This paper presents a master’s thesis project in which a system for anomaly detection on sound of a factory lathe has been developed and evaluated. The audio has been recorded with a microphone on site and has been analyzed using Fourier transforms and a Gated Recurrent Unit, developed to detect when this machine is running. An autoencoder has been used to determine if the gathered audio contains anomalies and thus indicates an error with the machine. The Gated Recurrent Unit has been evaluated using the metrics Precision, Recall and F1 score along with ROC curves and AUC, which has been used for comparison. To test the autoencoder, artificial anomalies has been generated and used to test if the algorithm gives a higher reconstruction error when these are present in the audio. Both neural networks shows promise, and with further development and training, could possibly work well in a real-life environment.
| Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:liu-162471 |
| Date | January 2019 |
| Creators | Berthold, Gabriel, Kinnvall, Jonas |
| Publisher | Linköpings universitet, Medie- och Informationsteknik, Linköpings universitet, Tekniska fakulteten, Linköpings universitet, Medie- och Informationsteknik, Linköpings universitet, Tekniska fakulteten |
| Source Sets | DiVA Archive at Upsalla University |
| Language | English |
| Detected Language | English |
| Type | Student thesis, info:eu-repo/semantics/bachelorThesis, text |
| Format | application/pdf |
| Rights | info:eu-repo/semantics/openAccess |
Page generated in 0.0015 seconds