By Jude Ferreira, Comcast; Kevin Dugan, Comcast; Maher Harb, Comcast; Mike O’Dell, Comcast; Larry Wolcott, Comcast
Automated Proactive Network Management (PNM) is no longer an afterthought or a luxury but is considered table stakes when it comes to maintaining Comcast’s vast HFC (hybrid fiber-coaxial) network. Our networks experience a wide range of conditions that can degrade their performance over time. These conditions include issues such as loose connections between components, cracks and breakages in lines, disruptive energy, and signal impediments, all of which are inherent challenges in maintaining our constantly evolving network. The process of early detection and efficient mitigation minimizes service disruptions, reduces downtime, and leads to a better customer experience. Identifying the specific nature and root cause of network impairments also enables us to route repair technicians to the appropriate location, and reduces Mean Time to Repair (MTTR), thereby driving operational efficiency. Deploying OFDMA (Orthogonal Frequency Division Multiple Access) in the mid-split region of the spectrum has allowed us to offer ~3-10x higher upstream speeds to customers and is also an important steppingstone toward offering multi-gig symmetrical services using FDX (Full Duplex) under our 10G roadmap. D3.1 OFDMA allows the use of higher modulation levels up to 4096-Quadrature Amplitude Modulation (QAM) and provides up to 2x efficiency increases when compared to Single Carrier QAM levels of 64-QAM. Profile Management Application (PMA) systems are being used to manage OFDMA Profiles as described in our previous SCTE contribution [1]. However, PMA can also mask network impairments at a cost to capacity as an inherent component of its functionality. Therefore, it is critical to develop systems to perform PNM to reduce the impact of network impairments and enable the highest possible capacities and speeds. In this paper, we describe the use of Convolutional Neural Networks (CNNs) to identify network impairments within the mid-split region and share the current performance of our machine learning (ML)models. This effort is similar to our previous efforts to identify network impairments in OFDM and Downstream Single Carrier QAM (DS-SC-QAM) sections of the spectrum as described in our previous SCTE contributions [2,3].