Page 33 - Proceedings of the 2018 ITU Kaleidoscope
P. 33
UNSUPERVISED LEARNING FOR DETECTION OF LEAKAGE FROM THE HFC
NETWORK
Emilia Gibellini ; Claudio E. Righetti 1
1
1 Telecom Argentina, Argentina
ABSTRACT signals get into the HFC, there is also a possibility that part
of the signals that should be contained in the cable are
egressing to the air, bringing noise into the radioelectric
In the context of proactive maintenance of the HFC networks, spectrum. Consequently, the identification of ingress
cable operators count on Full-Band Capture (FBC) to necessarily leads to a proactive detection of leakage.
analyze the downstream spectrum and look for impairments.
There exists one particular type of impairment, which is The identification and fix of the flaws that cause impairments
ingress, likely to happen along with leakage. Therefore, the have always been an issue for the field service. CableLabs
detection of the former leads to the identification of the latter. refers to the full set of impairment identification capabilities
We collect data from FBC tool, and use unsupervised as DOCSIS Proactive Network Maintenance (PNM) [1].
machine learning to group cable modems such that the CableLabs’ InGeNeOS (Intelligent Generation-Next
signal they receive show common patterns. This allows a Operational Systems) working group has been working on -
characterization of all cable modems in a service group. and continues to work on- a variety of techniques based on
Then, we use the modems’ locations to determine whether DOCSIS (Data-Over-Cable Service Interface Specification)
the root cause of the flaw is inside the homes or not. [2] to deal with impairments to simplify these tasks and
improve efficiency.
Keywords - Machine learning, unsupervised learning,
pattern clustering, spectral analysis, content distribution Modern cable modems, more specifically DOCSIS 3.0 and
networks, signal processing algorithms. DOCSIS 3.1, have the capability to measure the spectrum of
a downstream signal using a high-speed A-D converter (e.g.
1. INTRODUCTION 2.5 Gsamples/sec). The chipmaker Broadcom announced in
2011 [3] the first fully digital “Full-Band Capture” tuner chip
Hybrid Fiber/Coax (HFC) is the term that describes the - able to tune anywhere in the 50 MHz to 1 GHz downstream
service delivery architecture used by cable operators and spectrum.
Multi System Operators (MSO). The architecture includes a
combination of fiber optic cabling and coaxial cabling to Full-Band Capture (FBC) allows cable operators to analyze
distribute video, data and voice content from the headend to the spectrum of cable modems. Technicians and engineers
the subscribers, and vice versa. Folds, breaks, corrosion of would look at the data collected by this tool, in real time, and
connectors, among others, cause noise and interference, and look for signs of spectral impairment. Cable operators are
distort the transmission on the coaxial. This means that the looking for alternatives to the visual analysis; efforts go
spectrum inside the HFC shows impairments. mainly towards machine learning as it provides an automatic
and hence more precise and time-efficient analysis of the
Many home devices emit signals on the radioelectric spectrum [4].
spectrum at frequencies that match the HFC’s upstream band
(5 to 42 MHz) and downstream bands (50 MHz to 1 GHz). It is part of our role as scientists to evangelize about machine
These signals could enter the cable system through poorly learning technology within our company. In order to do so,
shielded cables or through the communication devices we look for applications that draw on the most intuitive
attached to the cable network within the home, causing a algorithms. We have found that the use of intuitive
particular type of impairment, which we simply call ingress. algorithms allows us to transfer knowledge to other areas in
an effective way.
The type of ingress may vary according to the kind of
damage in the physical network. For instance, a broken In order to develop this tool, we apply a well-known
coaxial cable may act as a radio antenna, bringing into the unsupervised machine learning technique, which is the k-
HFC spectrum some trace of FM radio signals. means clustering algorithm to create an easily replicable
analysis. The advantage of using k-means is that we can find
We know that ingress and leakage occur simultaneously. this algorithm in almost any software. The ultimate goal is to
This is quite intuitive because whenever there is ingress, as group signals in such way that through the identification of
978-92-61-26921-0/CFP1868P-ART @ 2018 ITU – 17 – Kaleidoscope