Professional Documents
Culture Documents
Keywords: Power consumption measurement, Electrome- The distributor’s critics sector uses an acquired image as
chanical meters, Digit recognition, HoG, SVM one of the ways to analyze inconsistencies in reading. This is
done in order to avoid errors in the charged consumption, theft
Abstract or fraud. That is, it seeks to avoid the non-technical losses that
according to [1] in Brazil represent a total of 44% of the total
The reading of the electrical energy consumption is carried out annual loss that is around 52 Terawatts / hour (TWh). On a
manually for the most part. One of the ways to validate the daily basis, a mean volume of 30,000 images is captured and
measurement performed by the meter reader on the field is sent for analysis before generating an invoice for the customer.
through the request of images acquisition in situations where Due to the large number of images to be evaluated, the audit
there is a divergence in the consumption informed by the meter is done by sampling. Yet the verification routine of the invoices
reader. However, this validation process in two Brazilian en- that were not printed on the field is done in its entirety. In cases
ergy companies (CEMAR and CELPA) entails a daily volume where inconsistencies with the images have been found, the
of roughly 30 thousand images to be analyzed. In this context, distributor needs to send another meter reader to the field to
this article presents a method for automatic consumption read- redo the reading or investigate the situation.
ing on electromechanical meters based on meter image using
Some works have been proposed aiming achievement of
HoG and SVM. The method obtained an accuracy of 79.52%
automatic consumption reading from electrical energy con-
in tests performed with more than 8000 digits.
sumer units based on images. In general, these works carry
out the acquisition of the image by fixing a camera positioned
1 Introduction in front of the meter, focusing on the region of the equipment’s
An electric meter or energy meter is a device or equipment digits and then using image processing techniques and machine
capable of measuring the electrical power consumption of a learning to recognize the digits from the reading region [2, 3].
residence, company or electrical appliance. The electricity There are few published works with a similar goal and present-
distribution companies perform a monthly reading of the dig- ing consistent results applied in a base with a sufficient amount
its which represent the consumption of the consumer unit in of images..
kilowatt-hours (kWh). Most of the time, the reading process Another approach for automatic consumption reading is
is performed manually by meter readers, which can result in proposed based on the use of smart meters. Smart meters are
reading errors and generate inconsistencies. devices that perform the reading and send it automatically via
The energy company of Maranho (CEMAR) and power data network (SMS, GSM, PLC or wireless) [4, 5, 6].
plants from Para S.A (CELPA) use a mobile device and a In this context, this article describes an automatic method
portable printer to respectively collect the consumption and to perform the reading of electric energy consumption in elec-
print out the costumer’s invoice. The meter readers typically tromechanical meters based on image. The result of this work
record on a mobile device application the consumption dis- is part of a Research and Development (R&D) project, hired by
played on the meter. If there is no inconsistency, the invoice CEMAR / CELPA (ANEEL PD-00371-0029 / 2016), executed
is generated, printed and then delivered to the customer. How- by the Applied Computing Center (NCA) from the Federal
ever, in some cases, the informed consumption is outside the University of Maranho (UFMA). The purpose of this project
consumer’s consumption average or the meter reader can not is to develop a Validation System (SIVAL) that is capable of
read. In these cases, the application requests the image acqui- automatically reading and processing images captured on the
sition from the meter and captures the geographical location of field, detecting inconsistencies (absence of the meter, incorrect
the reader as a record of the informed reading. The procedure meter identification, impossibility of reading the consumption
may contain failures, such as: image acquired without a me- and incorrect consumption) at the same time it performs auto-
ter, or present image of meter different from the corresponding matic recognition of consumption.
meter, etc. SIVAL has the objective of automatically processing and
Some information such as the number of digits (n), and mean 2.5 Digit Recognition
horizontal spacing between digits (dw), which were previously
defined. In order to reduce the computational complexity of the The segmented digits in the previous step are submitted to the
candidates search based on local similarity we used dynamic recognition step. In this step the Histogram of Oriented Gra-
programming [10]. The similarity are the information’s of the dients (HoG)[11] is used on the digits images for recognition.
spacing and alignment between digits. The method is robust The HoG could capture local shape feature while avoiding tex-
because it reduces the limitations such as the lighting variation ture noises. Initially it was found that there is no uniform size
from the external environment and digits dimension variation. for the extracted digits. To make the extraction of characteris-
The ET Energy from each simple blob combination is com- tics invariant to size, it was taken as a stepping stone that the
puted. Thus, the lowest energy combination (min ET), is cho- amount of cells, blocks and angles, HoG parameters , must be
sen for the digits segmentation. constant for all the extracted regions.
Thus each digit is centered on the smallest rectangular
bounding box having dimensions in power terms of 2, rang-
ET = Eh (bi , bi+1 )) + Ech (bi , bi+1 ) + EdistW (b1 , bi+1 ) + Eth (b1 , bi+1 ) to 512 pixels in length and height, which pre-
ing from 32
(1) sented maximum and minimum dimensions present on images
Knowing that: database. The empty spaces created are filled with zero (zero-
padding). In this way, there are no distortions in the extracted
Eh (bi , bi+1 ) = α1 |h(bi ) − h(bi+1 )| (2) regions and it is guaranteed that the central regions from the
bounding box contain the digit. Finally, the parameters used
for all digits, regardless of size, are 4 blocks, 8 cells per block
Ech (bi , bi+1 ) = α2 |Ch (bi ) − Ch (bi+1 )| (3) and overlap 1. Empirically, the number of directions used in
the histogram was 24 directions.
EdistW (bi , bi+1 ) = α3 |dw (bi ) − dw (bi+1 − dw )| (4) A total of 3,867 digit images were manually segmented and
used for training as described in Section 2.1. From the 3,867
Et h(bi , bi+1 ) = α4 |th(bi ) − th(bi+1 )| (5) images we have: 478 digit 0, 462 digit 1 , 454 digit 2, 354 digit
3, 372 digit 4, 362 digit 5, 350 digit 6, 336 digit 7, 356 digit
where (bi , bi+1 ) refers to adjacent blobs, Eh is the height dif- 8 e 342 digit 9. The characteristics extracted with the HoG on
ference energy, EC h is the energy of the difference from the the training images were submitted to the Support Vector Ma-
vertical distance between the centers of mass, EdistW is the chine Training (SVM) [12] using a radial kernel, configured
energy of the horizontal spacing variation and Et h is the en- with parameters C equal to 32 and γ equal to 0.0078125 esti-
ergy of the difference between the thresholds values used to mated with Grid search method using cross-validation with 10
obtain the binary blobs. Lastly, the bounding boxes from the folders. We choose SVM because literature prove that is a high
selected blobs are extracted and go to the recognition step. efficient method for pattern recognition in many different tasks.
Table 1. Digits recognition result (Test 1) Table 2. Digits recognition result (Test 2).
Fault by Fault by
digit Accuracy Amount digit Accuracy Amount
0 44.27% 835/1886 0 48.73% 843/1730
1 25.29% 477/1886 1 26.99% 467/1730
2 10.18% 192/1886 2 11.21% 194/1730
3 5.83% 110/1886 3 4.45% 77/1730
4 5.94% 112/1886 4 3.53% 61/1730
5 8.32% 157/1886 5 4.97% 86/1730
6 0.16% 3/1886 6 0.12% 2/1730
Success by Success by
digit 74.11% 6989/9431 digit 79.52% 6879/8651