Outcomes, Insights and Best Practices from IIC Testbeds: Deep Learning Facility Tesbed
A distributed learning system was used to
create parallel learning in the multiple
servers, enabling shared results of the
computations and learnings. Dell storage has
a very high network speed for exchanging
data, helping the many servers share their
data. This was done using computational
algorithms to scale the data and prevent
saturation that would otherwise be caused
by the number of servers. sensors have been installed on a couple of
older buildings on the SAS campus that
employ brownfield instrumentation. In some
ways, the older sensor technology is better
for targeting specific use cases, but it does
not match the breadth that accompanies the
newer buildings that have sensors on energy
circuits, water systems and some indoor
environment systems. Next steps may
involve retrofitting the buildings.
An autoencoder was used on the Toshiba
Kawasaki building for anomaly detection.
Because the building is new, there is little
defect data to be used to create a model for
anomaly detection. The autoencoder
therefore only uses data from normal
operating conditions, and if unusual data is
observed, it can be detected. In one case of
an anomaly detection, the autoencoder
picked up a dataset that had unusual data. A
data scientist was able to analyze the data
without any domain knowledge to find
which input affects the output of the
anomaly, and the building facility
management team confirmed which sensor
had the abnormal situation or defect. At the SAS campus, one industry standard
used for measuring energy efficiency is
energy intensity, which measures the energy
usage per square foot per year within a
facility. The SAS facilities are currently
running at 15 kWh per square foot per
year—a 3.3% improvement over the year
before.
Another area of technological achievement
is AI-at-scale. It became apparent that
employing one large neural network would
not give desired results as it would not
specify where to take action, nor indicate
what action was needed. Therefore, the
Deep Learning Facility Testbed employs a
large number of smaller neural network
models that can give insight into the
individual systems and designate on which
ones to act.
Phase 2 is underway with the introduction of
the SAS Smart Campus into the project. This
focuses on two of the newer buildings at the
SAS facility due to the large amount of useful
data they produce. Neural network models
have been running on those buildings for
over a year, and they were trained using
historical data that dated back about
another year. The anomaly detection
technology finds outliers, giving facility staff
an opportunity to address them. While the
number of anomalies cannot be predicted, a
target can be set for reducing energy
intensity in certain areas. Going forward,
IIC Journal of Innovation
SAS has some customer engagements
underway, one of which has been solidified
as a reference.
Key Challenges
From Toshiba’s perspective, one challenge
was attributed to the domain knowledge of
the building facility management team. In
typical deep-learning technology, supervised
learning is required, and so data must be
- 20 -