IIC Journal of Innovation 11th Edition | Page 22

Outcomes, Insights and Best Practices from IIC Testbeds: Deep Learning Facility Tesbed In addition to a focus on energy efficiency, SAS aligns with Toshiba’s mission to focus on equipment, predicting equipment failures and finding equipment malfunctions—a difficult mission analogous to finding a needle in a haystack. IIC’s Deep Learning Facility Testbed provides the avenue for these companies to fulfill this vision. establish a distributed deep-learning system platform. There are areas of experimentation in the Deep Learning Facility Testbed. Modern buildings have many IoT sensor systems for various systems within the building. Sensors monitor large-asset maintenance, granular energy efficiency and occupant usage. AI and Deep Learning can use all of the data from these individual systems to achieve higher- order objectives such as energy sustainability and occupant satisfaction. In addition to being deployed at the Toshiba Smart Community Center in Kawasaki, the Deep Learning Facility Testbed is also implemented at the SAS Smart Campus in Cary, North Carolina, USA. This corporate headquarters has 24 buildings, ranging from two months to 20 years old. Though primarily an office campus, there are also daycare centers, fitness facilities and restaurants. While the Toshiba Kawasaki building employs greenfield instrumentation, the SAS site also has brownfield. The testbed mostly covers buildings created within the last six years, but there are also projects underway to retrofit some of its older buildings. The Deep Learning Facility Testbed’s Phase 1 target at the Toshiba Smart Community Center is sensor-anomaly detection. The building facility management team is concerned with the shortage of people to monitor, control and manage all the building assets, but AI can support these tasks automatically. Deep-learning technology assists the management team in detecting anomalies in sensors and assets and analyzing data from 30,000 data points. SAS uses neural networks on all of the building data. The number of different measurements from the SAS buildings were similar to that of the Toshiba Kawasaki building. Neural networks help create a model that understands building energy usage for monitoring the building to determine whether it is performing as expected in accordance with its design and understanding where and why deviations from those expectations may occur. Another area of experimentation is trying to determine what is computationally feasible and how to apply various high-end technologies to maximize those computation capabilities. For example, autoencoders have been used for anomaly detection. Deep learning needs a substantial amount of time for computing resources, and Dell’s rapid storage system helps The testbed plans to publish implementation lessons learned, best practices and use case results. The algorithmic knowledge and innovations discovered through the implementations of this testbed are also intended to be applied commercially by the participating companies. IIC Journal of Innovation - 18 -