Accepted Manuscript

The determination of combustion engine condition and reliability using oil analysis by
MLP and RBF neural networks

Jakub Gajewski, David Vališ

PII: S0301-679X(17)30320-1
DOI: 10.1016/j.triboint.2017.06.032
Reference: JTRI 4792

To appear in: Tribology International

Received Date: 11 April 2017

Revised Date: 20 June 2017
Accepted Date: 22 June 2017

Please cite this article as: Gajewski J, Vališ D, The determination of combustion engine condition
and reliability using oil analysis by MLP and RBF neural networks, Tribology International (2017), doi:
10.1016/j.triboint.2017.06.032.

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
M
D
TE
EP
C
AC
 The determination of combustion engine condition and reliability using oil
ACCEPTED MANUSCRIPT
analysis by MLP and RBF neural networks
Jakub Gajewskia*, David Vališb

Notation and units

Nomenclature:

AES – atomic emission spectrometry

PT
ANN – artificial neural networks
APP [%] – anti-wear particles
AxP [%] – anti-oxidation particles
BFGS – Broyden–Fletcher–Goldfarb–Shanno network learning method

RI
CBM – condition-based maintenance
ci – set of determinable centres

SC
Cr [ppm] – chromium
Cu [ppm] – copper
Fe [ppm] – iron
Flash point [°C]

U
FTIR – Fourier transformation infrared
f(u) – activation function
AN
Ignition point [°C]
km – kilometres
LNF – laser net finder
Mh – motor-hours
M

MLP – multilayer perceptron neural network

Mo [ppm] – molybdenum
Na [ppm] – sodium
D

NP [ppm] – nitration products

NN – neural network
TE

OLVF – on-line visual ferrograph

Pb [ppm] – lead
PM – preventive maintenance
EP

RBF – radial basis function network

RUL – remaining useful life
Soot [%]
Sp [%] – sulphating products
C

TBN – total base number

ui – output signal
AC

vi – output signal of neuron

xj – input signal
yk – layer output signal
wab – values of weights
wij – suitable weights of synaptic connections
w0 – threshold value, bias

 – parameter determining a function width

a
Lublin University of Technology, Faculty of Mechanical Engineering, Department of Machine Design and
Mechatronics, Nadbystrzycka 36 Str., 20-618 Lublin, Poland *Corresponding author
b
University of Defence, Kounicova 65, 662 10 Brno, Czech Republic
 1. Introduction
ACCEPTED MANUSCRIPT
Given more and more demanding requirements of the quality and dependability of technical systems,
the expectations of system condition information obtained during operation are heightened (see e.g. [1]).
Non-destructive diagnostic signals of various forms, such as vibration, heat transfer and sound, are one of
the potential sources of information about system condition.
Condition-based maintenance (CBM) has become an attractive research area over in recent decades,
as a result of practical advances in this area (see e.g. [2–6]).
The literature on the subject offers a few works on the methods for obtaining more accurate
information about oil condition or system condition based on the information obtained from the oil itself.
There are several works on the ways of obtaining accurate information about oil or technical system

PT
condition involving the use of information from oil. However, the literature lacks studies on the prediction
of the condition of both oil and the technical system in which this oil is being used.
This study focuses on heavy transport systems, the operating environment of which is very difficult
and demanding. The specimens of oil were obtained from several dozen engines of heavy crawlers; the

RI
dataset contained over 350 records collected over a period of 10 years.
The data were acquired thank to systematic recording (see [7] for details) and properly organised
laboratory processing. The systematic analysis of specimens was performed by atomic emission

SC
spectrometry (AES) using a SPECTROIL M Q100 from SPECTRO Inc., Littleton, Massachusetts, USA,
while morphology observations were made by Fourier transformation infrared (FTIR) and laser net finder
(LNF), using a NICOLET AVATAR 330 manufactured by Thenno Electron Scientific Instruments

U
Corporation from Madison, Wisconsin, USA.
A certain level of inaccuracy and uncertainty is inevitable, due to the collection and laboratory
AN
instrumentation of oil specimens taken from the system during operation. For this reason, some traditional
methods of data analysis were not useful or applicable here.
Using these data and neural network (NN), we discuss the patterns which can point to material and
system deterioration. There have been no previous studies on operating objects of this size and set volume.
M

A few previous works – see e.g. [8] – provide limited perspective on observation of tribological systems. In
[8], only large mobile technical systems, such as diesel locomotives, mine lorries and warships, were
assessed, but using perfunctory approaches. There are no previous studies which have focused to such an
D

extent on special vehicles such as tracked, medium-sized lorries and common road vehicles. A few papers
(e.g. [9, 10]) describe uses of neural networks or other numerical methods for oil systems, but in other
TE

aspects of application.
Given the great potential of artificial neural networks (ANN) and a large statistical database of oil
samples as well as the limitations mentioned above, it seems justified to use a selection of neural networks
EP

suitable for this study. The aim of this study is to describe the degradation of combustion engines in the
above types of vehicles using neural networks.
We would like to prove whether there is a potential for change in time periods for hard time
preventive maintenance (PM), as we think that some PM time periods – especially for oil change – are rather
C

short. This is the main motivation behind this study, as there are a vast amount of operational data which
provide significant information about system condition. Therefore, the results of this study can be used for
AC

the establishment of procedures for operation optimisation and maintenance rationalisation.

 2. Present situation in neural networks, condition-based maintenance and
ACCEPTED MANUSCRIPT
tribodiagnostics

We are aware that there has already been much discussion in the literature on CBM, degradation, soft failure
occurrence and remaining useful life (RUL) prediction; CBM and degradation modelling and soft failure
occurrence play vital role in present reliability approaches. Many publications can be selected, but not all
suit our idea of raw oil field data elaboration. Our perspective is focused on the most applicable sources,
which we have classified. Some of the sources overlap, however we have organised their overview into the
following groups: (1) field and real-time data plus their elaboration and application for CBM purposes; (2)
elaboration of oil field data; (3) typical indicators of degradation plus deterioration whose data processing

PT
leads to RUL estimation; and (4) application of NN for data elaboration and creation of relevant reliability
models.
As mentioned in [11], assessment of reliability plus availability made in real-time is a key factor in

RI
CBM. Outcomes can modify decisions related to optimisation and the next maintenance steps. Park [2]
described some optimal prognostic criteria based on cost for CBM when he studied risk variables based on a
multiple logistic function. Unfortunately and typically, the function applied is only exponential. The system,

SC
which works in two-phase rhythm, is described by Maillart and Pollock [3]. They proposed a predictive-type
maintenance policy for it while applying a cost-minimising and satisfying approach. This was done for exact
determination of when system inspection should be conducted and the way in which financial resources for
multiple systems were allocated.

U
The approach by Yang [4] is based on the application of a state-space model of a Kalman filter to
predict failure for PM. Two degradation processes along with random shocks are the basis for multiple
AN
competing failure processes presented by Li and Pham [5] to create a generalised CBM model. Maintenance
actions of an imperfect kind for a degradation system were dealt by Phuc et al. [12] in order to propose a
CBM procedure. Computer aided maintenance optimisation based on real data is presented by Mazurkiewicz
M

[6].
A relatively large number of articles and papers from the tribology area are focused on the chemical
structure of oil and its chemical and physical properties. It is noted in the literature, as we performed the
D

survey on ‘on-line data acquisition’, that oil tribo data are mentioned as one possible indicator applicable in
diagnostics assessments – in addition to vibrations, thermal emission, acoustic signals, gas emissions, etc.
TE

The overview of the literature indicates the need for research on the use of oil in assessing the
technical condition of a machine.

Some oil data assessments as diagnostic data are elaborated and results presented by Wang [8] and
EP

Wang and Zhang [13], and in the publications in the reference lists of these articles. Hidden Markov theory
and stochastic filtering is a basis for wear prediction model which was developed by Wang [14]. In this
publication, Wang assumes that the real state of the system (which cannot be directly observed) is related to
C

metal contaminant concentrations from oil monitoring. He also works with the initial condition that discrete
time point observations are available. A detailed model of soot formation was developed by Zhao et al [15].
AC

This model is believed to reflect real soot formation processes and the dynamic kinetics of oxidation in large
scale diesel engine operations. The results are presented in the form of crucial factors for soot formation.
Growing interest in acquiring the size distributions in addition to the number density and mass fractions of
diesel engine soot particles was emphasised by Lee et al [16]. In this work, the operating conditions, typical
for a diesel engine, address the problem of particle size distribution, which was computed and validated. A
full-scale wind turbine gearbox test, as an inspirational example, is presented by Sheng [17], who elaborated
first-hand oil and wear debris. Stepped loading and stepped velocity approaches to simulating the true
operational conditions of petrol engines are presented by Cao et al [18]. Accelerated wear tests for piston
rings and cylinder liners were conducted via a self-made pin-on-disc wear machine equipped with an on-line
visual ferrograph (OLVF) for wear monitoring, used in varying lubricating conditions. The differences
between constant conditions and stepped changing conditions were used to extract wear coefficients for oil
monitoring. A similarity model for oil monitoring was created through grouping data sets for similar
working conditions. As a result of the outcomes, it was reported that real-time variations in pollutant
concentrations can be acquired by an on-line visual ferrograph monitoring system. Tests performed on a
sensor which measures the lubricant density dynamic viscosity, dielectric constant and temperature were
reported by Clark and Fajardo [19]. A specific temperature range was the basis for bench-top experiments
which were performed in order to ACCEPTED MANUSCRIPT
validate the accuracy of sensor outputs against the outcomes from
automated simulation test methods. Bench-top experiments were deliberately used in order to establish
correlations between fuel contamination levels and changes in lubricant properties. A diesel engine with on-
board sensors was used as an experimental and measurement unit in order to quantify changes in the
lubricant physical properties with respect to engine working period. An interesting principle for the
continuous measurement approach is presented by Durdag and Andle [20]. This approach was selected as
the best one respecting the immediate results records and feedback from quality control.

Although there are publications devoted to the problem of internal combustion engine oil, there are
few analyses related to the use of oil in vehicle engines and, especially, in heavy crawlers. Moreover, the
literature does not provide sufficient analyses regarding reliability, prediction of technical condition, failure

PT
occurrence, and modification of maintenance system parameters. Some potential concepts are presented in
the [21, 22].

RI
Our interest lies in the utilisation of diagnostics signals – oil data – for RUL estimation. We present
here only the most inspirational works in this respect. When the equipment condition is not accurately
known, a suitable approach may be one based on indirect condition monitoring. This is typically the case

SC
when proportional hazard model is applied for modelling of system residual life, which was presented by
Ghasemi et al. [23]. Rolling bearings were the tested elements. We can also find inspirational approaches on
non-direct diagnostics of electric items and methods for RUL estimation using various characteristics of

U
power in [24–26]. Mathematical degradation models based on deteriorated signals always consider a
population of components. Three main sources of information are typically used as combined inputs to
AN
stochastic models: (1) degradation characteristics of a component based on a deteriorated signal which was
acquired in real-time in situ; (2) the component´s population deterioration characteristic; and (3) condition
and status of the environment under which the component operated, also acquired in real-time. The
computed distribution of the time needed for a degradation signal peak to reach a predetermined threshold
M

when the distribution function is defined is in this work believed to be equivalent to estimation of residual
life distribution. Although several RUL and soft failure occurrence estimation approaches can be found in
the literature of the subject, there are no data concerning the respective level of fuzziness and uncertainty.
D

Oil data are always burdened with fuzziness and uncertainty due to their origin. Therefore, we try to
contribute with this study to the formulation of rules describing how RUL can be estimated when
TE

elaborating fuzzy field data.

Neural network structures are inspired by biology: the structures of neurons, synapses and natural
neural systems, especially those in the brain. Neural networks are mathematical models which perform
operations involving the processing of signals by different parts of the structure. Neural systems are
EP

sometimes regraded as ‘black boxes’ because they only produce neural network output information, on the
basis of certain system input signals. In the course of calculations performed by the network, it is possible to
determine an error function or verify, for example, values of neuron weights; however, from a global point
C

of view, we can access these results without the use of a mathematical apparatus.
There are many types of artificial neural networks which differ in terms of their operating
AC

principles and structure. One of the most popular network models is a sigmoid network that has the form of
an multilayer perceptron (MLP). This network offers a wide number of possibilities, for instance with
respect to classification and regression problems, which seems promising in the context of considerations of
oil operating time versus oil composition. The concept of this type of network is based on the arrangement
of neurons into layers. Neurons calculate the weighted sum of their inputs, and thereby determined values
are the arguments of an activation function calculating the values of neuron outputs. The structures of
neurons are unidirectional, that is, they do not provide any feedback. This type of network is regarded as an
input-output model, its parameters being weights and threshold values.
The radial basis function network (RBF), the other type of neural network used in the study,
exhibits a different operating characteristic. The RBF network structure is similar to sigmoidal functions
with one layer. Compared to the MLP, the RBF has a radial basis instead of hidden neurons, the shape of
this radial basis being different from that of sigmoid functions. Although a sigmoidal network may have
nonlinear neurons, radial networks are more varied in terms of their shape [27]. The differences in the
operation of these types of artificial neural networks are illustrated in Figure 1.
 ACCEPTED MANUSCRIPT

PT
Fig. 1. Difference between multilayer perceptron and radial basis function networks

RI
Artificial neural networks are an effective method for the classification and prediction of variables
describing operating parameters of a machine. At present, there are numerous publications available

SC
describing various applications for artificial intelligence methods in the assessment and prediction of time
series or, indirectly, of a system condition (see e.g. [28–32]). In recent years, advanced diagnostic methods
for engines and gearboxes have been developed [33–35]. Artificial neural networks can also be applied to

U
tribological problems. It is worth mentioning here the studies on the use of multilayer perceptrons for
predicting sliding friction [36].
AN
3. Oil data
M

With regard to the application of neural networks in the investigation of system condition using oil samples,
we can draw attention to the studies [37]. Presented here is an original approach to the problem of reliability
D

of object operation wherein oil is a carrier of information about a potential failure of the technical object. A
similar aspect of the problem was also raised in [38], where, based on selected parameters of oil and the
TE

particles contained in oil, an attempt was made to evaluate the possibility of engine diagnostics. The studies
used artificial neural networks and fuzzy systems for this purpose. Given the presented descriptions of the
systems and the records of their performance reported in the publications, it must be concluded that this type
EP

of analysis, although somewhat limited, seems to be promising. In this paper, the authors put forward a
broader hypothesis about the possibility of taking advantage of information contained in oil as a carrier of
symptoms describing the condition of a system such as an engine or gearbox. Previous works were based
only on individual indicators (properties) of engine oil based on a small number of samples obtained from
C

the tested technical systems. The study described in this paper is based on a collection of oil samples
regarded as the carriers of diagnostic information. This approach is important in the context of the number of
AC

preliminary oil characteristics which are considered and used in this analysis. The study was performed
using over 350 oil samples obtained from more than 30 vehicles. Each sample provides 32
characteristics/parameters acquired in laboratory conditions. Out of 32 parameters, we try to identify only
those which are the most significant with respect to the operating time of the tested engine system. A
selection of the parameters – elements (indicators) describing the tested oil – is given in Table 1.
 Table 1. Characteristics of oil and selected oil particles
ACCEPTED MANUSCRIPT
Ignition point AxP APP Fe Cr Pb Cu Na NP Sp Mo Soot
Days Mh Viscosity TBN
[°C] [%] [%] [ppm] [ppm] [ppm] [ppm] [ppm] [ppm] [%] [ppm] [%]
1 609 91 10.458 209 10.814 0.019 72.3 1.62 5.84 312.8 7.37 0.04 0.035 3.05 98.4
2 270 127 8.63 215 10.249 0.024 73.6 21.8 0.9 11 199.3 27.4 0.034 0.015 3.4 83.3
3 553 304 8.9178 209 10.056 0.014 73.6 48.9 2.3 5.4 319.6 43.4 0.053 0.037 2.9 72.5
4 552 288 8.5122 212 10.024 0.009 74.2 44.4 2.3 7 307.5 39.9 0.061 0.03 3.5 72.8
5 964 115 10.755 208 11.347 0.034 74.3 1.14 4.74 322.8 7.66 0.017 0.041 2.01 99.6
6 508 95 7.2156 209 10.153 0.09 74.5 20.6 0.9 4.5 269 12.6 0.022 0.04 2.9 78.9
7 581 324 8.0665 210 10.592 0.035 74.5 56.1 2.3 6.3 333.8 39.9 0.066 0.042 2.9 69.4
8 592 174 7.4114 211 9.958 0.096 74.8 24.2 1.3 11.9 514.4 28.9 0.067 0.039 3.6 81.8

PT
9 580 304 8.3047 218 10.041 0.021 75.3 43.7 2.4 6.7 276.7 37.2 0.062 0.038 3.1 72
10 639 71 10.455 209 10.319 0.019 75.8 9.56 0.97 5.21 259.2 4.09 0.057 0.023 4.07 99.2
11 602 89 10.784 209 10.657 0.048 75.8 0.87 5.92 341.6 5.71 0.018 0.036 3.23 99.3
12 498 163 8.6739 210 10.498 0.027 76.5 25 0.9 262.4 14.7 0.04 0.017 2.6 81.3

RI
13 571 169 7.3552 202 9.919 0.045 76.5 33.9 1.8 10.9 693.1 15.8 0.05 0.022 6.2 76
14 498 136 8.7094 213 10.408 0.021 76.9 17.8 0.8 239.4 16.1 0.038 0.028 2 83.3
15 806 266 9.1121 210 10.644 0.097 77.1 36.6 2.4 7.9 286.1 31.9 0.051 0.042 3 73.3

SC
16 797 359 9.3412 212 9.21 0.089 78.2 46.1 1.9 6.3 657.1 32.1 0.057 0.042 3.5 72
17 807 339 8.0012 208 0.008 78.3 56.4 2.4 6.8 356.7 34.5 0.058 0.03 2.9
18 11 1 9.2614 210 10.0255 0.017 5.54 0.04 3.46 89.39 5.79 0.009 0.03 1.34 98.2
19 539 196 8.4533 210 9.145 79.1 49.3 2.9 11.5 968 17.4 0.043 0.049 9 72

U
20 592 188 8.1546 210 10.157 0.011 79.6 27.4 1.8 9.6 620.4 32.4 0.02 0.012 3.2 78.9
21 580 306 8.122 214 10.111 0.04 80 49.1 2.1 6.3 537.5 0.07 0.049 3.6 72
AN
22 1367 85 8.3297 202 10.148 0.072 80.8 20.55 1.68 9.51 383.5 5.73 0.045 0.045 5.39 100
23 525 48 10.254 208 10.192 0.051 80.9 11.84 1.89 6.83 462.1 3.87 0.024 0.03 4.17 99.3
24 580 256 9.5526 215 10.671 0.101 80.9 33.8 1.9 7.5 302.1 29.3 0.054 0.045 3 80.3
25 580 280 9.2534 213 11.028 0.14 81.5 32 1.8 5.9 518.3 28.3 0.075 2.4 80.3
M

26 269 183 8.7218 217 11.111 0.005 81.6 29.9 1.7 6.1 256.1 43 0.034 0.015 2.9 76
27 160 41 8.7475 219 10.706 0.015 81.8 15.6 0.5 3 294.8 10.7 0.026 0.9 78.9
28 434 236 8.435 213 11.259 0.01 81.8 34.6 1.5 6 260.1 37.2 0.048 0.023 2.5 77.4
D

29 498 252 8.4116 213 10.239 0.018 81.8 39.5 2 5.6 271.5 42.8 0.047 0.032 2.5 70.3
30 359 39 8.2851 215 10.074 0.066 81.9 21.2 0.8 8.7 264.2 19.2 0.023 0.012 1.7 86.4
TE

31 553 321 8.6029 211 9.4399 0.068 81.9 40.9 2.1 7.6 424.4 24.2 0.044 0.043 4.1 75.2
32 660 97 8.4058 10.878 0 82 20.3 0.77 5.06 283.8 18.18 0.04 0.03 2.43 78.9
33 581 331 9.7423 215 9.2661 0.102 82.2 44.9 1.8 7.1 448.1 23.8 0.054 0.045 3.5 76
34 156 36 8.711 210 10.11 0.031 82.5 13.9 0.3 3.7 279.2 8.6 0.015 0.8 78.9
EP

35 477 135 8.0517 211 9.653 0.015 82.8 29.4 1.9 12.1 914.2 39.2 0.034 0.02 5 80.3
36 494 183 8.1078 217 10.426 0.012 82.8 33.5 1.8 11.6 705.9 26.7 0.044 0.023 4.2 76
37 806 312 8.2645 212 9.88 0.047 82.8 48.6 2.1 5.7 556.4 0.059 0.042 2.7
38 114 54 10.249 0.02 82.9 16.3 0.9 8.2 417.9 14 0.027 0 4.3 89.6
C

39 295 65 8.7356 210 10.569 0.028 83 22.4 0.7 4.5 272.2 14.2 0.016 0.013 2.2 81.8
40 552 264 8.2686 211 9.999 0.123 83 32.1 1.9 6.5 495.4 31.1 0.069 2.6 80.6
AC

41 239 42 8.3448 210 10.887 0.031 83.1 15.1 0.7 4.6 252.1 11.7 0.021 0 1.9 77.8
42 1241 130 7.8491 203 9.5543 0.042 83.2 32.66 2.98 14.31 794.4 8.01 0.033 0.042 99.8
43 806 316 8.5742 212 10.92 0.017 83.4 38.9 1.6 12.8 356.4 28.4 0.062 0.045 1.7 74.7
44 470 157 8.7165 214 10.662 0.04 83.6 25.1 1.7 9.3 555.6 32.4 0.034 3.4 84.8
45 484 182 8.673 211 10.719 0.061 83.6 27.8 1.6 10.3 649.3 30.6 0.049 0.019 3.1 76
46 981 130 7.9369 204 10.172 0.038 83.8 20.85 0.89 11.17 429.1 6.26 0.034 0.026 5.5 98.7
47 805 213 9.1044 215 9.692 0.097 83.8 33.1 1.7 10.3 683.5 28.2 0.054 0.049 5.4 76
48 511 185 9.2268 210 11.296 0.069 83.9 26.1 1.6 9.2 527.9 29.7 0.055 0.015 3.3 86.4
49 609 194 8.8706 207 10.799 0.042 84.2 29.3 1.8 10 649.9 30.6 0.051 0.026 3.5 76
50 … … … … … … … … … … … … … … … …
 The relationships between the content of essential – in terms of system condition assessment – oil
ACCEPTED
particles and the actual operating time are shown MANUSCRIPT
in Figures 2–5. Not all oil indicators obtained in the
analysis are relevant to system condition assessment. For this reason, we choose only these elements
(indicators) which are significant with respect to their correlation with operating time – Mh, days,
kilometres [km]. Such indicators also have the greatest potential for providing information about the
system condition and the oil itself. They are particles which, in terms of system condition assessment, are
the products of wear of kinematic joints, such as iron – Fe, lead – Pb, copper – Cu, etc., and which – with
respect to oil condition assessment – are the products or characteristics of its ageing and external pollutants
such as anti-wear particles (APP), anti-oxidation particles (AXP), soot, flash point, etc. The concentration
of these particles increases or decreases depending on the mechanism which describes their origin or
designation.

PT
RI
U SC
AN
M
D

Fig. 2. Na content vs. operating time in Mh Fig. 3. Fe content vs. operating time in Mh
TE

0,45

0,40
EP

0,35

0,30
Soot content SC [%]
C

0,25

0,20
AC

0,15

0,10

0,05

0,00
0 100 200 300 400 500
Operating time [Mh]

Fig 4. SP content vs. operating time in Mh Fig 5. Soot content vs. operating time in Mh

The relationships between the content of particles which are essential from a point of view of
system condition assessment and the operating time expressed in days are illustrated in Figures 6–9.
 ACCEPTED MANUSCRIPT

PT
RI
U SC
Fig. 6. Na content vs. operating time in days
AN Fig. 7. Fe content vs. operating time in days
0,45

0,40
M

0,35

0,30
Soot content SC [%]
D

0,25
TE

0,20

0,15
EP

0,10

0,05
C

0,00
0 200 400 600 800 1000 1200 1400
Operating time [days]
AC

Fig. 8. SP content vs. operating time in days Fig. 9. Soot content vs. operating time in days
 ACCEPTED
4. Artificial neural networks MANUSCRIPT
– applied mathematical model

The proposed approach is based on the use of two specific types of neural networks: the MLP and
RBF. Based on the analysis of the literature and the authors’ experience [39–44], the first stage of the study
involved the use of the MLP. The MLP enables the modelling of functions of nearly any complexity. This
requires, naturally, an appropriate number of neurons and layers. Their number depends on the complexity
of the problem being considered; if experimental data is required, a large number of tests will have to be
conducted.

PT
One of the most widely used neuron models is that created in the 1940s by McCulloch and Pitts,
wherein the neuron is a binary unit. The McCulloch–Pitts model is the first mathematical description of the
neuron [45].
The output signal of the neuron yi is determined as:

RI
 = ∑  +  (1)

SC
where:
xj is the input signal for  ∈ 
wij are suitable weights of synaptic connections, here: signal direction from node j to node i

U
w0 is a threshold value, bias
AN
In this model, the activation function f(ui) is discrete, it is equal to 1 when ui is greater than 0, otherwise,
when ui ≤ 0, it equals 0. The output signal of adder ui for a perceptron is defined as:

 = ∑
M

 (2)

The aim of multilayer perceptron training is to determine the values of the weights wab in all layers in
D

order to obtain for a given input vector x output signals yi which correspond as much as possible to the
present values of di. The output signal of neuron i in the hidden layer can be described as:
TE

 = ∑   (3)

In the output layer of neuron k, the output signal has the form:
EP

 =  ∑


   = ∑

 
 ∑ 

  (4)
C

Weights of hidden layer neurons are denoted by the upper index ' while those of the output layer by '' [46].
AC

In the tested network, we used an algorithm for neural network training known as back propagation.
The back-propagation algorithm defines a method for selecting weights of multilayer network neurons using
gradient optimisation methods. The objective function of the algorithm is the sum of squares of differences
between the values of the network’s output signals and the present values [27].

An objective function for input nodes N with neurons K in the hidden layer and neurons M in the
output layer can be described as:

 
 = ∑%
 ! ∑

  

− # $ = ∑ % 
 &∑

  ∑ 
  − # ' (5)
 The other network used in the numerical analysis is the RBF. Networks of this type consist of a
ACCEPTED
single hidden layer with radial neurons and a linearMANUSCRIPT
output layer (with scalar product). The hidden layer
consists of radial neurons which, in the case in question, serve for the modelling of a Gaussian response
surface. A function of any shape can be modelled using only one hidden layer. This is due to the fact that
functions are nonlinear. For a network to create an effective model of a given function, it is necessary to
ensure a sufficient number of radial neurons in the network’s structure. As a consequence of a sufficient
number of radial neurons, every vital detail of the function being modelled can be ascribed a suitable radial
neuron, which results in obtaining a solution which represents the present function with satisfactory
accuracy.

In the radial basis function network, for the input vector x, the hidden neuron realises the function
changing radially around a selected centre c, ( x = ( ‖x − c‖, which takes non-zero values only in a

PT
neighbourhood of the centre. The input neuron, based on the hidden neurons radially representing the space
around individual points, represents a multidimensional space.
The approximate output of the RBF network, for K basis functions, can be expressed as:

RI
 = ∑
 ( ‖x − c ‖ (6)

SC
where ci is a set of determinable centres (i = 1, 2, 3,..., K).

For the applied Gaussian radial basis function with a centre at point ci we can define that:

U
‖/012 ‖3
( x = ( ‖x − c ‖ = exp .− 5 (7)
423
AN
where
 is the parameter determining a function width [46].
M

The scope of this paper covers the determination of vital components of the input vector of the
diagnostic signal. Based on the above assumptions, the processed data obtained from the laboratory analysis
will be used for the training, testing and validation of the artificial neural networks.
D
TE

4.1 Prediction of oil operating time (in Mh) using RBF and MLP
networks
EP

Given the effectiveness of the neural model, as shown in Table 2, one can observe that the content
C

of Na, Fe and Mo has the most significant effect on the MLP and RBF networks. The results of the
sensitivity analysis listed in Table 2 point to certain differences between the analysed network structures.
AC

For this reason, it was decided that successive sections of this paper will be devoted to separate numerical
analyses of the tested network structures and neural network training methods.
 Table 2. Sensitivity analysis results of the learning sample of a network system
Na Fe Mo Soot ACCEPTED
Sp Cr MANUSCRIPT
APP Pb NP Cu
Viscosity TBN
AxP Ignition
Network [ppm] [ppm] [ppm] [%] [ppm] [ppm] [%] [ppm] [ppm] [ppm] [%] [°C]
1.RBF 14-18-1 1.38315 1.22449 1.010618 1.433772 1.365922 1.002516 1.082428 1.014427 1.034584 0.993904 1.022431 0.972390 0.928994 1.018908
2.RBF 14-17-1 1.41290 1.34593 1.043086 1.492943 1.454354 1.070236 1.125525 0.992352 1.074683 1.021436 1.006521 0.998310 1.024452 1.018722
3.MLP 14-15-1 11.23579 10.46833 7.990665 4.752885 4.234095 3.889246 3.615035 3.303397 2.997006 2.270204 2.174474 2.087912 1.986993 1.423396
Average 4.67728 4.34625 3.348123 2.559867 2.351457 1.987332 1.940996 1.770058 1.702091 1.428515 1.401142 1.352871 1.313480 1.153675

Figures 10–13 illustrate the relationships between the operating times predicted by the radial basis
functions and the real operating times of the oil expressed in hours. If the fit were ideal, the points in the
plots would be located on a straight line. The real values of the points under the straight line are higher than
their predicted values, while to the left of the secant line the predicted values are higher than the real data.

PT
One must draw attention to the RBF 14-15-1 model (Figs. 12 and 13) which shows the most accurate fit of
the real and numerical data. The coefficient values listed in Table 2 also point to this model as being the
most accurate representation of the problem. One can see here a close relationship between the network and

RI
the input parameters.

U SC
AN
M
D
TE

Fig. 10. Predicted operating time vs. real operating Fig. 11. Predicted operating time vs. real operating
time for RBF network model time for RBF network model
C EP
AC

Fig. 12. Predicted operating time vs. real operating Fig. 13. Predicted operating time vs. real operating
time for RBF network system time for RBF network system
 ACCEPTED MANUSCRIPT
4.2 Prediction of oil operating time (in Mh) using MLP

Analysing the numerical results obtained with the perceptron network in relation to operating time, it must
be concluded that the particles which serve as input vectors for the network model are: Fe, Mo and Na. They
are typical particles describing the mechanical wear of kinematic parts of the tested system. At this stage one
may hypothesise that an effective model for assessing oil life can be primarily based on these parameters.
This hypothesis is confirmed by the results listed in Tables 3 and 4. Table 3 contains the values of
coefficients which provide information about the significance for a model of individual input variables. The
higher the value, the greater the significance of the variable for the model. The value of 1 is the boundary

PT
value, i.e., the variable is useless when its value is less than 1. At 98% efficiency prediction, the multilayer
network models reveal (Table 3) that the coefficient resulting from the sensitivity analysis for the above
mentioned elements justifies basing the system’s operation on their potential. As regards other factors such

RI
as soot, Sp, total base number (TBN) and APP, their effect on the network’s operation is much lower;
however, when their value is greater than 1, they can be considered useful. Taking into account the need to
design a universal system, it seems that these input variables can prove useful.

SC
Table 3. Sensitivity analysis of the MLP network

U
Coefficients
Fe Na Mo Soot Sp APP NP Cr Cu Pb AxP
AN
TBN Viscosity Ignition
Network [ppm] [ppm] [ppm] [%] [%] [%] [ppm] [ppm] [ppm] [ppm] [%]
1.MLP 14-18-1 26.42723 8.797089 7.655707 6.301260 3.156748 4.339834 3.249297 3.092994 3.356885 2.533647 2.222783 1.908717 2.077052 1.170414
2.MLP 14-15-1 3.04697 8.626455 2.924758 2.526548 2.719535 1.274193 1.803887 1.832758 1.556058 2.000148 1.610761 1.507340 1.250322 1.176304
3.MLP 14-12-1 7.39514 6.571596 4.411822 2.938764 2.268212 1.858505 2.209027 1.967467 1.738055 2.029506 1.836293 1.513617 1.322407 1.250048
M

Average 12.28978 7.998380 4.997429 3.922191 2.714832 2.490844 2.420737 2.297739 2.217000 2.187767 1.889946 1.643225 1.549927 1.198922

Table 4 compares the training quality of individual networks in terms of validation, train and test
D

sets. The closer the value is to 1, the closer to the ideal the quality of the neural network is.
TE

Table 4. Comparison of model quality

Quality Quality Quality Training Error Activation Activation
No. Network
(training) (testing) (validation) algorithm function (hidden) (output)
MLP 14-
EP

1 0.989833 0.963090 0.985736 BFGS 53 SOS Exponential Tanh

18-1
MLP 14-
2 0.981802 0.958737 0.981210 BFGS 40 SOS Exponential Tanh
15-1
C

MLP 14-
3 0.985388 0.961477 0.985808 BFGS 48 SOS Exponential Tanh
12-1
AC

Figures 14–17 illustrate the relationships between the operating times predicted by the MLP
networks and the real operating time, expressed in hours. The plots show a good fit between the
experimental data and the predictions of the neural systems. As was already mentioned, the closer to the
straight line it is, the better fit between the theoretical and experimental values. From the plots it can be
concluded that the network’s operation is correct for the specific data used for model training and its
validation.
 ACCEPTED MANUSCRIPT

PT
Fig. 14. Predicted vs. real operating time – MLP Fig. 15. Predicted vs. real operating time – MLP
network model network model

RI
U SC
AN
M

Fig. 16. Predicted vs. real operating time – MLP Fig. 17. Predicted vs. real operating time – MLP
network model network model
D
TE

4.3 Predicting oil operating time (in days) using an RBF network
EP

Similarly to the findings reported in the previous section, Tables 5 and 6 list values of the coefficients
describing the sensitivity and quality of the RBF models. They also provide information about error
C

functions, activation in the hidden and output layers, and the training algorithm. As mentioned previously,
the values of the quantities listed in Table 5 which are greater than 1 indicate the suitability of the variables
AC

for model operation. The higher the value is, the more significance the variable has. As expected, the quality
of system performance is lower than in the case of the models based on moto hours. The ageing of oil is a
long process, one that is difficult to be unambiguously described by means of particle contents in oil. The
results (Table 6) should however be considered correct. The quality of the RBF models for the validation
sets amounts to approx. 90%. It is also necessary to underline the significant effect of Sp particles and soot
content in the oil as well as ignition point and APP. Table 5 indicates that in the investigated case the model
is based on input data which differ from the input variables mentioned in Subsections 4.1 and 4.2. The
process of network training was implemented in compliance with the algorithm described in (6) and (7).
 Table 5. Sensitivity analysis as expressed in dimensional qualities
Ignitio ACCEPTED MANUSCRIPT
Sp Soot APP Cr Pb Na Viscosi Cu Mo NP Fe AxP
n TBN
[ppm] [ppm] [%] [ppm] [ppm] [ppm] ty [ppm] [ppm] [ppm] [ppm] [%]
Network [°C]
1.RBF 14- 1.3724 1.1424 1.1212 1.0655 1.1297 1.0807 1.0517 1.07169 1.0518 1.0152 1.0214 1.0273 1.0145 0.9899
28-1 08 93 79 42 72 42 23 6 25 74 77 45 15 50
2.RBF 14- 1.3808 1.1577 1.1263 1.1683 1.0808 1.0767 1.0340 1.00640 1.0182 1.0445 1.0267 1.0160 1.0121 0.9955
19-1 59 19 81 05 56 05 72 6 72 49 66 83 19 21
1.3766 1.1501 1.1238 1.1169 1.1053 1.0787 1.0428 1.03905 1.0350 1.0299 1.0241 1.0217 1.0133 0.9927
Average
33 06 30 24 14 23 98 1 48 11 22 14 17 36

Table 6. Quality of selected RBF models

PT
Quality Quality Quality
Network Network Training Error Activation
(training
ID name algorithm function Activation (output)
) (testing) (validation)

RI
(hidden)
0.95050
1 RBF 14-28-1 0.835204 0.885709 RBFT SOS Gaussian Linear
7

SC
0.98113
2 RBF 14-19-1 0.863279 0.903373 RBFT SOS Gaussian Linear
0

U
AN
M
D
TE
C EP
AC

Fig. 18. Predicted operating time vs. real operating time

Figures 18–20 reveal that the use of the RBF models to predict oil operating time yields positive
results. However, due to a lower quality of the network, the fit between the real and predicted values is not
as effective as that described in the previous sections; individual wrong cases can be observed. However,
taking into account the capability of neural networks for generalisation, the models in Table 6, especially the
network containing 19 neurons in the hidden layer, can be used effectively. The large number of neurons in
the network layer is due to the specific nature of this model and the high number of input variables.
 ACCEPTED MANUSCRIPT

PT
Fig. 19. Predicted operating time vs. real operating Fig. 20. Predicted operating time vs. real operating

RI
time for RBF network with 28 neurons in hidden time for RBF network with 19 neurons in hidden layer
layer

SC
4.4 Predicting oil operating time (in days) using an MLP network

U
AN
Tables 7 and 8 present the operation of the MLP models with 8 and 13 neurons in the hidden layer with
respect to the prediction of oil operating time. As can be seen in Table 7 and Figure 21, this appears to be a
more difficult task than was the case with predicting oil operating time expressed in moto hours. The values
M

of fit coefficients are lower. However, this problem is very important in terms of an operating time of the
system for signalling oil change in case of its extended use, at a short real operating time of the technical
object. Again, as was the case with the RBF network, the most important variables are Sp, soot and ignition
D

point. The effectiveness of other input vectors, especially for MLP 14-8-1, is questionable. It appears that
with such a large number of input variables, eight neurons in the hidden layer are not enough to render the
TE

nonlinearity of the problem. The networks have also different activation functions in the input and hidden
layers.
EP

Table 7. Sensitivity analysis of the best models of MLP network

Ignitio
Sp Soot APP Pb Cr Na NP Cu AxP Mo Viscosi Fe
n TBN
C

[ppm] [ppm] [%] [ppm] [ppm] [ppm] [ppm] [ppm] [%] [ppm] ty [ppm]
Network [°C]
1.MLP 14-8- 1.3139 1.1682 1.0593 0.9930 1.0592 0.9953 0.9997 1.0036 0.9872 0.9958 0.9811 0.99043 0.9764 0.9406
AC

1 52 51 64 24 88 25 67 27 92 02 81 0 56 39
2.MLP 14- 1.4386 1.2705 1.2581 1.1573 1.0522 1.0986 1.0777 1.0436 1.0194 1.0086 1.0214 1.00760 1.0092 1.0123
13-1 05 74 54 49 62 72 51 99 29 07 06 7 40 78
1.3762 1.2194 1.1587 1.0751 1.0557 1.0469 1.0387 1.0236 1.0033 1.0022 1.0012 0.99901 0.9928 0.9765
Average
79 12 59 87 75 98 59 63 60 05 94 8 48 09

Table 8. Training quality of the models with the highest quality coefficients
Quality Quality Quality Training Error Activation Activation
Network
(training) (testing) (validation) algorithm function (hidden) (output)
MLP 14-8-
0.795114 0.893245 0.889196 BFGS SOS Linear Exponential
1
MLP 14-
0.859980 0.939252 0.893553 BFGS SOS Logistic Logistic
13-1
 ACCEPTED MANUSCRIPT

PT
RI
SC
Fig. 21. Predicted vs. real operating time for selected MLP models

U
AN
4.5 Prediction of oil operating time (in Mh and days) using an MLP
M

Taking into consideration the numerical results of the previous approach, it was decided to develop a model
containing two neurons in the output layer, estimating both the operating time of the object expressed in Mh
and the total operating time, described as a period during which the oil is inside the technical object that is
D

ready for use. Table 9 lists the results of a sensitivity analysis of the two best MLP models. It can be
observed that the most important input variables of the two models are similar to those used for predicting
TE

operating time expressed in days. The particles of sodium, iron and molybdenum have lower coefficient
values than Sp and soot. This is an important conclusion drawn from the study especially that the coefficient
values for Sp and soot content listed in Tables 2 and 3 point to their significance in the two analysed cases.
EP

Table 9. Sensitivity analysis results for two best models of MLP network
C

Sp Soot Ignition APP Na Pb Cr Mo NP Cu Fe AxP

Viscosity TBN
Network [ppm] [ppm] [°C] [ppm] [ppm] [ppm] [ppm] [ppm] [ppm] [ppm] [ppm] [%]
AC

1.MLP 14-16-2 1.412737 1.461493 1.129979 1.169092 1.133397 1.162712 1.127489 1.113838 1.084249 1.097521 1.048031 1.032355 1.019373 1.023704
2.MLP 14-10-2 1.375089 1.115361 1.163643 1.099361 1.070598 1.032248 1.019342 1.020515 1.026343 1.009506 1.009629 1.002026 1.003186 0.976392
Average 1.393913 1.288427 1.146811 1.134227 1.101998 1.097480 1.073415 1.067177 1.055296 1.053513 1.028830 1.017190 1.011279 1.000048

As was already mentioned, amongst the analysed multilayer perceptrons, the most promising
models were those listed in Tables 9 and 10. The network learning method adopted was the Broyden-
Fletcher-Goldfarb-Shanno (BFGS) algorithm due to the fact that it is effective and at the same time
insensitive to all kinds of minimisation inaccuracies (in accordance with (1)–(5)). The analyses were
performed on the networks with a varying number of neurons in the hidden layer. According to Table 10, the
smallest error values can be observed for the models having 16 and 10 neurons, respectively, in the hidden
layer. Although a higher number of neurons ensures a higher agreement between the numerical results and
the real data, when the number of neurons in the hidden layer is too high, the model is exposed to
overtraining. This entails too much attachment of the network to the training data and thus the loss of
generalisation the model aims to achieve.
 ACCEPTED
Table 10. Quality of the applied MLP MANUSCRIPT
network models

Quality Quality Quality Learning Error Activation

Network Activation (output)
(training) (testing) (validation) algorithm function (hidden)
MLP 14-16-2 0.910823 0.951433 0.930144 BFGS 35 SOS Tanh Tanh
MLP 14-10-2 0.889318 0.944418 0.930112 BFGS 10 SOS Tanh Exponential

The neural network calculates the output values based on input variables. It must be remembered
that in the perceptron model, which is described by the number of input neurons, neurons in the hidden layer
and output neurons, respectively (e.g., MLP 14-16-2), weight values modified during the training process

PT
play an important role for specific connections between the neurons. This enables the approximation of very
complex nonlinearities.

RI
U SC
AN
M
D
TE

Fig. 22. Real operating time vs. MLP-predicted Fig. 23. Real operating time vs. MLP-predicted
operating time [days] operating time [Mh]
C EP

Figures 24–25 show plots illustrating the relationships between the individual input and output
values of the MLP neural network models. It was decided that, due to a high sensitivity coefficient, the first
AC

plots would illustrate the effect of Sp and soot content on the predicted operating time of oil (Figure 24).
 ACCEPTED MANUSCRIPT

PT
RI
SC
Fig. 24. SP and soot content vs. predicted operating
Fig. 25. APP and ignition point vs. predicted
time [Mh]
operating time [Mh]

U
AN
M
D
TE
EP

Fig. 26. Ground cut values of Sp and soot content

Fig. 27. Ground cut values of APP and ignition
during an operating time of 200 Mh
point during an operating time of 200 Mh
C
AC

Figures 26 and 27 show the plots illustrating the occurrence of individual particles during an
operating time of 200 Mh.
 ACCEPTED MANUSCRIPT

PT
Fig. 28. SP and soot content vs. predicted oil Fig. 29. APP content and ignition point vs. predicted

RI
operating time [days] oil operating time [days]

SC
Figures 28 and 29 illustrate the effect of the above input variables on the operating time [days] of oil.

U
4.6 Prediction of oil operating time (in Mh and days) using an RBF
network
AN
The last stage of this study involved the use of a radial base function network model to predict the operating
M

time expressed in days and Mh. The sensitivity analysis results are given in Tables 11 and 12. As in the
previous cases, the parameters which are the most useful in predicting the operating time of oil are Sp, APP
and soot content, as well as ignition point. This indicates that the experiments have been performed
D

correctly. Taking into account the results of sensitivity analysis, the network training quality coefficients
obtained for the selected RBF models (Table 12) should be considered correct.
TE

Table 11. Sensitivity analysis for a selection of RBF networks

EP

Sp Soot APP Ignition Cr Na Pb Mo Fe NP Viscosi Cu AxP

TBN
Network [%] [%] [%] [°C] [ppm] [ppm] [ppm] [ppm] [ppm] [ppm] ty [ppm] [%]
1.RBF 14- 1.2800 1.3045 1.2188 1.0910 1.0158 1.0759 1.0335 1.0708 1.0449 1.0613 1.04688 1.0273 1.0093 0.9832
C

27-2 66 89 56 93 37 04 02 76 49 07 3 92 83 39
2.RBF 14- 1.3555 1.1336 1.1722 1.0686 1.1372 1.0534 1.0940 1.0338 1.0443 1.0186 1.02687 1.0227 0.9671 0.9724
AC

15-2 35 11 30 97 51 10 86 69 38 24 0 95 78 07
1.3178 1.2191 1.1955 1.0798 1.0765 1.0646 1.0637 1.0523 1.0446 1.0399 1.03687 1.0250 0.9882 0.9778
Average
00 00 43 95 44 57 94 72 43 65 6 93 80 23

Table 12. Training quality of the tested RBF models

Quality Quality Quality Learning Error Activation Activation

Network
(training) (testing) (validation) algorithm function (hidden) (output)
RBF 14-27-2 0.904474 0.965036 0.942030 RBFT SOS Gaussian Linear
RBF 14-15-2 0.888787 0.960043 0.945991 RBFT SOS Gaussian Linear

The relationships between the operating time as predicted by the neural networks and the real
operating time are shown in Figures 30 and 31. The solid line in the plot denotes a perfect fit while the
position of individual point indicates the accuracy of fit of the data predicted by the neural model compared
to the real values obtained for eachACCEPTED MANUSCRIPT
case. As mentioned above, a more accurate fit is obtained for the
predicted operating time expressed in moto hours (Figure 31). As regards the predicted operating time
expressed in days (Figure 30), there are more cases in which the model’s prediction differs from the real
data.

PT
RI
U SC
Fig. 30. Operating time [days] vs. RBF model Fig. 31. Operating time [Mh] vs. RBF model
prediction prediction
AN
Figures 32–37 show plots illustrating the relationships between individual input and output values of
M

the RBF models.

D
TE
C EP
AC

Fig. 32. Sp particles and soot content vs. predicted Fig. 33. APP particles and ignition point vs.
oil operating time [Mh] predicted oil operating time [Mh]
 ACCEPTED MANUSCRIPT

PT
RI
Fig. 35. APP particles and ignition point vs. predicted oil
Fig. 34. Sp particles and soot content vs. operating time [days]
predicted oil operating time [days]

U SC
AN
M
D
TE

Fig. 36. Ground cut values of Sp particles and Fig. 37. Ground cut values of APP particles and ignition
soot content during an operating time of 720 point during an operating time of 720 days
days
EP

5. Discussion and conclusions

During the operation of a technical system, the concentration of single elements and substances
C

increases. This is not only due to effective machine operation, but also as a result of oil ageing. As the wear
curve shows, the concentration of some elements increases significantly at the beginning of operation (e.g.
AC

Fe) which might be later connected with an initial period. In other cases the amounts of various pollutants
rise in the oil independently of the system working period. When using conventional methods, however, it is
rather difficult to say whether the operating time or a total period of time during which the oil was used have
an impact on this situation.
The moment of soft failure occurs when the particle concentration reaches its critical value. In this
study it is estimated that this moment will occur at 200 Mh or 720 days as given by the maintenance
specification. The operating time of the system is then based on the input vectors and indications of the
designed network model.
Applying artificial neural networks, the authors show that predicting a likely oil operating time based
on the input data is possible. The NN approach introduced in this article is expected to foster ideas in
tribological data assessment. The main outcome is in application of NN to model particle concentration
values which in combination correspond to both operating and calendar time elapsed. The NN have also
been used to indicate the most significant pollutants and/or particle elements out of the whole vector of input
variables (see e.g. Tables 1,2,3 and 5) which are the most important for modelling both oil and system
condition. We would like to emphasise the fact that our results are based on a statistically valid set of oil
field data. Figures 26, 27, 36 and 37ACCEPTED MANUSCRIPT
show plots illustrating the occurrence of individual particles for the
tested operating times: 200 Mh and 720 days, respectively. It is assumed that the critical concentrations of
the particles with the highest informative potential will occur here. Based on the data given in the plots, it
can be assumed that when the critical values of, for example, ignition point and APP are exceeded, the
object should no longer be operated.
In the analysis at hand, we believe we can determine the moment of soft failure as it normally occurs
at either 200 Mh or 720 days. Therefore, we created as examples in Figures 26, 27, 36 and 37 the ground
cuts of the respective values and observed variables concentrations. The ground cuts correspond to the safe
set-up operation time for soft failure and to the values of combined observed variables concentrations. In
practice, however, the pollutant concentrations may vary depending on an operation style, an environment, a
driver’s behaviour, etc., and may not reach the critical value at exactly 200 Mh or 720 days. We assume this

PT
will occur much later, but, in principle, it may be at some other time than 200 Mh or 720 days. Therefore,
the value and novelty of this study lies in the utilisation of higher-dimensional observed variables vector (in
the form of pollutant concentrations) obtained from oil. The goal is to create an adaptive model which

RI
considers various levels of fuzziness and uncertainties. Therefore, the time to soft failure – as one of the
fundamental reliability measures – is estimated based on the moment when the set of observed pollutants
(not isolated but in the presented combinations) will reach their critical values, which in our models

SC
corresponds to either 200 Mh or 720 days. We consider this to be the main novelty and the most significant
contribution of our article. The outcomes could be developed a lot more when selecting carefully significant
oil characteristics, however, performing soft failure estimation based on the results presented in Figures 26,

U
27, 36, 37 is perfectly sufficient.
The remaining useful life of a system is the time interval from actual time point of a soft failure to
AN
the time point of hard failure, or generally the time interval from an observing time point to the time point of
hard failure. We are capable of estimating the time of soft failure but, unfortunately, we do not know the
time of hard failure occurrence yet. Therefore, the RUL cannot be exactly determined now. However, we
will continue developing approaches to study the RUL of this observed system.
M

The advantage of using artificial neural network models, which here are closely connected with
experimental data, is that they enable obtaining reliable results without the necessity of examining the
relationships between input and output variables. The neural models developed here are based on historical
D

oil sample data and have been positively verified. The conclusion is obviously based on the data obtained
from real systems, hence the authors are confident that all these results might be applied to technical systems
TE

of similar operating conditions.

It was proved that if the active oil operating time is closely connected with the increase in the amount
of sodium, iron or molybdenum elements, the oil ageing in redundant systems can be predicted on the basis
EP

of Sp or soot concentration. Using the calculations, we can conclude that when it comes to predicting the
operating time [Mh] or the calendar time [days] of motor oil in systems that are not working continuously,
we have to rely mainly on the percentage content of Sp, soot and APP, and the flash point.
C

This study draws upon a specific understanding of oil pollutants in terms of describing system
condition. Here we combine various observed variables which are significant for the purpose. The set of
AC

these variables includes the data which describe: (1) system condition (wear particles), (2) oil quality (oil
physical measurements) and (3) external contamination (external pollutants).

Given the tools and methodology for future diagnostic assessment, the design of system
maintenance structures, the modification of operational processes and the cost-effectiveness of service life,
the results can be considered promising. The study has demonstrated that a neural model based on the
historical data of diagnostic particles contained in oil can be used as a tool for the assessment of technical
object condition. The use of neural networks enables a multidimensional approach to the problem that is
extremely difficult to solve by means of mathematical description, while its numerical analysis is
complicated and time-consuming. The development of methods for quick and easy determination of the
effect of individual variables on system condition may lead to a significant reduction in costs and the
establishment of improved maintenance and operational policies.
Acknowledgements
ACCEPTED MANUSCRIPT
This article has been created with support of ‘Institutional Development Intention’ K202 _
MOBAUT, University of Defence, Brno, the Czech Republic.

References
[1] Labeau PE, Smidts C, and Swaminathan S. Dynamic reliability: Towards an integrated platform for
probabilistic risk assessment, Reliability Engineering and System Safety 2000; 68 (3): 219–254.

PT
[2] Park KS. Condition-based predictive maintenance by multiple logistic function, IEEE Transactions on
Reliability 1993; 42 (4): 556–560.

RI
[3] Maillart LM and Pollock SM. Cost-optimal condition-monitoring for predictive maintenance of 2-phase
systems, IEEE Transactions on Reliability 2002; 51 (3): 322–330.

SC
[4] Yang SK. A condition based failure-prediction and processing-scheme for preventive maintenance, IEEE
Transactions on Reliability 2003; 52 (3): 373–383.

U
[5] Li W, and Pham H. An inspection-maintenance model for systems with multiple competing processes,
AN
IEEE Transactions on Reliability 2005; 54 (2): 318–327.

[6] Mazurkiewicz D. Computer-aided maintenance and reliability management systems for conveyor belts,
Eksploatacja i Niezawodnosc – Maintenance and Reliability 2014; 16(3): 377–382.
M

[7] Vališ D, Koucký M, and Žák L. On approaches for non-direct determination of system deterioration,
Eksploatacja i Niezawodnosc – Maintenance and Reliability 2012; 14 (1): 33–41.
D

[8] Wang W. Plant residual time modelling based on observed variables in oil samples, Journal of the
TE

Operational Research Society 2009; 60: 789–796.

[9] Hosseini Sadegh, Ahmadi Najafabadi Mehdi n, Akhlaghi Mehdi. Classification of acoustic emission
signals generated from journal bearing at different lubrication conditions based on wavelet analysis in
EP

combination with artificial neural network and genetic algorithm, Tribology International 2016; 95: 426–
434.
C

[10] Zhu X, Zhong C, and Zhe J. Lubricating oil conditioning sensors for online machine health monitoring
– A review, Tribology International 2017; 109: 473–484.
AC

[11] Jardine AKS, Lin D, and Banjevic D. A review on machinery diagnostics and prognostics
implementing condition-based maintenance, Mechanical Systems and Signal Processing 2006; 20 (7): 1483–
1510.

[12] Phuc D, Voisin A, Levrat E, and Iung BA. Proactive condition-based maintenance strategy with both
perfect and imperfect maintenance actions, Reliability Engineering & System Safety; 133:22–32.

[13] Wang W, and Zhang W. A model to predict the residual life of aircraft engines based on oil analysis
data, Naval Logistic Research 2005; 52: 276–284.

[14] Wang W. A prognosis model for wear prediction based oil-based monitoring, Journal of the
Operational Research Society 2007; 58: 887–893.
[15] Zhao FY, Yu WB, and Su WH. Sensitivity study of engine soot forming using detailed soot modelling
ACCEPTED
oriented in soot surface growth dynamic, Fuel, 2016;MANUSCRIPT
198: 91–90.

[16] Lee J, Sung NW, and Huh KY. Prediction of soot particle size distribution for turbulent reacting flow in
a diesel engine, International Journal of Engine Research, 2011; 12(2): 181–189.

[17] Sheng, SW. Monitoring of wind turbine gearbox condition through oil and wear debris analysis: A full-
scale testing perspective, Tribology Transactions 2016; 59(1): 149–162.

[18] Cao W, Chen W, Dong GN, Wu JY, and Xie YB. Wear condition monitoring and working pattern
recognition of piston rings and cylinder liners using on-line visual ferrograph, Tribology Transactions 2014;
57(4): 690–699.

PT
[19] Clark RJ, and Fajardo CM. Assessment of the properties of internal combustion engine lubricants using
an onboard sensor, Tribology Transactions 2012; 55(4): 458–465.

RI
[20] Durdag K, and Andle J. Real-time viscosity measurement for condition-based monitoring using solid-
state viscosity sensor, Tribology Transactions 2008; 51(3): 296-302.

SC
[21] Valis D, Žák L, Pokora O, and Lánský P. Perspective analysis outcomes of selected tribodiagnostic data
used as input for condition based maintenance, Reliability Engineering and System Safety 2016; 145(1):

U
231–242. AN
[22] Vališ D, Zák L, and Pokora O. Failure prediction of diesel engine based on occurrence of selected wear
particles in oil, Engineering Failure Analysis 2015; 56: 501–511.

[23] Ghasemi A, Soumaya SY, and Ouali MS. Evaluating the reliability function and the mean residual life
M

for equipment with unobservable states, IEEE Transaction on Reliability 2010; 59 (2): 426–439.

[24] Glowacz A. Diagnostics of direct current machine based on analysis of acoustic signals with the use of
D

symlet wavelet transform and modified classifier based on words, Eksploatacja i Niezawodnosc –
Maintenance and Reliability 2014; 16 (4): 554–558.
TE

[25] Glowacz A. Diagnostics of synchronous motor based on analysis of acoustic signals with the use of line
spectral frequencies and k-nearest neighbor classifier, Archives of Acoustics 2014; 39 (2): 189–194.
EP

[26] Gebraeel N, and Pan J. Prognostic degradation models for computing and updating residual life
distributions in a time-varying environment, IEEE Transaction on Reliability 2008; 57 (4): 539–550.
C

[27] Haykin S. Neural Networks, a Comprehensive Foundation. Macmillan College Publishing Company,
New York, 1944.
AC

[28] Mohammadi R, Ghomi, S. M. T. Fatemi, and Zeinali F. A new hybrid evolutionary based RBF
networks method for forecasting time series: A case study of forecasting emergency supply demand time
series, Engineering Applications of Artificial Intelligence 2014; 36: 204–214.

[29] Balestrassi PP, Popova E, Paiva AP, and Marangon Lima JW: Design of experiments on neural
network’s training for nonlinear time series forecasting, Neurocomputing 2009; 72: 1160–1178.

[30] Donate JP, and Paulo Cortez P. Evolutionary optimization of sparsely connected and time-lagged neural
networks for time series forecasting, Applied Soft Computing 2014; 23: 432–443.

[31] Donate JP, Cortez P, Sanchez GG, and de Miguel, AS. Time series forecasting using a weighted cross-
validation evolutionary artificial neural network ensemble, Neurocomputing 2013; 109: 27–32.
[32] Leema N, Radha P, Vettivel SC, and Khanna Nehemiah H. Characterization, pore size measurement
ACCEPTED MANUSCRIPT
and wear model of a sintered Cu–W nano-composite using radial basis functional neural network, Materials
and Design 2015; 68: 195–206.

[33] Filipek P. Estimating air-fuel mixture composition in the fuel injection control process in an SI engine
using ionization signal in the combustion chamber. Eksploatacja i Niezawodnosc – Maintenance and
Reliability 2013; 15 (3): 259–265.

[34] Jedlinski L. Multi-channel registered data denoising using wavelet transform. Eksploatacja i
Niezawodnosc – Maintenance and Reliability 2012; 14 (2): 145–149.

[35] Jedlinski L, Caban J, Krzywonos L, Wierzbicki S, and Brumercík F. Application of the vibration signal

PT
in the diagnosis of the valve clearance of an internal combustion engine. Vibroengineering PROCEDIA
2014; 3: 14–19.

RI
[36] Gyurova LA, and Friedrich K. Artificial neural networks for predicting sliding friction and wear
properties of polyphenylene sulfide composites, Tribology International 2011; 44: 603–609.

SC
[37] Ramezani S, and Memariani A. A fuzzy rule based system for fault diagnosis, using oil analysis results,
International Journal of Industrial Engineering & Production Research 2011; 22 (2): 91–98.

U
[38] Vianna GK, and Thomé ACG. Neuro-fuzzy system for diagnosis of engines, based on oil samples
analysis, Jul 2014, ICAI'14 –The 2014 International Conference on Artificial Intelligence.
AN
[39] Gajewski J, and Jonak J. Towards the identification of worn picks on cutterdrums based on torque and
power signals using Artificial Neural Networks. Tunn Undergr Space Technol 2011; 26: 22–28.
M

[40] Gajewski J, Golewski P, and Sadowski T. Geometry optimization of a thin-walled element for an air
structure using hybrid system integrating artificial neural network and finite element method, Composite
Structures 2017; 159: 589–599.
D

[41] Golewski P, Gajewski J, and Sadowski T. Optimization of a thin-walled element geometry using a
TE

system integrating neural networks and finite element method, Arch. Metall. Mater. 62 (2017), 1, 445-452

[42] Jonak J, and Gajewski J. Identifying the cutting tool type used in excavations using neural networks,
Tunn Undergr Space Technol 2006; 21: 185–189.
EP

[43] Jonak J, and Gajewski J. Identification of ripping tool types with the use of characteristic statistical
parameters of time graphs, Tunn Undergr Space Technol 2008; 23: 18–24.
C

[44] Litak G, Gajewski J, Syta A, et al. Quantitative estimation of the tool wear effects in a ripping head by
AC

recurrence plots, Journal Of Theoretical And Applied Mechanics 2008; 46 (3): 521–530.

[45] McCulloch WS, and Pitts WH. A logical calculus of ideas immanent in nervous activity. Bull. Math.
Biophysics 1943; 5: 115–119

[46] Osowski S. Sieci neuronowe do przetwarzania informacji, Warszzawa: Oficyna Wydawnicza PW, 2013
[in polish].
 ACCEPTED MANUSCRIPT
We provide contribution to system failure occurrence based on real oil field data

We analysed specific oil characteristics which are correlated on more dimensions

We propose challenging input to preventive maintenance resulting in cost savings

Applications of special forms of neural networks are presented

PT
RI
U SC
AN
M
D
TE
C EP
AC