Modern Techniques of Wear Debris Analysis
Ing.
Stanislav Machalík[1]
Abstract: Lubricants
used in mechanical parts must accomplish contradictory requirements on their
function in many cases and, at the same time, they must often work in extreme
conditions with longer service life. Increasing the reliability and economy of
machine use is closely connected to monitoring the condition and state of
technical parts of used lubricant with the purpose of diagnostics. Computer
image analysis of wear particles is a useful supporting tool for detail
analysis of oil samples. Presently, laboratory methods of analysing the individual
elements under a microscope are used most frequently. Modern methods, including
machine learning, provide possibilities of automatization of wear debris
analysis.
The
goal of this paper is to outline the possibilities
of improvement in image analysis with the help of automatic evaluation of wear
particles by using modern methods of artificial intelligence.
Keywords: image analysis, wear debris analysis, machine
learning methods, analytical ferrography
Constant progress of production and
development of engines results also in the growth of requirements on increasing
the power of aggregates on one hand and on decreasing of their size and weight,
which leads to increasing of their mechanical stress, on the other hand. These
requirements influence not only the properties of materials used at production,
but also the properties of lubricants that play a more important role from the
functionality point of view. The most frequent reason of failure is
insufficient lubrication of exposed parts.
The goal of this paper is to present
the possibilities of computer image analysis of particles that are extracted
from lubricants taken from a lubrication system. By evaluation of results of
analysis, it is possible to find suitable intervals for oil exchange, to
specify dominant wear mechanisms, to identify catastrophic particles and to
predict the beginning of critical wear leading to a damage. Wear particles are
characterized by their shape and size.
The most frequent techniques of
identification and correct diagnosis of wear debris elements is, apart from the
laboratory approach of identifying individual elements using a microscope, the usage
of neural networks or other methods of artificial intelligence that are used in
an element classifier employing expert systems which “learn” how to recognize
individual elements based on their morphological parameters or image patterns.
2 Acquisition and preparation of
oil samples for analysis
In
areas related to evaluation of operational materials and constructive materials
used in transport, it is possible to use the analysis of images acquired from
systems consisting of a microscope and a digital camera very effectively.
Among methods that are used most
often for obtaining high-quality bases (most often it is an image of lubricants
with scattered wear particles), analytical ferrography is used most extensively.
From the quality point of view, the results of the analysis depend on adhering
to specific procedures of oil sample extraction. The sample must be taken
always from the same place, always during the same system mode, before filtering,
at the time of dynamic balance. If there is any delay between taking the sample
and separating the particles, it is necessary to warm the sample up, to homogenize
it and to dilute the sample in order to acquire required viscosity, at which
wear particles are spread all over the area of the ferrograph. Further description
of the technology leading to the acquisition of applicable input data (images
of wear particles) is out of range of this paper. If the procedure is correct, the ferrogram
consists of particles extracted from lubricants spread according to their size as
a result of an inhomogeneous magnetic field being in effect. The
ferrogram is traditionally examined with a microscope, one of its important
parts being a video camera or a high-quality digital camera that allows scanning
images of the ferrogram with wear particles. These images are examined by
computer image analysis.
Presently,
methods leading to the identification of wear debris are being discovered. One
of the most important features of these methods is the independence on the
skills of staff; they are automated as much as possible.
2.1
Possibilities of data (wear particles) collection for classification
It is possible to use several
tools for collecting the wear particles which will be examined by image
analysis presently. They depend on the type of information that is the goal of the
analysis. One of the simplest methods is using an optical particle counter that
measures the quantity of light passing through the oil sample.
Most of the
contemporary systems used in tribodiagnostics for wear particle analysis are based
on the usage of neural networks that create a tree of rules for decision making
from morphological parameters of particles. As a result, an identification of
particles takes place without the necessity of human operation. Presently, the
state-of-the-art system for wear particle classification is the laser particle
counter, which is based on the usage of neural networks for “teaching” the system.
2.1.1 Ferrography
Ferrography
is a tribodiagnostic method based on separation of heterogeneous particles
included in oil filling of lubrication systems from the oil itself. It is based
on sedimentation of particles on a special bottom (film, glass board) during
the flow of the oil sample through a strong inhomogeneous magnetic field.
2.1.2 Optical particle counter
Optical
particle counter uses a special method of particle counting based on the analysis
of the light that is blocked by the particle. Each particle blocks the quantity
of light corresponding to its size. Based on the parameters of the ray of light
passing through the oil sample, an electrical signal is generated by the detector.
Changes of electrical signal are compared to a calibration table, according to
which the results (the count of particles and their size) are acquired.
2.1.3 Laser particle counter
Laser
particle counter (LNF, Laser Net Fines) uses a special laser technology and
advanced software based on neural networks for the analysis and particle
identification. After extracting the oil sample from lubrication system, the
LNF gets an oil sample using an automatic pump with the flow cell which is
lighted with a pulse laser diode that allows to obtain an image
documentation. The image is continuously scanned with a video camera with
magnifying optics. Photographs are subsequently examined with an application
program using a special neural network that is able to recognize specific types
of wear particles.
3 Modern methods of image analysis in
tribodiagnostics
Most
of the contemporary procedures used for wear particle analysis rely on the
skills of the operator. The influence of the human factor is indispensable, not
mentioning the time-consuming “manually operated” particle analysis. Modern
trends head for an automatic evaluation of wear particles contained in the image
based on parameters of particles themselves.
Presently,
one of the ways of possible development of automatic image analysis is the usage
of methods of applied artificial intelligence. Methods of machine learning that
allow searching the context hidden at the first sight also in very complicated
data having binary, visual or also numerical form seem to be prospective.
Machine learning methods are based on controlled or uncontrolled learning from
training data patterns using an adequate algorithm.
Among
the methods of machine learning that could simplify the process of particle
classification in tribodiagnostics significantly, neural networks seem to be
the most perspective; they are described further. There are also other suitable
methods worth mentioning but their introduction is not a part of this paper:
·
Cluster Analysis,
·
Principal Components Analysis (PCA),
·
Support Vector Machines (SVM),
·
Boosting and Adaptive Boosting
(AdaBoost).
3.1 Neural networks
Neural
networks are used as a base in most applications and tools that are used for an
automatic evaluation of the wear debris (particles) classification. The state-of-the-art-tool
– laser particle counter – is based on the principle of neural networks, too.
Neural
networks are beneficial for general diagnostics especially because of the fact
that many parameters can be followed in real time. The first presumption of
successful system development based on neural networks is the selection of
suitable parameters (in case of wear particles, these may be e.g. shape
factors), analysis of mutual relation of particles and wear type, influence of
parameter changes on type and range of wear and evaluation of usability of
these parameters for the next analysis. As the second step, the analysis of
possible states that can occur in the monitored device and a reasonable classification
of these states for the needs of the neural network is carried out. Together
with this analysis, the basic selection of a suitable type, topology and the
total arrangement of the neural network should be made. The next step of working
with the neural network is teaching the net using the set of measured data acquired
during the parameter analysis. Changes in topology and characteristics of neural
networks may occur as late as during the teaching. At the end of this process,
the system (neural network) should be able to classify the wear particles based
on the acquired knowledge of the particle parameters.
4 Conclusion
In
the area of wear particle image analysis, a fast development focused on several
areas is taking place presently. One of the most significant goals is to create
a system which would be able to perform the automatic wear particle
classification without the human factor influence (and without related
inaccuracies). The advantage of this solution is the fact that the data acquired
through the image analysis offer not only the diagnostic information, but also the
predictive information, i.e. they allow to foresee the situations of damage so
that they can be dealt with even before they happen.
Modern
methods of image analysis application in the area of tribodiagnostics help to acquire
information not only about basic characteristics of evaluated ferrograms or
single particles, but also lots of other data that would be very complicated or
almost impossible to obtain using common procedures. This information can
contribute to fundamental expansion of the knowledge about systems which the
oil samples were extracted from. Some particular applications of machine
learning methods seem to be very perspective.
[1] University of Pardubice, Jan Perner Transport Faculty, Department of Informatics in Transport, Studentská 95, 532 10 Pardubice, Czech Republic, tel.: (+420) 466 036 181, e-mail: stanislav.machalik@upce.cz