The purpose of this thesis is to analyze the process of creating a reliable mechanical audio anomaly detection system, thus contributing to research regarding how to shape the system into rendering reliable results with a low quantity of training data. The aim is to research how the performance of said system change depending on both the quantity and the type of training data provided.

To fulfill the purpose, both a prestudy and a experiment was conducted. The prestudy’s aim was to investigate audio features and determine which features were suitable for use in a mechanical audio anomaly detection system. The experiment consisted of a evaluation of the system, where it was trained on data of varying type and quantity. The data was collected from automatic gates provided by the company ITAB.

The feature prestudy investigated 27 feature values where 12 were determined to be valuable for the system. The experiment showed that the developed system performed with great accuracy. There was generally a lower accuracy when the system was trained with a low quantity of data, however it rapidly improved and the accuracy reaches to about 95 percent between 35-50 training samples.

In regards to the given results, the authors conclude that a reliable1 audio anomaly detection system has been created. The authors can also conclude that the type of training data, in this use case, do not greatly affect the performance of the systems, thus rendering viable even without prior knowledge regarding faults in the mechanical system.