Detection of anomalies in up- and download speed and latency of Internet connections


This projects’ main goal is to detect Quality of Service issues in general end-user Internet connections. The application could potentially increase awareness and certainty of such issues. To achieve this a convolutional neural net (CNN) autoencoder is used to detect anomalies on three inputs. The inputs are upload speed, download speed and latency and a model is trained for each one separately. A detection is done by analyzing the prediction error of the model when feeding it with one of the metrics. To do so an error threshold is selected and data points in a certain temporal vicinity that are breaching that threshold are collected together. At a given size of such a collection it is considered to point to an anomaly. As a software foundation the Keras library, which is based on Tensorflow 2.3.1, will be used to construct and use the model. For training and testing a dataset that’s been recorded over a weeks time using a DSL/LTE Hybrid connection is used. The connection is inherently unstable and is hence useful for detection purposes. For training purposes it is filtered for anomalies beforehand. The results show a degree of applicability after tuning of the model but further experimenting would be needed to create a reliable and well-rounded application.