International Journal of Mining, Metallurgy & Mechanical Engineering (IJMMME) Volume 2, Issue 2 (2014) ISSN 2320–4060 (Online)
Back Propagation Neural Network Approach for
Electricity Usage Meter Numeral Recognition
Andi Sudiarso, and Rierien J. Merischaputri

implement BP and to calculate the accuracy rate of the
algorithm for numeral recognition.
Abstract —Automatic meter reading has become a requirement
of technology improvement for reading electricity usage meter
automatically. Artificial intelligence method can be used to recognize
numerical digits on a postpaid kWh meter. The purpose of this paper
is to implement and to calculate the accuracy of artificial neural
network method using back propagation algorithm through numeral
recognition of electricity usage meter reading. The network was
trained to learn by adjusting the interconnection strengths on every
iteration. There are 33 samples that has been tested for this study.
The back propagation algorithm is able to perform with 100%
accuracy after some iteration processes. The results revealed that as
the number of variation of samples increased, more accuracy is
achieved by the trained network.
II. RESEARCH METHOD
A block diagram of the recognition system is shown in the
Figure 1.
Keywords—Artificial neural network, backpropagation,
electricity usage meter, image processing, numeral recognition.
I. INTRODUCTION
ECOGNITION of numeral images has been a popular
research area because of its various potential applications.
At this moment, the Indonesian electricity meter reading still
use a repetitive and time consuming process that using a
mobile phone to store the amount of electricity usage, a mobile
phone camera to capture images of kWh usage, and a set of
papers to record activities. Moreover, some human operators
are not actually present to collect data and sometimes record
wrong electricity usage deliberately [1]. This study presents an
artificial neural network (ANN) method to automate numeral
recognition of electricity meter reading using back propagation
algoritm.
Numeral recognition is defined as an electronic translation
process of images from handwritten, typewritten, or printed
digit into a format understood by user for the purpose of
editing, indexing/searching, and a reduction in storage size [2].
The related problems of numeral recognition are the existing
noisy inputs, image distortion, and differences between
typefaces, sizes, and fonts. For solving this problem, a back
propagation (BP) algorithm of ANN approach is used. The
very general nature of the back propagation training means
that a back propagation network (a multilayer, feedforward
network is trained by back propagation) can be used to solve
problems in many areas. The purpose of this study is to
R
Fig. 1 Block diagram of recognition system
In the first step of research, it is required to determine the
numeral samples as the inputs of network training. The
training process will be completed when the correct targets has
been achieved. For the best performance, an automated system
needs an excellent image preprocessing stages including
filterring and filling process. The preprocessing stage is one of
the key successes of recognition system. Some of the steps
follow a reference [3].
A. Numeral Samples
An observation has been made for 33 numeral type of
Indonesian kWh meter at Yogyakarta district. The Indonesian
kWh meter mostly used typewritten arial font. Numeral
samples used are shown in Figure 2.
Fig. 2 Numeral samples from 0 to 9
A. Sudiarso is with the Department of Mechanical and Industrial
Engineering, Universitas Gadjah Mada, Jl. Grafika 2 Yogyakarta,
INDONESIA (phone:+62274521673; e-mail: a.sudiarso@ugm.ac.id).
R.J. Merischaputri is currently a graduate student at the Department of
Mechanical and Industrial Engineering, Universitas Gadjah Mada, Jl. Grafika
2 Yogyakarta, INDONESIA (e-mail: jessyntha.putri@gmail.com).
90
International Journal of Mining, Metallurgy & Mechanical Engineering (IJMMME) Volume 2, Issue 2 (2014) ISSN 2320–4060 (Online)
B. Back Propagation (BP)
TABLE I
PARAMETER CONDITIONS USED IN THE EXPERIMENT
Back Propagation Algorithm
Fig. 3 BP architecture (www.mathworks.com)
Backpropagation (BP) is a supervised training algorithm
which looks for the minimum of the error function in weight
space using the method of gradient descent. BP is a method
for propagating information about errors at the output units
back to the hidden units and input units. The architecture of
BP is ilustrated in the Figure 3. Although a single hidden layer
is sufficient to solve any function approximation problem,
some problems may be easier to solve using a network with
two hidden layers or more. In such a network, the first hidden
layers often serves to partition the input space into regions and
the units in the second hidden layer represent a cluster of
points [3]. The output units and the hidden units also may have
biases. This bias terms act like weights on connections from
units whose output is always 1. Network is initialized with
randomly chosen weights and will be corrected by gradient of
the error function.
In the BP framework, each computation for each neuron
requires the derivative of activation function. One cycle
through the complete training set forms one epoch. The
number of epoch elapsed while reaching the desired MSE
convergence is counted. The BP training process ended when
all the target performance have achieved. Strength of BP
algorithm stands on its feed-forward and back propagation
error evaluation which give more accurate outputs [4]. A
neural network is well suited for such application because of
its ability to learn adaptively.
The usual motivation to apply BP algorithm is to achieve a
balance between correct responses to new input patterns, it is
not beneficial to continue training until the total target of error
achieved. The concept is the error computed by the trainingtesting method. As long as the error for the training-testing
patterns decreases, training continues. When the error begins
to increase, the network is starting to memorize the training
patterns too spesifically (and starting to lose its ability to
generalize). Setting of optimum internal parameter used in the
experiment is shown in Table I. Some of the settings follow
previous experiments [6].
No
Setting Variable
Amount / Type
1
2
Number of Nodes in Hidden Layer
Activation Function in the Input Layer
5
Log Sigmoid
3
Activation Function in the Output Layer
Log Sigmoid
4
Training
Gradient Descent
5
6
7
Number of Epochs
Number of Goal
Coefficient Momentum
50000
1e-5
0.95
8
Performance
MSE
III. RESULTS AND ANALYSIS
A series of test has been performed on the system to
recognize the numeral digits. The performance and reliability
of the system is based on the error rate. Experiments carried
with a set of inputs of size 10 x 300, where 10 is the number of
classes and 300 is the feature vector length. The network is
considered to have learned a pattern if all computed output
values are within a specified tolerance of the desired values (0
or 1). For the result, the response of a unit was considered
correct if its activation was no greater than the tolerance and
the pixel (in the training pattern) corresponding to that unit
was ‘off’. Sample size of 33 are taken for testing purposes.
Figure 4 shows the example of ANN numeral recognition.
Fig. 4 BP numeral recognition
91
International Journal of Mining, Metallurgy & Mechanical Engineering (IJMMME) Volume 2, Issue 2 (2014) ISSN 2320–4060 (Online)
From 33 set experiments conducted, the BP algorithm had
reached 100 % accuracy rate. Proposed system recognizes
100% samples with average number of iterations is 3 times.
The repetitive process of recognition caused by bad images
which consist of glare and incomplete segmentation. Other
character failures are caused by oblique, small, and shift
characters. The example of image recognition and iteration
shown in Table III.
Fig. 4 BP numeral recognition (continued)
TABLE III
EXAMPLE OF NUMERAL RECOGNITION AND ITERATION
To determine the level of accuracy, set and individual
calculation are conducted between the characters. For
individual evaluation of system performance is shown in Table
II. For testing, there were 33 samples used.
No.
Image
Results
Data
Segmentation
1
TABLE II
ACCURACY RATE NUMERAL RECOGNITION
Accuracy
Electricy
(%)
ID Barcode
Usage
Read As
100%
521080026578
2107
2107
100%
521080115774
2730
2730
449280
Iteration
2
0
9260
2
521080120549
2880
2880
100%
5
521080018127
2904
2904
100%
4
2
521080052302
3175
3175
100%
521080082602
3952
3952
100%
3
521080127815
4076
4076
100%
1
521080084408
4117
4117
100%
2
3
521080007253
4375
4375
100%
521080118970
4558
4558
100%
3
...
11
...
...
...
9280
521080127417
4606
4606
100%
1
5005
5005
100%
IV. CONCLUSION
521080077527
3
521080022200
6254
6254
100%
0
521080022587
6853
6853
100%
0
4
The proposed method is mainly designed for real-time
electricity usage reading on postpaid kWh meter. From the
automated system built, the system could recognize the
electricity usage correctly, with 100% accuracy. The repetitive
iteration may be caused by bad images which consist of glare,
oblique, small, shift characters and inapropriate shooting
distance. Further research will focus on searching algorithm
for integrated automated numeral recognition and billing
calculation of electricity usage.
521080061884
8880
8880
100%
521080047200
8972
8972
100%
3
2
521080106901
9769
9769
100%
521080015959
10843
10843
100%
3
1
521080125055
12273
12273
100%
521080146119
3952
3952
100%
0
521072425567
1000
1000
100%
2
REFERENCES
2
[1] M. Setiawan, Teknik Pengendalian Kualitas Pembacaan Angka kWh
Meter Menggunakan I-MR (Individual-Moving Range) Control Chart,
Telaahan Staf, PT.PLN (Persero), Metro Lampung, 2010
[2] S. Asthana, F. Haneef, and R.K. Bhujade, “Handwritten Multiscript
Numeral Recognition using Artificial Neural Networks”, International
Journal of Soft Computing and Engineering (IJSCE), Vol 1, March
2011, pp. 1-5
[3] C.J. Lakshmi, A.J., Rani, K.S. Ramakrishna, and M. Kantikiran, “A
Novel Approach for Indian License Plate Recognition System”,
International Journal of Advanced Engineering Sciences and
Technologies, Vol 6 (1), 2011, pp. 10-14.
[4] R. Rojas, Neural Networks, Springer-Verlag, Berlin, 1996.
[5] J.J. Siang, Jaringan Syaraf Tiruan dan Pemrogramannya
Menggunakan MATLAB, Penerbit Andi, 2009.
[6] A.F. Cahyadi, Otomasi Sistem Pemindai Meteran Air PDAM untuk
Meningkatkan Akurasi Pembacaan Tagihan Air dengan Jaringan
Syaraf Tiruan Bertipe Backpropagation, Bachelor Thesis, Universitas
Gadjah Mada, 2013.
521061513866
2604
2604
100%
521080051893
2991
2991
100%
0
2
521080146119
3952
3952
100%
521070089179
4731
4731
100%
2
1
1
521072432195
4828
4828
100%
521061376257
9055
9055
100%
100%
521061376258
9087
9087
3
521061376272
9179
9179
100%
521061376259
13310
13310
100%
10
521061376268
13110
13110
100%
3
521061376271
10733
10733
100%
6
9087
100%
3
521061376258
9087
1
92