4. Déploiement du réseau LoRa ®
dans 18 agglomérations
françaises et progressivement
au niveau national
A fin S1 2016, dans 18 agglomérations
soit 1200 communes
A fin janvier 2017, dans 120
agglomérations soit 2600 communes
Capacité d’étendre cette couverture
par une offre site
8. 8 Interne Orange Orange Labs Research Exhibition 2016
B2B client-centric view of IoT
• multi tenant
• multiple services
• provide a B2B client centric view of the
functioning of the IoT network
• how are my set of devices functioning?
• KPIs adapted to the services
• All this on unlicensed bands for LPWA i.e. radio
frequencies that are free for everyone to use if a few
conditions are respected (transmit power, duty cycle,)
9. 9 Interne Orange Orange Labs Research Exhibition 2016
IoT NMS – Data analysis
• data collection
• data cleaning
• exploratory data analysis
visualization
scatter plot
correlation
• Modeling
• Machine learning
10. 10 Interne Orange Orange Labs Research Exhibition 2016
IoT NMS – machine learning aspects
Activity Machine Learning techniques Examples
Anomaly detection k-means clustering. DCs that need more analysis. Can be
extended to use external open data sets
such as road works and meteorological
inputs.
Model construction Random forest , Open ML in the
future?
Establishing the variables that most
influence a KPI (such as packet delivery
rate), which model to use?
Event prediction in
a streaming context
Incremental learning on non-
stationary streams – concept drift,
Adaptive Hoeffding Trees
The goal is for the model to adapt itself
dynamically to potentially changing
environments. The prediction is verified
against the real label and the model
adapted accordingly.
13. Challenges - stream processing
• Integration of ML libraries such as Samoa with
Stream processing engines
• Delayed/Missing labels
• Missing features – imputation?
• Concept Drift (change in seasons, new
building sites)
14. Challenges – system view
• Prediction model is at the level of devices or
links.
• How do we go from these atomic predictions
to network level and system level views?
• Use traffic pattern profiles and map low level
prediction to KPIs associated with the profiles