An Empirical Study on Root Cause Analysis and Prediction of Network Failure Using Deep Learning

This paper proposes a pragmatic deep learning-based method to not only analyze the root cause of network failure, but also predict possible failures in the microwave (MW) network utilizing alarms, status, and performance information of weather, terrain, and network equipment in South Korea. In general, it is incredibly difficult for network engineers to distinguish the temporal MW communication problem, which is originated from bad weather and is usually recovered automatically from the actual network failure caused by physical or software issues of the equipment. First, to solve the problem, we propose a deep learning-based MW Root Cause Analysis System (MW-RCAS) which performs RCA actions whenever a new network failure occurs. The system efficiently utilizes weather and terrain information in addition to data from the network, such as status, alarms, and performance data of each MW equipment. Compared with the 80% RCA accuracy of highly professional network engineers, the proposed MW RCA system shows 95% accuracy. Lastly, we propose the MW Network Failure Prediction System (MW-NFPS) based on an long short-term memory (LSTM) which can forecast future occurrences of MW communication failures at least 15 minutes in advance using weeklong data of network equipment, weather, and terrain.