Page 39 - Proceedings of the 2018 ITU Kaleidoscope

Machine learning for a 5G future

In C4, there is radio, as well as 3G-downlink signal ingress. 7. AKNOWLEDGMENTS
These cases show tilt, which is another type of spectral
impairment. The cable modems in clusters C2, C3 and C6 do We thank Gabriel Davila Revelo and José Machao, who took
not show clear evidence of impairment; the group C2 time to induce us into some fundamental PNM concepts, as
contains noisier or more variable signals, but no evident well as provided feedback on our interpretation of results.
pattern. C3 contains the cases of Internet service only.
8. REFERENCES
We end our analysis with a look at the spatial location of the
cable modems in the C1, C4 and C5 groups. The points are [1] CableLabs. “PNM Best Practices: HFC Networks
sparse and occupy the total area of the service group, so we (DOCSIS 3.0)”, July 2016. [ONLINE] Available at:
conclude that there might be some damage outside the clients’ https://apps.cablelabs.com/specification/proactive-
homes. network-maintenance-using-pre-equalization [Latest
retrieved June 20, 2018].
6. CONCLUSIONS
[2] CableLabs. “DOCSIS 3.0 Operations Support System
Cable operators are looking to transmit signals beyond 1 Interface Specification”, July 2017. [ONLINE]
GHz on the HFC network, which may lead to leakage to Available at:
some 5G bands. Consequently, it becomes necessary to open https://apps.cablelabs.com/specification/CM-SP-
the discussion on ingress/leakage detection in 5G. OSSIv3.0 [Latest retrieved June 20, 2018].

We used the k-means clustering algorithm and the data [3] Telecompaper. “Broadcom launches digital Full-
retrieved by the FBC tool to improve efficiency in spectral Band Capture technology”, June 2011. [ONLINE]
analysis. We carried out data processing as well as clusters Available at:
calculation using open software. https://www.telecompaper.com/news/broadcom-
launches-digital-full-band-capture-technology--
In order to inform about potential sources of leakage, the 809318 [Latest retrieved June 20, 2018].
focus is on the frequency bands where we know that it is
possible that ingress takes place. This eliminates any [4] Dorairaj, S.; Burg, B.; Pinckernell, N., Comcast
unnecessary computational effort. The scaling of data points Corporation; Bastian, C., SCTE. “Simplifying Field
was key to get an algorithm that groups signals according to Operations Using Machine Learning”, U.S., 2017.
patterns, and not according to signal strength levels.
[5] ENACOM. “Cuadro de Atribución de Bandas de
In the context of PNM framework, our new tool helped us Frecuencias de la República Argentina (CABFRA)”.
achieve a rapid identification of six common patterns, three Argentina, pp. 101-185. April 2018.
of which involve ingress of FM radios, digital TV and 3-G
downlink signals. We do not find evidence of LTE ingress in [6] Society of Cable and Telecommunications Engineers
this service group. Considering the location of the clusters in (SCTE). “DOCSIS 3.0 Part 2: MAC and Upper Layer
the city map, we have concluded that the cause of ingress Protocols”, Exton, Pennsylvania, U.S., pp. 30. 2013.
was not inside the subscribers’ homes.
[7] R Core Team. “R: A Language and Environment for
This tool works as a successful example of machine learning Statistical Computing”, R Foundation for Statistical
application among the teams of field service, and we hope Computing, Vienna, Austria, 2017.
that this easily replicable PoC may also work as a
recommended proceeding for cable operators to detect [8] Venables, W. N.; Ripley, B. D., “Modern Applied
leakage on their networks. Statistics with S”, Springer, New York, U. S., 2002.

On the other hand, we want to remark that it is an early stage [9] Peña, D. “Análisis de datos multivariantes”,
development, and the analysis is not finished. McGraw-Hill/Interamericana de España, Spain, pp.
228-231. 2002
We will re-use the identified patterns to train a supervised
algorithm that detects the impairments on another set of [10] Gove, R. “Using the elbow method to determine the
modems. We will also continue to investigate the impact of optimal number of clusters for k-means clustering”
the number of samples per modem, as well as the variation [ONLINE] Available at:
of signal patterns in time, in order to determine the minimum https://bl.ocks.org/rpgove/0060ff3b656618e9136b
conditions that need to be met to apply the method. We [Latest retrieved June 20, 2018]
expect that in time this could lead to a standardization of the
measurement process. [11] Lebart, L.; Morineau, A.; Piron, M. “Statistique
Exploratoire Multidimensionnelle”, Dunod, Paris,
France, pp. 152-154. 1995.

– 23 –