Joel Abrahamsson and Markus Möller will present their master's thesis

Title: Investigating Audio Signatures of Pump Failure for Predictive Maintenance in Utility Environments

Authors: Joel Abrahamsson and Markus Möller

Supervisor: Anders Heyden

Examiner: Kalle Åström

Abstract: 

Damaged or partially damaged pumps can cause issues that at best is an
inconvenience and at worst a catastrophe. In order to prevent such issues
it is necessary to catch any potential damage as early as possible. One
possible solution is to use a sensor that can be placed on or near a pump
in order to monitor it using audio. Audio based fault detection is still a
relatively new field of research, with a wide verity of possible methods.
In this study we assess some of them and compare them in a verity
of tests in order to try and evaluate their strengths and weaknesses.
This is done using two datasets, the MIMII datasets and a dataset we
constructed ourselves that contain a verity of faults. Our results show
that it is possible to detect faults to a very high accuracy in a controlled
environment but difficulty arises when these models are used outside of
the controlled environment. Our conclusion is that fault detection using
audio and machine learning bears much potential but there are some
key challenges that must be either overcome or accommodated.