Abstract
Objectives
Motor failure in multi-leaf collimators is a common reason for unscheduled accelerator
maintenance, disrupting the workflow of a radiotherapy treatment centre. The work outlined in
this thesis sets out to develop and validate a data-driven model which would predict MLC
replacement needs ahead of time, allowing for proactive maintenance.
Novelty
Whilst there have been previous works which tackled the issue of using various data sources
to predict medical linear accelerators going out of specification, all the literature at the time of
writing focused on utilising parameters individually. This work takes a novel approach to the
issue by considering multivariate statistical approaches toward fault prediction and isolation.
This would allow to identify not only individual parameter errors but also identify
uncharacteristic parameter behaviour, in reference to the other parameters in the system. This
approach would allow for identifying faults which normally would go amiss, if their behaviour
was considered in isolation.
Results
A multivariate statistical model based on linear accelerator log files has been developed and
validated. It has been shown to detect between 61% and 73% of the multi leaf collimator
maintenance needs ahead of time, whilst being able to correctly identify which parameter is at
fault in a subsection of those cases.
Conclusions
It was shown that it is possible to utilise multivariate, data driven approaches to develop a
quality assurance check for proactive linac maintenance, albeit at the cost of a false alarm rate
similar to that of maintenance requirements.