1. An Identification Method of IR Signals
to Collect Control Logs of Home Appliances
〇Yuta Takahashi1，Teruhiro Mizumoto1
1. Nara Institute of Science and Technology
2017 ACIS Conference Series BCD
July 11, 2017

2. Background & Motivation
❖Control logs of home appliance
2
❖More intelligent smart home
Log
18:00
Cold
24℃
- ON/OFF
- Channel
- Volume
- Temperature …
Home which can understand user’s preference
- Automation
- Energy efficient
- Recommendation Smart home
Goal

3. Method for collecting control logs
❖Information appliance
3
❖Estimation by electric consumption
〇 Accurate logs
〇 Remote control
Products are not diverse
〇 Compatible with various products
Need for attachments (smart mater)
Difficult to estimate detail usage

4. IR signal & Problems
❖Collecting IR signals
4
❖Problems of identification
▪ Many protocols (NEC, AEHA…)
▪ Repeater functions
▪ Environmental noise
- Various appliances are controlled by IR
- Installing IR receiver to each room
Difficult to identifying

5. Proposed method
❖Process of IR signal
5
IR remote
controller
Preprocess
Comparison
IR Database
Identification of
appliance type
Identification of
command type
Unknown
signal
No
match
Command
type
Appliance
type
IR receiver
Identifying by
statistical model

6. Preprosess
❖Raw IR signal
▪ Consist of high/low pulses (PWM, Pulse Width Modulation)
▪ High memory-capacity for devices
▪ High computation for identifying
6
Raw IR
Pulse width
sequence
❖Pulse width sequence
▪ Consist of length of
high/low pulses
▪ Range is 0 to 255
▪ Easy to handle
▪ Low memory-capacity

7. Comparison method of two signals
7
Two signals
𝑆1
0
𝑆1
1
𝑆1
2
𝑆1
3
𝑆1
4
𝑆1
5
𝑆1
6
𝑆1
7
𝑆1
0
𝑆1
1
𝑆1
2
𝑆1
1
𝑆1
2
𝑆1
3
𝑆1
2
𝑆1
3
𝑆1
4
𝑆1
𝑆𝑠𝑢𝑏
𝑆2 𝑆2
0
𝑆2
1
𝑆2
2
𝑆2
0
𝑆2
1
𝑆2
2
𝑆2
0
𝑆2
1
𝑆2
2
𝑀𝐴𝐸0, 𝑆𝐴𝐸0 𝑀𝐴𝐸1, 𝑆𝐴𝐸1 𝑀𝐴𝐸2, 𝑆𝐴𝐸2
𝑝 = arg min(𝑀𝐴𝐸 𝑛) 𝑴𝑨𝑬 𝒑, 𝑺𝑨𝑬 𝒑
A captured signal
A referenced signal
𝑀𝑒𝑎𝑛 𝐴𝑏𝑢𝑠𝑜𝑙𝑢𝑡𝑒 𝐸𝑟𝑟𝑜𝑟 =
1
𝑁
෍
𝑖=0
𝑁
|𝑆𝑠𝑢𝑏
𝑖
− 𝑆2
𝑖
|
Sum 𝐴𝑏𝑢𝑠𝑜𝑙𝑢𝑡𝑒 𝐸𝑟𝑟𝑜𝑟 = σ𝑖=0
𝑁
|𝑆𝑠𝑢𝑏
𝑖
− 𝑆2
𝑖
|
(long)
(short)

8. Dataset
8
14 appliances
↓
140 commands
10 signals
↓
1,400 signals
irMagician
1400
2
= 979,300 combinations

9. Error frequency of same appliance and other appliance
9
Same appliance (any command) :
Other appliance (any command) :
A appliance
A1 command
A appliance
A2 command
A appliance
A1 command
B appliance
B1 command
Small
overlapped
Difficult to
fit a model
(over fitting)
Constructing a model of “same appliance” of MAE

10. Model for identifying appliance type
❖Fitting
▪ Inverse gaussian, Gamma, Inverse gamma,
Weibull, Chi and F distributions
▪ Maximum likelihood estimation
▪ AIC (Akaike's Information Criterion)
▪ Inverse gamma (k=3) and F (k=4) are best
fitting
10
❖Decision
▪ 95% confidence interval
▪ 𝑒 ≤ 𝑒𝑡ℎ : same appliance
▪ 𝑒 > 𝑒𝑡ℎ : other appliance
Bad fitting (Weibull) Inverse gamma
95% 5%
3.72
𝑒𝑡ℎ

11. Error frequency of same command and other command
11
Same command (same appliance) :
Other command (same appliance) :
A appliance
A1 command
A appliance
A1 command
A appliance
A1 command
A appliance
A2 command
Good shape
of histogram
Constructing a model of “same & other command” of SAE

12. Model for identifying command type
❖Fitting
▪ Inverse gaussian, Inverse
gamma and F are better
than other
▪ We chose Inverse gamma
as well as model of
appliance type
12
❖Decision
▪ Bayes’ decision
𝑙𝑜𝑔
𝑝 𝑦 = "same"|𝑥
𝑝 𝑦 = "other" 𝑥
▪ Positive : same command
▪ Negative : other command

13. Evaluations
1. Accuracy of identifying appliance type
▪ Verifying by 10-fold cross validation
2. Accuracy of identifying command type
▪ Verifying by 10-fold cross validation
3. Simple simulation
▪ Identification depends on signals in database
▪ Constructing database randomly
▪ Check how many signals are needed in database
13

14. Identification accuracy
14
❖Accuracy of appliance type (total support : 199,778)
❖Accuracy of command type (total support : 12,636)

15. Result of simple simulation
▪ Simulating 1,400 signals in each number of appliances
▪ Correct match rate is stable if 6 signals, or more, are
included in the database
15
Stable

16. Conclusions
❖Proposed method for identifying IR signal by statistical model
❖Identifying appliance accuracy is 95.5%
❖Identifying command accuracy is 92.0%
❖Identification stability is achieved when 6 signals, or more, of
each appliance are included in database
❖We plan to collect and identify the IR signals in real environment
16

17. Simple simulation
Process
1. Construct database from signals of each appliance
2. Identifying the test signals
3. Increment the number of signals in database
17
❖Matching method
▪ One appliance type most identified is selected
▪ No match : Several types are estimated or no types of identification
Signal:TV
Signal:TV
Signal:TV
Signal:Fan
→ TV
→ TV
→ Fan
Signals identified as same appliance
TV
Test
Compared signals
Labeled appliance &
command to signals
Database
Result