Convolutional Neural Networks for Proactive Network Management (2020)

By Jude Ferreira, Maher Harb, Karthik Subramanya, Bryan Santangelo & Dan Rice, Comcast

The signal quality on HFC networks can degrade over time, from an impairment perspective, if not proactively maintained. Comcast manages hundreds of thousands of miles of network throughout the world in which we experience a wide array of conditions that can degrade the performance of the network. From connections loosening and cracks forming, to lines getting cut, destructive energy and signal impediments are part of maintaining modern networks. Early detection, mitigation, and routing fix agents efficiently, to the right location, can not only improve customer experience but also enable operational efficiency.

Adaptive Profile Management Application (PMA) systems continue to be deployed to manage downstream and upstream network capacity and network stability. However, and perhaps ironically, PMA systems can mask the degradation of the network, as an inherent function of the optimization and mitigation process. It therefore has become increasingly important to develop systems that can support automated Proactive Network Maintenance (PNM) to reduce the impact of network impairments on customer experience and enable the highest possible capacity and performance.

In this regard, Comcast has invested heavily in data platforms and data science functions across organizations, to become more data driven and to incorporate Machine Learning (ML) approaches into the network. In this paper, we will describe the use of Convolutional Neural Networks (CNNs) to identify network impairments within DOCSIS 3.1 (D3.1) channels with a high degree of accuracy. It is beneficial to classify the various network impairments (shown in Figure 1) as they may warrant different responses from techs in the field to enable fastest possible Mean Time To Repair (MTTR). In addition, clustering of these impairments across geographic locations and network topology may be exploited to identify the root cause impacting multiple customers that share common points in the network. Note that the latter requires a second layer model to be built on top of the classification model described in the paper.

The model we describe improves on the rule-based approaches currently being used to identify Mobile Wireless Ingress and Sweep Generator patterns. Notifications from the current rule-based model for detecting Mobile Wireless Ingress are sent across a notification bus to other Comcast OSS tools, to ensure that technicians are dispatched to the right hubs, network segments, and homes to remediate issues. The rule-based approach also provides a baseline for evaluation for the ML approaches. We also developed a real time version of the algorithm that techs can use to check that issues have been fixed after remediation. The same workflow described in this paper could also be used with alternate CNN-based models.

By clicking the "Download Paper" button, you are agreeing to our terms and conditions.

Similar Papers

A Machine Learning Pipeline for D3.1 Profile Management
By Maher Harb, Jude Ferreira, Dan Rice, Bryan Santangelo & Rick Spanbauer, Comcast
Full Scale Deployment of PMA
By Maher Harb, Bryan Santangelo, Dan Rice & Jude Ferreira, Comcast
Deploying PMA-Enabled OFDMA in Mid-Split and High-Split
By Maher Harb, Dan Rice, Kevin Dugan, Jude Ferreira, Robert Lund; Comcast
How Network Topology Impacts Rf Performance: A Study Powered By Graph Representation Of The Access Network
By Maher Harb, Karthik Subramanya, Ramya Narayanaswamy, Sanket Walavalkar & Dan Rice, Comcast
Practical Implementation of Profile Management Application
By Brady Volpe, NimbleThis and The VolpeFirm
A Deep Learning Approach for Detecting RF Spectrum Impairments and Conducting Root Cause Analysis
By Kevin Dugan, Justin Evans, Maher Harb; Comcast
What Gets Measured Gets Done / What Gets Analyzed Gets Transformed: Analytics for a Wider/Deeper Network View
By Venk Mutalik, Dan Rice, Karthik Subramanya & Jon-en Wang, Comcast
Optimizing DOCSIS 3.0 Configuration in the Upstream through Applied Reinforcement Learning
By Kevin Dugan, Maher Harb, Dan Rice & Robert Lund, Comcast
Photon Avatars in the Comcast Cosmos: An End-to-End View of Comcast Core, Metro and Access Networks
By Venk Mutalik, Steve Ruppa, Fred Bartholf, Bob Gaydos, Steve Surdam, Amarildo Vieira, Dan Rice; Comcast
Time is Ripe for D3.0 Farming: Achieving Optimal Spectral Efficiency by Allocating D3.0 Spectrum to OFDM
By Maher Harb, Chad Humble, Sebnem Ozer, Dan Rice; Comcast
More Results >>