Anindya Dey Xiao Ling Adnan Syed Yuewen Zheng Bob Landowski David Anderson Kim Stuart Matthew E Tolentino Intelligent Platforms amp Architecture IPA Lab University of Washington 121216 ID: 802456
Download The PPT/PDF document "Namatad: Inferring Occupancy From Buildi..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Namatad: Inferring Occupancy From Building Sensors Using Machine Learning
Anindya Dey, Xiao Ling, Adnan Syed, Yuewen Zheng, Bob Landowski, David Anderson, Kim Stuart, Matthew E. TolentinoIntelligent Platforms & Architecture (IPA) LabUniversity of Washington
12/12/16
UW Intelligent Platforms & Architecture (IPA) Lab
1
Slide2Motivation
12/13/16UW Intelligent Platforms & Architecture (IPA) Lab 2
Collaborating on a project with Tacoma Fire Department to augment search and rescue operations.
Smart Buildings can assist by providing the location and count of the people inside
Straightforward approach
Instrument
the building with additional sensors(like video cameras etc
.)
Problems :
1
) privacy issues
, too many protocols (permissions)
2) additional cost of instrumentation and maintenance
Therefore, use existing environmental sensors and see how they can help with occupancy predictions
Slide312/12/16
UW Intelligent Platforms & Architecture (IPA) Lab 3Motivation
Cherry Parkes Building
Built 1890, LEED certified in 2003
Labs, faculty offices, classrooms, retail
Control System Sensors
Slide4Conclusion
12/13/16UW Intelligent Platforms & Architecture (IPA) Lab 4Namatad : Platform for managing and leveraging dynamic IoT data streams
.Accurately (95%) predict building occupancy using minimal environmental sensors
in Real TimeMaximize value of existing sensor deployments
Minimize cost of deploying new sensorsExtracting hidden
insights
Slide512/12/16
UW Intelligent Platforms & Architecture (IPA) Lab 5Current State Of Art
Message Broker
Persistent Data
Store
1
ETL (Extract Transform Load)
Model Training
Analytics / Predictions
2
Control
Systems
3
Slide6Issues with SOA
12/13/16UW Intelligent Platforms & Architecture (IPA) Lab 6Offline predictions = higher time to prediction
Lose temporal utility of IoT sensor dataIn
the fire hazard scenario latency = DEATHLimited
real time analytics within IoT data streams
Slide712/13/16
UW Intelligent Platforms & Architecture (IPA) Lab 7Real-time streaming analytics systemExtract value instantly before data gets staleManage
dynamic IoT data streams50B devices ?
Dynamically create analytical pipelinesAugment traditional building control systems
Create analytical models at multiple levels of granularity
Namatad Goals
Slide8Approach
12/13/16UW Intelligent Platforms & Architecture (IPA) Lab 8
Message Broker
Persistent Data
Store
1
ETL (Extract Transform Load)
Model Training
Analytics / Predictions
2
Control
Systems
3
ETL (Extract Transform Load)
Namatad
Predictions
Slide9Platform Components
12/13/16UW Intelligent Platforms & Architecture (IPA) Lab 9Leveraged OSS packages to put together end-to-end system
Apache AvroApache Kafka
Apache NiFiPython, R
Apache SparkConnectors to control systems
Slide10Initial Observations
12/12/16UW Intelligent Platforms & Architecture (IPA) Lab 10
Strong correlation CO2, Air Volume with Occupancy
Weak correlation
Room Temperature, Auxiliary Temperature with occupancy
Slide11Evaluation Methodology
12/13/16UW Intelligent Platforms & Architecture (IPA) Lab 11
Dataset : Environmental sensor readings taken every 5 minutesGround Truth : Class Schedule
Five classes of
occupancy:Class-0 : zero people in a
room
Class-1 : 1 - 5 people
Class-2 : 6 - 10 people , and so on
Features Used : CO2, Air Volume, Air Temperature, Room Temperature
Algorithms Used : Random Forest , KNN
Evaluated
the prediction accuracy of the model using 10-fold cross validation
.
Replayed data collected over 9 months to rerun the experiments repeatedly
Slide12Results
12/13/16UW Intelligent Platforms & Architecture (IPA) Lab 12Supervised Learning : Random
ForestTop feature : CO2The
data sets for each room differed in terms of sparsity.CP 103 : sparse, CP 106 : moderate, CP 108 : denseReflects heterogeneous nature of
IoT deployments.Prediction :
92% with CO2
only (scale)
95%
with
all sensor
types
KNN (Unsupervised Learning)
Could not find strong correlations with the installed sensor typesMaybe useful for uncovering patterns in newly deployed sensors
Slide13Future work
12/13/16UW Intelligent Platforms & Architecture (IPA) Lab 13Alternative streaming topologiestime
to insight effect on control
systemsscalingOther IoT use cases
Additional sensor typesAlternative machine learning algorithms
Slide14Conclusion
12/13/16UW Intelligent Platforms & Architecture (IPA) Lab 14Namatad : Platform for managing and leveraging dynamic IoT data streams
.Accurately (95%) predict building occupancy using minimal environmental sensors
in Real TimeMaximize value of existing sensor deployments
Minimize cost of deploying new sensorsExtracting hidden
insights
Slide15Contact Information
12/12/16UW Intelligent Platforms & Architecture (IPA) Lab 15Anindya Dey : andy1602@uw.eduProf Matthew Tolentino : metolent@uw.edu
Slide16Backup
12/13/1616UW Intelligent Platforms & Architecture (IPA) Lab
Slide17Deployment Options
12/13/16UW Intelligent Platforms & Architecture (IPA) Lab 17
Gateway
FOG Server
Cloud
Sensors
Namatad 1
Namatad 2
Namatad 3
Can be deployed anywhere
Hierarchical modeling between tiers