REPOZITORIJ > REZULTATI

Povzetek

Mechanical drives are the most ubiquitous items of equipment in almost all
industrial branches. Wear, excessive operational loads or errors in assembly
can cause premature unexpected failures resulting in partial or total production
downtime, damaged equipment or even loss of lives. Proper condition
monitoring of such equipment is therefore of great practical importance.
The field of condition monitoring of mechanical drives is very well developed.
Over the last decades, a large number of articles addressing every segment in
the field were published: mathematical models describing various faults; signal
processing methods for extraction of characteristic fault features and estimation
methods for the remaining useful life of the monitored drive. Notwithstanding
the significant results in every segment, several issues still remain
open. Probably the most notable are robust fault detection and isolation and
prognostics under variable and presumably unknown operating conditions.
This dissertation addresses the problem of robust feature extraction under
speed and load variations. Two novel techniques are suggested: one utilising
the probability density of vibration envelope signal and the other attempting
to identify specific vibrational patterns using the theory of temporal point processes.
On top of that, this thesis presents a possible solution to the problems
of fault isolation and evaluation based on evidential reasoning.
The analysis of the envelope distribution was done using wavelet packet coefficients
of the vibration signals. Based on these coefficients, alterations in
the distribution were quantified using Rényi entropy and α-Jensen divergence.
Features based on these quantifiers turned to be sufficiently sensitive to the
changes in the condition and, in the same time, they tend to be robust to the
variations in the operating regime.
Temporal point processes were employed to model the repetitive impacts patterns
that occur in cases of bearing surface damage. The analysis showed
that these impacts can be modelled as temporal point process with Inverse
Gaussian interevent distribution. The analysis of such a point process provides
an explanation why Rényi entropy quantifiers are suitable choice for
fault detection under variable operating conditions.
These two approaches greatly simplify the fault detection process since they
require no information about the operating conditions. In the same time,
the employed methods impose no limitations on the statistical properties of
the analysed signals, hence making them applicable both for stationary as
well as for non-stationary signals. Fault detection capabilities of the proposed
methods are demonstrated on a two-stage gearbox operating under different
rotational speeds and loads with various artificially built-in mechanical faults.
Finally, based on information contained in the calculated feature set, an evidential
reasoning based solution for fault isolation and evaluation is proposed.
This methodology mimics the reasoning process of the maintenance personnel
thus making it intuitive and open for seamless integration of experts knowledge.
The method fuses information contained in the feature set into a ranked
list of possible faults as well as an single overall assessment of the utility of the
monitored system. A prototype version of the system is validated on a test
batch of 130 electronically commutated motors, demonstrating high diagnostic
resolution and accuracy.