Design, implementation, and evaluation of learning algorithms for dynamic real‐time network monitoring

Abstract

Network monitoring is necessary so as to ensure high reliability and availability in telecom networks. One of the main challenges posed by state‐of‐the‐art monitoring tools is the creation of network baselines. Such baselines include thresholds that can be used to determine whether monitored values (with a given context, e.g., time) represent normal network operation or not. The size and complexity of current (and future) networks make it infeasible to manually determine and set baselines for each network operator and metric, let alone adapting the thresholds to changes in network conditions. This leads to the use of default baselines and/or setting baselines only once and never changing them throughout the lifetime of network elements. This does not only cause inefficient operation but could have implications for network reliability and availability. In this paper, we present the design, implementation, and evaluation of DARN: a collection of analytics and machine learning‐based algorithms aimed at ensuring that network baselines are automatically adapted to different metric evolution. DARN has been comprehensively evaluated on a deployment with real traffic to confirm accuracy of generated baselines, a 22% improvement in accuracy due to baseline adaptation and a 72% reduction in false alarms.

Workforce capacity planning for proactive troubleshooting in the Network Operations Center

“…Typically, the variance of the time series itself is used to compute confidence bands [25,26] whereby the actual value of the measure under study should lie in under the normal network performance [27]. And to predict such a normal performance, there is an extensive body of research in the area of time series forecasting.…”

“…The results are evaluated on self-supervised labeled traces, performing with an accuracy of nearly 100%. Finally, in [26], it is presented a real-time alerting system where the threshold of normality is estimated by exploiting also LSTM neural networks and validated with a dataset gathered from Facebook Prophet.…”

JADE: Data-Driven Automated Jammer Detection Framework for Operational Mobile Networks

“…Different approaches are taken for this thresholding in the literature. Some works do this manually, relying on human expert feedback to set/adapt thresholds (e.g., [27]- [29]) but this is unviable in our setting. Feng et al [30] simply use the median reconstruction error as the threshold, which is again not robust.…”

Towards the Automatic and Schedule-Aware Alerting of Internetwork Time Series

“…LSTM is an artificial recurrent neural network (RNN) that allows information from previous processing steps to persist and be part of the prediction of new inputs. Thanks to the property of identifying interdependencies in inputs, recurrent networks have been successfully applied in time series prediction [10], [29], [30], speech recognition [31], translation and language modeling [32], areas where context information is necessary to achieve the best results. Figure 2 illustrates the structure of the recurrent cells of an LSTM network [33].…”

“…For example, Prometheus [9] suggests considering time series as Gaussian processes and sets thresholds according to excursions over a certain number of standard deviations over/below the average. The authors in [10] refine that approach by defining the width according to the standard deviation of the noise of the time series-the difference between real and predicted values. Interestingly, the prediction is carried out through a trained Long Short-Term Memory (LSTM) neural network, similar to other works such as [11], [12].…”