Industrial IoT Edge Architecture for Machine and Deep Learning
3.
4.
5.
6.
via cached or locally generated content,
can be orders of magnitude greater than
from a core data center.
Data reduction: By running applications
such as ML/DL at the edge, operators and
application vendors can substantially cut
down the amount of data that has to be
sent upstream. This cuts costs and allows
for other applications to transfer data.
Isolation: A number of environments are
not always connected to the Internet over
high speed links. The edge is able to
provide services during periods of
degraded or lost connections.
Compliance: Edge applications can help
with privacy or data location laws.
Accuracy: Combine local results at the
edge with global results at the platform to
get a more comprehensive view of the
data in real-time.
Challenges
The new ideas have the following challenges.
1. Most edge gateways do not have the
computation capability to perform Deep
Learning. In those cases, we need to fit the
computation possible at the edge. These
can be simple rule-based methods and
auto thresholding algorithms or more
complex feature engineering and Machine
Learning algorithms such as decision tree
or logistic regression.
2. This can increase the cost of edge
hardware by requiring more computation
at the edge and drive processing away
from platform based computation. This
will increase investments that may not
have sufficient monetization.
3. This will require a new infrastructure for
processing, storing and securing data.
- 60 -
4. There does not exist many mature
orchestration and management solutions
for applications that involve the edge and
Platform tiers.
5. There are not enough ML/DL application
vendors for the edge and especially those
that cover both edge and Platform tiers.
The software platforms and ecosystems
are immature. There is a lack of expertise
and skills in this area.
Edge Computing Cost-Benefit
1. Third Party: Edge computing allows third
party applications to coexist with operator
applications at the edge. Third party
applications will unleash new innovations
and services like Machine and Deep
Learning at the edge.
2. Real-time: Some applications simply
cannot tolerate latency more than in the
order of 10s of milliseconds. Additionally,
these applications are also sensitive to
jitter (the variation in latency). Industry
4.0, which in turn depends on Machine
Learning will only be possible through low-
latency
connectivity
to
compute
resources.
3. Cost: ML/DL applications could run in the
cloud, but the upstream bandwidth
required to transmit the data to the cloud
to power them is not economical. Video
surveillance, face recognition, vehicle
recognition, IoT gateway are ML/DL
applications that belong to the edge. In an
IoT gateway, where even though the
bandwidth may not be high, sending
billions of events to the cloud would be
expensive and inefficient vs. handling
them at the edge with an IoT gateway,
especially considering that most of these
September 2017