Abnormality detection in diagnostics data from network cameras

Sammanfattning: For data-driven companies, there is a need to efficiently navigate through big quantities of collected data. In our case; detecting changes in the behaviour of data. We have investigated whether machine learning could be applied to automate the process of finding abnormal behaviour (anomalies) in collected data. Some simpler methods were also investigated and compared to the results of machine learning methods and the manual findings. In order to compensate for low-quality data, a method was created to synthesize data similar to the real data. Our results suggests that some methods did a better job than others on the anomaly-finding task. The overall most accurate method was the forecasting method Theta Forecasting closely followed by non-machine learning algorithms. The results also suggests that training the model on the synthesized data is worse than only using it to tune the hyper parameters. Future work could investigate additional algorithms and synthesis in order to raise performance even further.