
indicate that the proposed method obtains a high agreement between monitoring using gray level and integrated morphological features, Wear 316 (1-2)
the real values of debris density and oil viscosity. Compared with (2014) 19–29.
existing oil condition monitoring methods, this proposed approach of [18] W. Cao, G. Dong, Y.-B. Xie, Z. Peng, Prediction of wear trend of engines via on-line
wear debris monitoring, Tribol. Int. 120 (2018) 510–519.
fers multiple monitoring parameters of the oil simultaneously by using a [19] Y. Peng, Z. Wu, G. Cao, S. Wang, H. Wu, C. Liu, Z. Peng, Three-dimensional
single sensor, thus achieving reliable, efficient and more comprehensive reconstruction of wear particles by multi-view contour fitting and dense point-
oil condition monitoring. cloud interpolation, Measurement 181 (2021) 109638.
[20] H. Shi, H. Zhang, L. Ma, L. Zeng, A multi-function sensor for online detection of
contaminants in hydraulic oil, Tribol. Int. 138 (2019) 196–203.
CRediT authorship contribution statement [21] A. Becker, Health indicator metrics applicable to inductive wear debris sensors,
Proc. Inst. Mech. Eng. Part J: J. Eng. Tribol. 231 (5) (2016) 583–593, https://doi.
org/10.1177/1350650116665047.
Zhenzhen Liu: Methodology, Software, Data curation, Writing – [22] H. Shi, H. Zhang, C. Gu, L. Zeng, A multi-parameter on-chip impedance sensor for
original draft, Writing – review & editing, Funding acquisition. Yan Liu: the detection of particle contamination in hydraulic oil, Sens. Actuators, A 293
(2019) 150–159.
Writing – review & editing, Supervision. Hongfu Zuo: Conceptualiza
[23] H. Shi, H. Zhang, W. Wang, L. Zeng, G. Sun, H. Chen, An Integrated Inductive-
tion, Funding acquisition. Han Wang: Methodology. Cheng Wang: Capacitive Microfluidic Sensor for Detection of Wear Debris in Hydraulic Oil, IEEE
Resources. Sens. J. 19 (23) (2019) 11583–11590.
[24] L.i. Du, J. Zhe, An integrated ultrasonic–inductive pulse sensor for wear debris
detection, Smart Mater. Struct. 22 (2) (2013) 025003.
Declaration of Competing Interest [25] B.o. Li, Y. Xi, S. Feng, J. Mao, Y.-B. Xie, A direct reflection OLVF debris detector
based on dark-field imaging, Meas. Sci. Technol. 29 (6) (2018) 065104.
[26] Z. Liu, Y. Liu, H. Zuo, H. Wang, H. Fei, A Lubricating Oil Condition Monitoring
The authors declare that they have no known competing financial System Based on Wear Particle Kinematic Analysis in Microfluid for Intelligent
interests or personal relationships that could have appeared to influence Aeroengine, Micromachines 12 (7) (2021) 748.
the work reported in this paper. [27] Y. Peng, T. Wu, S. Wang, Y. Du, N. Kwok, Z. Peng, A microfluidic device for three-
dimensional wear debris imaging in online condition monitoring, Proc. Inst. Mech.
Eng. Part J: J. Eng. Tribol. 231 (8) (2017) 965–974.
Acknowledgements [28] H. Wu, R. Li, N.M. Kwok, Y. Peng, T. Wu, Z. Peng, Restoration of low-informative
image for robust debris shape measurement in on-line wear debris monitoring,
Mech. Syst. Sig. Process. 114 (2019) 539–555.
This work is supported by the National Natural Science Foundation [29] H. Wu, N.M. Kwok, S. Liu, R. Li, T. Wu, Z. Peng, Restoration of defocused
of China (U1933202), the Postgraduate Research and Practice Innova ferrograph images using a large kernel convolutional neural network, Wear 426-
tion Program of Jiangsu Province (KYCX21_0235) and the Postgraduate 427 (2019) 1740–1747.
[30] P. Peng, J. Wang, Wear particle classification considering particle overlapping,
Innovation Experiment Competition Program of NUAA (CXJS21_068). Wear 422–423 (2019) 119–127, https://doi.org/10.1016/j.wear.2019.01.060.
[31] J. Wu, J. Mao, W. Cao, Y.-B. Xie, Characterization of wear-debris group in on-line
References visual ferrographic images, Proc. Inst. Mech. Eng. Part J: J. Eng. Tribol. 228 (11)
(2014) 1298–1307.
[32] R. Ghelich, M.R. Jahannama, H. Abdizadeh, F.S. Torknik, M.R. Vaezi, Central
[1] P. Lu, H.E. Powrie, R.J.K. Wood, T.J. Harvey, N.R. Harris, Early wear detection and
composite design (CCD)-Response surface methodology (RSM) of effective
its significance for condition monitoring, Tribol. Int. 159 (2021) 106946.
electrospinning parameters on PVP-B-Hf hybrid nanofibrous composites for
[2] J.M. Wakiru, L. Pintelon, P.N. Muchiri, et al., A review on lubricant condition
synthesis of HfB2-based composite nanofibers, Compos. B Eng. 166 (2019)
monitoring information analysis for maintenance decision support, Mech. Syst. Sig.
527–541.
Process. 118 (2019) 108–132, https://doi.org/10.1016/j.ymssp.2018.08. 039.
[33] M. Mia, Mathematical modeling and optimization of MQL assisted end milling
[3] J. Sun, L. Wang, J. Li, F. Li, J. Li, H. Lu, Online oil debris monitoring of rotating
characteristics based on RSM and Taguchi method, Measurement 121 (2018)
machinery: A detailed review of more than three decades, Mech. Syst. Sig. Process.
249–260, https://doi.org/10.1016/j.measurement.2018.02.017.
149 (2021) 107341.
[34] O. Altıntaş, M. Aksoy, E. Ünal, O. Akgöl, M. Karaaslan, Artificial neural network
[4] X. Zhu, C. Zhong, J. Zhe, Lubricating oil conditioning sensors for online machine
approach for locomotive maintenance by monitoring dielectric properties of engine
health monitoring – A review, Tribol. Int. 109 (2017) 473–484, https://doi.org/
lubricant, Measurement 145 (2019) 678–686.
10.1016/j.triboint.2017.01.015.
[35] W. Hong, W. Cai, S. Wang, M.M. Tomovic, Mechanical wear debris feature,
[5] B. Yu, N. Cao, T. Zhang, A novel signature extracting approach for inductive oil
detection, and diagnosis: A review, Chin. J. Aeronaut. 31 (5) (2018) 867–882.
debris sensors based on symplectic geometry mode decomposition, Measurement
[36] A. Kumar, S.K. Ghosh, Size distribution analysis of wear particles in the
185 (2021) 110056, https://doi.org/10.1016/j.measurement.2021.110056.
transmission system of mining equipment, Proc. Inst. Mech. Eng. Part J: J. Eng.
[6] H. Liu, T. Li, W. Hong, S. Wang, Using multi-window correlation to improve
Tribol. 232 (8) (2017) 921–926, https://doi.org/10.1177/1350650117746804.
sensitivity and adaptability for oil debris detections, Measurement 176 (2021)
[37] A. Kumar, S.K. Ghosh, Size distribution analysis of wear debris generated in HEMM
109236, https://doi.org/10.1016/j.measurement.2021.109236.
engine oil for reliability assessment: A statistical approach, Measurement 131
[7] M. Soleimani, Thick film sensors for engine oil acidity detection, University of
(2019) 412–418, https://doi.org/10.1016/j.measurement.2018.09.012.
Southampton, 2014.
12
Z. Liu et al. Measurement 195 (2022) 111152
[38] I. Hutchings, P. Shipway, Lubricants and lubrication, Tribology (2017) 79-105. [42] M.A. Khan, A.F. Cheema, S.Z. Khan, et al., Image based portable wear debris
[39] F. Wardle, in: Ultra-Precision Bearings, Elsevier, 2015, pp. 37–146. analysis tool, Ind. Lubric. Tribol. 67 (4) (2015) 389–398, https://doi.org/10.1108/
[40] W. Zhao, A. Ma, J. Ji, X.u. Chen, T. Yao, Multiobjective Optimization of a Double- ILT-11-2014-0127.
Side Linear Vernier PM Motor Using Response Surface Method and Differential [43] T.H. Wu, J.H. Mao, J.T. Wang, J.Y. Wu, Y.B. Xie, A New On-Line Visual Ferrograph,
Evolution, IEEE Trans. Ind. Electron. 67 (1) (2020) 80–90. Tribol. Trans. 52 (5) (2009) 623–631.
[41] Y. Peng, U. Khaled, A.A.A.A. Al-Rashed, R. Meer, M. Goodarzi, M.M. Sarafraz, [44] H. Fan, S. Gao, X. Zhang, X. Cao, H. Ma, Q.i. Liu, Intelligent Recognition of
Potential application of Response Surface Methodology (RSM) for the prediction Ferrographic Images Combining Optimal CNN With Transfer Learning Introducing
and optimization of thermal conductivity of aqueous CuO (II) nanofluid: A Virtual Images, IEEE Access 8 (2020) 137074–137093.
statistical approach and experimental validation, Physica A 554 (2020) 124353.
13