We present the design, implementation and evaluation of an enabling platform for locating and querying physical objects using existing WiFi network. We propose the use of WiFi management probes as a data transport mechanism for physical objects that are tagged with WiFi-enabled accelerometers and are capable of determining their state-of-use based on motion signatures. A local WiFi gateway captures these probes emitted from the connected objects and stores them locally after annotating them with a coarse grained location estimate using a proximity ranging algorithm. External applications can query the aggregated views of state-of-use and location traces of connected objects through a cloud-based query server. We present the technical architecture and algorithms of the proposed platform together with a prototype personal object analytics application and assess the feasibility of our different design decisions. This work makes important contributions by demonstrating that it is possible to build a pure network-based IoT analytics platform with only location and motion signatures of connected objects, and that the WiFi network is the key enabler for the future IoT applications.
3. Utku Acer, Aidan Boran, Claudio Forlivesi, Werner Liekens, Fernando Perez-cruz and Fahim Kawsar
Internet of Things Research
Sensing WiFi Network for Personal IoT Analytics
Presented by: Till Riedel, KIT
The IoT has arrived
3
4. Waste Pollution Traffic
Quantified Self Quantified Home Quantified City
Search for physical objects’ location and state is one of the basic services that provides foundation for many applications.
4
5. What we learnt in the past
Dedicated Sensing Infrastructure (ZigBee, RFID, Mote, etc.)
High deployment and management costs
Search range is limited to the smart phones’ proximity
Bluetooth Discovery with Smart Phone
5
6. WiFi is the most Pervasive Sensor Network. Its available literally everywhere.
6
7. Can WiFi network be used as a platform for personal IoT analytics?
Research Objective
Premise
Connected objects’ movement data extracted from WiFi network signals carries vital
information to model their spatio-temporal usage pattern
7
8. Key Design Decisions
Proximity
Ranging
Privacy Aware
Architecture
WiFi based Proximity Ranging with high accuracy and minimum
energy overhead.
Rich Semantic Location Labels for Better User Experience
Seamless
Ingestion
Uniform Features for Scale and Heterogeneity
Management Location and State (of Use) Aware Physical
Things.
Scalable Cloudlet Driven Architecture
Node.js + WebSocket
Data Locality
Privacy by Design
Dynamic Light Weight In-Memory DB for Spatio-Temporal-
State Data Representation. Data remains in the local
gateway.
Using WiFi Management Probes for Data Transport to
Minimise Energy expenditure
8
10. ôMobile Object Tags
Static Object Tags
Home Node
- Proximity Ranging Service
- Data Storage Service
EF5
Personal Object AnalyticsQuery Server
Anchor Points Query Service
Index Service
System Architecture
System Components
Object Tags
Attached to physical objects and emit the location and state-of-use of the physical objects.
Home Node
Hosted in the residential home gateways, provide proximity ranging service, and stores objects data.
Query Server
Hosted in the cloud, maintains persistent connection with home node and provides query interface to
personal analytics applications.
10
11. Object Tags
• Mobile Object Tags
• Attached to mobile objects
• Model state-of-use based on motion signature
• Transport data using WiFi probes
• Static Object Tags
• Attached to static objects
• Model state-of-use based on vibration signature
• Capture probes from mobile tags and forward to home node.
Prototype mobile tag with ESP8266 SoC with
integrated WiFi, flash, MCU and accelerometer
(Freescale MMA8452Q).
Prototype static objects tag with Raspberry PI. The
tags are powered and have two identical WiFi
interfaces (Ralink 5370), one for capturing probes
and the other for the connection to home node.
• Data Transport by WiFi probes for minimum energy expenditure
The object tags leverage the IEEE 802.11 standard’s management
frames to propagate the states to the home node.
We use a dedicated SSID and a single bit in the probe in that
SSID value to reflect the state of the object.
11
12. Home Node
• Home node is implemented with a Meshlium Gateway from Libelium
• Features a 500MHz x86 processor
• Two WiFi interfaces
• Runs an embedded Debian Linux operating system
• Both services are implemented with Node.js and a SQLite database is
used to store data.
Proximity Ranging Service
• Collects probes from object tags and annotates them with a coarse-grained location using a One-vs-One SVM
Classifier using a universal Gaussian kernel.
Data Storage Service
• Stores the objects location and state of use data in a local storage.
The space is divided into K overlapping zones, where each zone is represented by a Static Object Tag, here Xi
represents the vector of the RSSI values of a mobile object tag received by different static objects tags.
12
13. Query Server
ô EF5
Personal Object AnalyticsQuery Server
Anchor Points Query Service
Index Service
The query server and its components are implemented in Node.js with mongoDB and runs in Docker container.
Anchor Points
• An anchor point maintains a persistent connection with a home node using web sockets with failover support.
• Over this connection, an anchor point forwards queries to a home node and receives the query responses.
Index Service
• A B+ based key-value store that is used to locate the home node that contains the data relevant to the query.
Query Service
• A Stateless service that acts as a query broker
• Uses Index Service to identify home node for relevant query, and uses Anchor Point to fetch data from home node.
• Supports aggregated query, no raw data (e.g., RSSI, accelerometer reading) is provided.
13
14. Prototype Personal IoT Analytics Application
“Quantify the Spatio-Temporal Usage of Personal Object”
IoT Analytics
G
14
15. One Touch Tagging
Registration steps include
• Attach the tag
• Search and discover the new
tag via the application
• Provide a friendly name
15
20. Energy Assessment of Object Tags
• Using probes at MAC layer as data transport gives gains in the energy footprint compare to application
level processing.
Current(mA)
0.01
0.02
0.03
0.04
0.05
0.06
0.06
0.07
0.08
0.09
Consecutive Samples (100ms intervals)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Probe Request
Current(mA)
68.0
68.2
68.4
68.6
68.8
69.0
Configuration
Baseline Probe Application
68.72
68.53
68.10
EstimatedBatteryLife(Days)
0
100
200
300
400
Battery Capacity
100mAh 500mAh 850mAh
349.71
205.71
41.14
30%
• Essentially there is one significant peak for each outgoing packet, which is significantly smaller in comparison to
application level processing where multiple connections are maintained.
• A tag can achieve almost 1 year battery life based on 850 mAh battery with a premise that a household object
reports on an average 10 updates (state-of-use and location changes) with probes per day. This is based on the
published deep sleep consumption of 78uA, a transmit consumption of 70mA, a 80% efficiency of the regulator.
20
21. Zone A Zone B Zone C Zone D Zone E
Zone A 0.781 0.216 0.002 0.000 0.001
Zone B 0.253 0.719 0.016 0.002 0.009
Zone C 0.010 0.046 0.895 0.040 0.009
Zone D 0.001 0.003 0.015 0.914 0.068
Zone E 0.002 0.009 0.006 0.019 0.964
(a)
Zone A Zone B Zone C Zone D Zone E
Zone A 0.737 0.258 0.002 0.000 0.002
Zone B 0.213 0.736 0.030 0.005 0.017
Zone C 0.016 0.052 0.855 0.065 0.013
Zone D 0.002 0.002 0.016 0.955 0.026
Zone E 0.006 0.013 0.012 0.046 0.923
(b)
Confusion Matrix based on 140K observations
Proximity Ranging Assessment
• We have used One-vs-One SVM Classifier using a universal Gaussian kernel.
• 5 Non overlapping zones with 3 mobile object tags and 5 static objects
tags and one home node is used in the experimental setting.
• One tag is used to train the model and the other two to test the model.
• The algorithm achieved classification accuracy.84%
21
23. Summary
1
2
3
We are the first to show the feasibility of developing a WiFi-only solution for connected object analytics
which radically minimises deployment and management cost.
We present the design, implementation and evaluation of an end-to-end personal object analytics
platform build around only location and motion signatures of connected objects. We show that these
two pieces of information are enough to develop useful IoT applications by illustrating a personal object
analytics application.
We demonstrate how the design decisions and corresponding algorithms used in this work address
critical technical challenges with respect to energy, inference accuracy and computational overhead.
23
25. Gerd Kortuem and Fahim Kawsar "Market-based User Innovation
for the Internet of Things"; Internet of Things 2010 Conference
Afra Mashhadi, Fahim Kawsar, and Utku Acer
“Human Data Interaction in IoT: The Ownership Aspect;". The
IEEE World Forum on Internet of Things 2014
You Own Your Data, You Sell Your Data
• Our cloudlet based design scheme coupled with WiFi management frame based data transport
offer implicit privacy and data protection as Data remains in the home gateway and this provides
users with the control of their own data to do whatever they want to do with them – delete, sell
or share.
• An advantage of these design schemes is that, it opens up opportunity for wilful monetisation of
personal data
25