Current Sensor based Non-intrusive Appliance Recognition

Current Sensor based Non-intrusive Appliance Recognition
The 23rd International Technical Conference on Circuits/Systems,
Computers and Communications (ITC-CSCC 2008)
Current Sensor based Non-intrusive Appliance Recognition for Intelligent Outlet
Takeshi Saitoh1 , Yuuki Aota1 , Tomoyuki Osaki1 , Ryosuke Konishi1 and Kazunori Sugahara1
1
Department of Information and Electronics, Tottori University
4-101 Koyama-minami, Tottori 680-8552, Japan
E-mail : 1 {saitoh, konishi}@ele.tottori-u.ac.jp
Intelligent Outlet
Abstract: This paper presents the current sensor based nonintrusive appliance recognition method for intelligent outlet.
Our system has two main functions; one is the remote control
function of power supply through the Internet. The other is
monitoring function observe the state of appliance. In this paper, the monitor function is especially focused. To recognize
the state of appliance, we extract nine features based on measured current signal. In the experiment, we gathered a number
of signals with various appliances, and found that three features Ipeak , Iavg , and Irms yield valid recognition results of
81.3%, 84.0%, and 87.4% for classifying the state of appliance into three categories.
PC
Internet
PIC
AC100V
relay circuit
current sensor
outlet
appliance
plug
1. Introduction
The number of products with communication capability will
be available at our home by the development of IT technology
in not only a computer or a cellular phone but home appliance.
By the appliances’ being connected with the home network
becomes possible, we can control the appliance from remote
place. There are ECHONET and OpenPLANET on the home
network intended for the white goods [1], [2]. These use Ethernet, IEEE1349, and wireless LAN etc. as a communication
device. Therefore, it is necessary to equip these with the communication device to control an existing product that doesn’t
have the communication capability.
Our aim is to develop the intelligent outlet which has the
remote monitoring function of appliance to analysis state and
to reduce the standby electricity. The system that alerts to
unplug the appliance by gathering and analyzing of electric
power information was proposed. Yoshimoto et al. proposed
the non-intrusive load monitoring system by Neural Networks
[3]. Murata et al. proposed the method for determining the
on-off operation of appliance with support vector machine
[4], [5]. Moreover, Ito et al. proposed a system of appliance
detection and control using power consumption measurement
[6]. Nakamura et al. also proposed load monitoring system
of electric appliances based on Hidden Markov Model [7].
However, the recognition method of individual state of appliance is not treated to the control or monitoring. Then, this
paper proposes the method for recognizing the state of the
appliance and the appliance for a number of appliances.
2. System configuration
The system configuration of developed intelligent outlet is
shown in figure 1. In this figure, the red line means the voltage
flow, and the blue line means the signal flow. This system is
composed of a Peripheral Interface Controller (PIC) which is
a kind of micro controller, a relay circuit to control on-off of
power supply, and a current sensor that measures the amount
349
Figure 1. System configuration.
140mm
relay circuit
110mm
current
sensor
AC 100V
appliance
DC 5V
outlet
L1 modular
connector
RJ-45
appliance
PIC
Figure 2. System overview.
of the current that flows to the appliance. The state of the
system can be obtained through the Internet.
The overview of developed prototype system is shown in
figure 2. The dimension of prototype system is 110(W) ×
140(H) × 46(D) mm. Though, this system is used as the posterior device, the type of the embedded in wall is also in our
target. Our system is consisted of two outlets, two current
sensors, two relay circuits and one PIC. We use CTL-6-V-Z
as current sensor made by U.R.D. co., Ltd, and PIC18F67J60
made by Microchip Technology Inc.
The current sensor transduces the current signal that flows
to connected appliance into the voltage signal. This signal is
transduced the positive value through the Op Amp and noninverting amplifier, and input to the A/D converter of PIC.
Here, we set the resolution of A/D converted to 10 bit. Then,
we can measure the voltage signal of 3.3V in the maximum
voltage and the resolution 3.2mV. Moreover, we set the sampling frequency to 4.4 kHz. The number of samples that can
be measured at a time is 80 samples, that is, 1.33 seconds according to the specification of our system. The commercial
original signal
x(i)
phase shifting
y(i)
feature extraction
recognition
Figure 5. Flowchart of recognition algorithm.
Figure 3. Browser window.
0.30
0.30
11_01_01
66_02_01
0.25
0.25
66_02_02
11_01_02
11_01_03
66_02_03
0.30
Voltage [V]
66_02_01
0.25
11_01_02
0.25
66_02_02
11_01_03
66_02_03
0.20
Voltage [V]
Voltage [V]
0.20
0.15
0.10
0.20
0.20
11_01_01
0.15
Voltage [V]
0.30
0.15
0.10
0
0
0
10
20
30
0.05
0.05
0
10
20
30
40
sample
(a) fan
50
60
70
80
40
sample
50
60
70
0
80
0
10
20
30
40
sample
50
60
70
10
20
30
40
sample
50
60
70
80
(b) LCD display
(a) fan
0
0
0.10
0.05
0.05
0.10
0.15
80
Figure 6. Signals after phasing.
(b) LCD display
Figure 4. Signals before phasing.
3. 1 Phase shifting
Even if it is the same appliance and the same state, the phase
of the measured signal might be different as shown in figure
4. It is preferable to obtain coherent signal. Then, we apply
the phase shifting process. Here we call the measured signal
on original signal x(i), and applied phasing signal that took
out only one cycle on target signal y(i). The target signal is
shifted based on the peek value in original signal. Figure 6
shows the target signal form that applies the phasing process
to figure 4.
frequency in west Japan is 60 Hz. Then, the data at almost
one cycle can be observed by measuring one time.
Our system has two main functions; one is the remote control function of power supply through the Internet. The other
is monitoring function observe the state of appliance. In this
paper, the monitor function is especially focused.
Figure 3 shows the browser window of our system. In the
upper part of this window, there is an ON/OFF switch of the
relay. In this figure, one port is ON and the signal of this port
is displayed in the lower part of this window. The other is
OFF and its signal is not displayed. If the user turns on the
switch of port 2, the signal port 2 is displayed.
Figure 4 shows the result of the measurement to a fan and
a LCD display each three times. The horizontal axis indicates
the number of samples and the vertical axis indicates the voltage value. This paper describes the method for recognizing
the category and the state of the appliance from these signals.
3. 2 Features
This paper applies recognition process to recognize the category of appliance and its state, by using the following nine
features calculated from y(i).
There are a peak value Ipeak , an average value Iavg , and
a root mean square value Irms as a typical features of the
current. These values are calculated as follow equations.
Ipeak = max y(i)
i∈N
(1)
N
3. Recognition algorithm
The flowchart of our algorithm is shown in figure 5. We first
apply the phase shifting to obtain coherent signal. Next, we
extract nine features. Then, we apply the normalization process, and the nearest neighbor method to recognize the state
of appliance. The details of each process are described in the
following sections.
1 y(i)
N i=1
N
1 =
y(i)2
N i=1
Iavg =
(2)
Irms
(3)
Where, N is a number of samples of one cycle. Moreover,
there are a crest factor CF , a form factor F F in which a time
350
concentration of the signal, and a peak to average ratio Fpta .
These values are calculated as follow equations.
Ipeak
Irms
Irms
FF =
Iavg
CF
Ipeak
Fpta =
=
FF
Iavg
CF =
Table 1. Target appliances.
number of
appliance
product state #1 state #2
desktop PC
3
2
1-2
laptop
3
2
1-2
CRT display
3
2
1-2
LCD display
3
3
2
laser printer
3
3
2-3
ink-jet printer
3
3
2-3
television
3
2-2
1-3
portable television
1
2
2
projector
3
2-3
2-3
projector
1
2
2
digital video camera
3
4
3
video cartridge recorder
3
4
4
radio cassette recorder
3
3-4
2-3
refrigerator
3
2
2
rice cooker
3
3
2
pot
3
2
2
pot
1
2
1
coffee maker
3
2
1
cooking stove
1
4
3
washing machine
3
3
2
portable washing machine
1
2
1
cleaner
3
2
1
dehumidifier
3
2-3
1-2
hair drier
3
3
2
shaver
3
2
2
warm toilet seat
3
2
1-2
heaters
carbon heater
1
4
3
1
2
1
ceramic heater
1
3
2
stove
electric blanket
1
2
1
electric carpet
3
2
1
electric fan
3
4
3
portable electric fan
1
3
2
kotatsu
3
2
1
telephone
3
2
2
battery charger of cellular phone
3
2
1
game console
3
2-3
1-2
desk lamp
3
2
1
total
94
243
188
(4)
(5)
(6)
In this research, we define three another features, the ratio of
low-level at one cycle to peak value rl , the ratio of high-level
at one cycle to peak value rh , and the gradient angle θu [deg]
of rising edge.
3.3 Recognition method
The recognition process is applied the well-known method,
the Nearest Neighbor method after applying normalization.
For each feature, the observed values are normalized using
the average μ and standard deviation σ.
4. Experiment
In this experiment, each signal of 20 samples of 243 all states
was measured from 94 appliances used the office and household as the personal computer, the television, the refrigerator,
the fan, as shown in table 1. For instance, if the fan has two
rotational speeds (strong mode and weak mode), we can measure three states; off state, strong mode, and weak mode.
Here, we excluded 55 states that the current hardly flows
while turned off. Figure 7 and figure 8 show two states, off
state and on state, of the washing machine and cleaner, respectively. Observing figure 7(a) and figure 8(a), both signal
values are almost zero, and it is difficult to discriminate each
other. However, observing figure 7(b) and figure 8(b), these
are on state and it is easy to discriminate each other.
We classified into three kinds that showed 188 measured
states in the following. In table 1, the number of state #1
denotes 243 all states included 55 states, and the number of
state #2 denotes 188 states.
1. 188 categories considered all states of each appliance to
be another category.
2. 94 categories considered to be one category without distinguishing state of each appliance.
3. 35 categories in which appliance was classified by kind
We applied the leave-one-out method to obtain high recognition accuracy with a few data. Namely, out of 20 samples
for each category, 19 samples are a training set and the remaining one sample is a recognition set. By varying one sample, the total number of recognition trials is 20 for each category. The resulting recognition rate with nine all features was
76.3%, 80.5%, 85.5%, into 188 categories, 94 categories, and
35 categories, respectively.
The next experiments were to determine which features
among nine features are more effective for recognition. First
only one feature was used as the recognition feature. The feature yielding the highest recognition rate was identified and
then a second feature with the first feature yielding the highest recognition rate was determined. This process was carried
out for all the nine features. The most to least effective features are Irms , Iavg , and Ipeak in this order. Figure 9 shows
the result, where three curves (pattern 1, pattern 2, and pattern
3) are shown. As the result, we obtained the highest recognition rate of 81.3%, 84.0%, and 87.4% with three features, into
188 categories, 94 categories, and 35 categories, respectively.
Here, 243 states including off state were classified into
three patterns similar to the above mentioned. As the result,
each highest recognition rate was 64.1%, and was 66.1%, and
69.1%, respectively. Since the current hardly flows in off, the
discrimination with other states is difficult. Then, the recognition rate was decreased.
351
0.06
1.8
40_00_01
0.05
40_01_01
1.5
40_00_02
40_01_02
40_00_03
40_01_03
1.2
Voltage [V]
Voltage [V]
0.04
0.03
0.02
0.01
[2]
0.9
0.6
[3]
0.3
0
0
0
10
20
30
40
sample
50
60
70
80
0
10
20
30
40
sample
50
60
70
80
(a) OFF
(b) ON
Figure 7. washing machine.
0.06
1.8
45_01_01
45_00_01
0.05
45_01_02
1.5
45_00_02
45_01_03
45_00_03
1.2
Voltage [V]
0.04
Voltage [V]
[4]
0.03
0.02
0.01
[5]
0.9
0.6
0.3
0
[6]
0
0
10
20
30
40
sample
50
60
70
80
0
10
20
(a) OFF
30
40
sample
50
60
70
80
(b) ON
Figure 8. cleaner.
[7]
100
recognition rate [%]
90
80
70
pattern 3 (35 categories)
pattern 2 (94 categories)
pattern 1 (188 categories)
60
50
40
30
20
10
0
1
2
3
4
5
6
7
8
9
Irms
Iavg
Ipeak
rl
FF
Fpta
θu
CF
rh
the number of feature
Figure 9. Recognition results.
5. Conclusion
This paper proposes current sensor based non-intrusive appliance recognition method for intelligent outlet. We gathered
a number of signals with various appliances, and found that
three features Ipeak , Iavg , and Irms yield valid recognition
results of 81.3%, 84.0%, and 87.4% for classifying the state
of appliance into three categories.
This paper describes the recognition method, then, the future work is implemented proposed method into the intelligent outlet, and work toward practical use of whole system.
References
[1] S. Yamada. The trend of household electric business:
network technology of hte household electric applicances
352
”ECHONET” and application. Trans. of the Japanese Society for Artificial Intelligence, 16(3):349–354, 2001. (in
Japanese).
Y. Nakanishi. The conception of the Open PLANET.
Trans. of the Japanese Society for Artificial Intelligence,
16(3):355–360, 2001. (in Japanese).
K. Yoshimoto, Y. Nakano, Y. Amano, and B. Kermanshahi. Neural networks applied to non-intrusive load
monitoring system. IEEJ, 122-C(8):1351–1358, 2002.
(in Japanese).
H. Murata, T. Onoda, K. Yoshimoto, Y. Nakano, and
S. Kondo. Non-intrusive electric appliances load monitoring system — experiment for real housefold —.
IEEJ Trans. on Electronics, Information and Systems,
124(9):1874–1880, 2004. (in Japanese).
H. Murata, T. Onoda, K. Yoshimoto, and Y. Nakano.
Comparison of machine learning techniques for estimating the power consumption of household electric appliances. IEEJ Trans. on Electronics, Information and Systems, 123(7):1350–1355, 2003. (in Japanese).
M. Ito, H. Ohmata, S. Inoue, H. Shigeno, K. Okada, and
Y. Matsushita. A method and system of appliance detection and control using power consumption measurement. Trans. of Information Processing Society of Japan,
44(1):95–104, 2003. (in Japanese).
H. Nakamura, K. Ito, and T. Suzuki. Load monitoring
system of electric appliances based on hidden morkov
model. IEEJ Trans. on Power and Energy, 126(12):1223–
1229, 2006. (in Japanese).
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement