1. Idea: A System for Efficient Failure Management in Smart IoT Environments
Palanivel Kodeswaran, Ravi Kokku, Sayandeep Sen, Mudhakar Srivatsa
MobiSys 2016
Presented by: Zaeem Hussain
CS 3720
Feb 20, 2018

2. Basic Idea
Sensors fail over time, needing replacement or some other form of failure management
Cost of failure management increases with the scale of deployment
Inherent sensor redundancy in detecting Activities of Daily Living (ADLs)
Leverage this redundancy to increase the time between repairs

3. Setup

4. Motivation
ADLs generally trigger multiple sensors e.g. Cooking {Fridge, Microwave, Cupboard, Stove}
Not all sensors may be necessary to detect an activity
Failure of a sensor may or may not significantly degrade ADL detection
System can monitor sensors and delay failure management measures if detection performance does not fall below acceptable level

5. Architecture of Idea Online Offline ADL Detection Application
ADL Signature Generation
Sensor Failure Detection
ADL Rules
Maintenance Scheduling
Sensor Impact Estimation

6. Datasets Name # User Duration (days) # ADL (# Activity) # Sensor
(# Event)
KasterenA 1 25
16 (283)
14 (2006)
KasterenB 15
25 (172)
27 (22595)
KasterenC 19
27 (254)
23 (39861)
Aruba
219
11 (6477)
39 (805268)
Twor9-10 2
249
25 (3745)
100 (711421)
Twor2009 57
14 (499)
100 (136504)
TworSmr 63
8 (1016)
100 (366075)
AdlNorm 24 84
5 (120)
39 (6425)

7. Signature Extraction
Identify frequently occurring subsets of sensors for each activity
Extract rules of the form: (Fridge,Groceries_Cbrd,Plates_Cbrd) ⇒ Prepare_Breakfast <support:0.05,confidence:0.7>
Depending on the number of sensors triggered, an ADL may have multiple rules
Augment the rules with average Time of Day and Duration

8. ADL Detection Assume sensor trigger sequence partitioned by activity
Matching rule: Set of sensors in the rule is a subset of sensors in the activity
Sensor score of ADLk: Sum (confidence value of all matching rules of ADLk) x Prob(ADLk)
Temporal matching: Activity’s temporal feature lies within 1.5 standard deviations of the rule’s temporal feature
Temporal score: Sum (confidence value of temporally matching rules)
Final score: Weighted sum of temporal and sensor scores

10. Failure Detection
For periodic sensors, generate alert if time elapsed since last report from the sensor exceeds some predetermined threshold
For 100% impact sensors, if time elapsed since last detection of ADL for which the sensor is critical exceeds the threshold, generate alert
For rest of the sensors, keep track of rarity score ρ
Rarity score: Probability that observed sequence of activities happens without triggering the sensor in consideration
Once rarity score for a sensor falls below some threshold, generate alert
Choice of threshold for rarity score represents tradeoff between false alerts and detection latency

12. Maintenance Scheduling
Cloud collects following information for each home
Impact of failed sensor
Location of home
Degradation in ADL detection accuracy
Optimize for labor cost and minimum distance travelled by mechanic(s)

13. Experiments Split data into 80% train and 20% test sets
Longer duration traces generated by replaying the original sequence multiple times
Failures emulated by removing the failed sensor’s trigger events from the dataset, starting from the time of failure
Sensor failures follow Weibull distribution with MTBF of 1 year

20. Questions?