Machine Vision Based Wireless Link Layer Anomaly Characterization

As the number of wireless end and edge devices increases, so does the volume of data to be monitored in view of predicting or detecting malfunctions. Furthermore, as the networks become more complex, the more context can be provided around a certain anomaly, fault or malfunction, the easier will be to establish mitigation actions in a fast and efficient manner. While detecting and classifying shapes of anomalous link behaviour from time series has already been investigated, the precise localization in time and characterization in duration and amplitude from actual data traces that would provide more context related to the anomaly is outstanding.

In this paper, we study the performance of time-series to image transformation techniques combined with machine vision to precisely localize, describe and classify link layer anomalies in wireless networks. Our evaluation shows that the proposed approach is able to characterize an anomaly by precisely localizing its start and end, as well as its amplitude with mean absolute error of less than 1%. Furthermore, the classification scores of the method are comparable with state of the art classifiers, reaching overall F1 scores of 0.97. This is the first attempt at characterizing anomalous shapes that appear in networking time-series data with potential applications in real-time monitoring systems, predictive maintenance where more context may lead to better prioritization and faster mitigation.