Turbine blades are subjected to various damage mechanisms with fatigue as the primary contributor. During
operation, damage accumulates in the form of crack initiation and propagation. This may lead to catastrophic
failure, which is cause for concern in terms of availability and safety of the turbine. To optimize
the maintenance schedule and to provide operational
exibility of the turbine, the state of health of the
blades is monitored. This is usually accomplished through non-destructive testing (NDT) during outages.
Conventional NDT techniques for in-situ inspection of turbine blade and disk assemblies is di cult and
often ine ective, due to limited access to areas of concern, as well as the complex geometries of blade roots.
O -site inspection can be costly if the blades are still assembled in the turbine disk since the process of
removing and reinstalling these blades is critical and labour-intensive, increasing the turbine downtime and
These problems could potentially be overcome by employing inspection techniques that o er the prospect
of assessing obstructed areas through monitoring the global dynamic characteristics of the structure, which
provide relatively easily interpretable data. With global inspection methods, a degree of measurement
sensitivity is forfeited but the potential to detect more severe damage, without prior knowledge of the
precise damage area location, exists.
In this dissertation, the feasibility of a vibration-based structural damage identi cation technique that could
be usable in support of conventional NDT to detect cracks in pinned turbine blades during o -line in-situ
inspection, is evaluated. The investigation was limited to considering uninstalled single blades only, and
thus o -site inspection of this component is regarded above the turbine disk assembly. This is clearly a
simpli ed case and does not address the critical case from a practical perspective of having a large number
of blades mounted onto a disk with pins, which is really the circumstance under which the technique could
become useful. This study must thus be considered as a rst step towards addressing the real practical
problem. In this simpli ed problem, the following questions are answered: Is it possible to detect damage
in an unconstrained and isolated blade using vibration response, and if so, can di erent damage scenarios
be identi ed? The proposed vibration-based damage detection method entails a multi-class support vector
machine classi cation procedure in which the natural frequencies are employed as the discriminatory feature
for damage detection and identi cation of di erent single-location damage scenarios.
The natural frequencies were acquired from accurate experimental modal analysis of freely supported individual
pinned turbine blades through impact testing. To con rm and predict the expected behaviour of the
blades, a healthy numerical model was built and validated whereafter defects and damage were introduced.
This includes geometrical variability at the root, observed in the procured blades, and the anticipated worstcase
single-location damage at the most probable locations near or on the root, obtained from literature
and discussions with experts in the industry. Arti cial damage, i.e. a uniform 1mm notch, was introduced
in the root at the upper pinhole on the leading edge pressure side; and just above the root at the aerofoil
base on the trailing- and the leading edge. To establish the discriminative quality of the modal property
natural frequency, it was necessary to determine its sensitivity to geometrical variability and damage. It
was also required to establish the damage-speci c behaviour or damage trend in the experimental data of
Feasibility of Vibration-based Damage Detection for Pinned Turbine Blades
these damage scenarios to conclude their distinctiveness. This analysis was extended to outlining the feature
quality by exploring the separability of class clusters for the healthy and damage scenario(s).
The feasibility of the proposed method is assessed using experimental data through simple hypothesis testing
regarding the detection and identi cation of both geometrical variability in healthy blades, and damage. It
was found that healthy blades are very similar, as geometrical variability cannot be detected. This is because
the distributions of natural frequencies fall within a range about a mean value in an ambiguous cluster. In
contrast to this, the damage scenarios were found to be distinct, and thus discernible from the healthy
blades. These classes formed discrete clusters, each with a similar distribution than the healthy blades. The
conclusion of the feasibility study serves as proof of concept.
Dissertation (MEng)--University of Pretoria, 2018.