Wireless Network Fault Localization via Metric Classification

The wireless ecosystem tends to be saturated and complicated due to a great number of network devices and applications deployed. Fault localization is critically important for the reliability of the wireless network system. However, it is difficult to localize reoccurring wireless network faults due to the inconsistent quality of monitoring metrics as well as the overlap of multiple fault root causes. This paper proposes FTCFL, a new wireless network fault localization framework based on feature engineering, the neural-network-based feature extractor and the time series classifier. FTCFL enhances raw metrics with incon-sistent quality based on feature engineering including the pseudo standardization, the Kalman filter and the feature selection. Then FTCFL applies the neural-network-based feature extractor and the time series classifier to localize reoccurring wireless network faults based on enhanced metrics. The experimental evaluation is performed in a real-world telecommunication dataset collected from live 5G wireless networks. The experimental results show that FTCFL is able to localize multiple fault root causes in the wireless network system with the MicroF1 of 0.876.