Using machine learning to monitor network performance

We describe a new approach, using machine learning, to automate performance monitoring in massively interconnected communications networks. The information obtained from monitoring network performance over time can be used to maintain the network preactively by detecting and predicting chronic failures and identifying potentially serious problems in the early stages before they degrade. We have applied this machine learning approach to the detection and prediction of chronic transmission faults in AT&T´s digital communications network. A windowing technique was applied to large volumes of diagnostic data, and these data were analyzed and decision rules were induced. A set of conditions has been found that is highly predictive of chronic circuit problems. Through continuous monitoring of the network at regular intervals using the new approach, we have also been able to identify several local network trends of specific chronic problems while they were in progress