This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON SMART GRID 1

Smart Transformer for Smart Grid—Intelligent

Framework and Techniques for Power
Transformer Asset Management
Hui Ma, Member, IEEE, Tapan K. Saha, Senior Member, IEEE, Chandima Ekanayake, Member, IEEE,
and Daniel Martin, Member, IEEE

Abstract—Condition monitoring and diagnosis have become an adopt high performance liquid chromatography to measure
essential part of power transformer asset management. A variety 2-furfuraldehyde content for evaluating paper insulation aging
of online and offline measurements have been performed in util- condition in transformers [5], [6]. Polarization-based mea-
ities for evaluating different aspects of transformers’ conditions.
However, properly processing measurement data and explicitly surements including polarization and depolarization current
correlating these data to transformer condition is not a trivial measurement and frequency domain spectroscopy have been
task. This paper proposes an intelligent framework for condi- also used in utilities for estimating aging and water con-
tion monitoring and assessment of power transformer. Within tent in transformer oil-paper insulation [4], [7]. Frequency
this framework, various signal processing and pattern recogni- response analysis has been applied for winding deformation
tion techniques are applied for automatically denoising sensor
acquired signals, extracting representative characteristics from and displacement detection in transformers [8]. To provide
raw data, and identifying types of faults in transformers. This dynamic behavior of a transformer’s condition during its
paper provides case studies to demonstrate the effectiveness of operation, online sensor-based measurements have also been
the proposed framework and techniques for power transformer performed. Examples include partial discharge (PD) mea-
asset management. The hardware and software platform for surement for transformer insulation diagnosis and acoustic
implementing the proposed intelligent framework will also be
presented in this paper. measurement for transformer on-load tap-changer (OLTC)
fault identification [9], [10].
Index Terms—Asset management, denoising, dielectric Due to the complexity of a transformer’s structure and its
response, dissolved gas analysis (DGA), insulation, partial
discharge (PD), pattern recognition, power transformer. degradation mechanism, interpreting the measurement data
obtained from the above techniques and subsequently evaluat-
ing the transformer’s condition is a challenging task, especially
I. I NTRODUCTION for nonexperts. Therefore, there is a necessity to investigate
intelligent techniques and their applications for automatically
OWER TRANSFORMER is one of the most important
P assets in an electricity grid. During its operation, a power
transformer is subjected to thermal, mechanical, and electri-
analyzing measurement data and identifying fault types in
the transformer. Various investigations on applying intelligent
techniques to transformer monitoring and diagnosis have been
cal stresses, which can eventually deteriorate the transformer
reported in [11]–[15]. For example, numerous artificial neu-
and cause the failure of the whole unit. Moreover, electrical
ral networks (ANNs), neural-fuzzy systems, and evolutionary
utilities around the world are currently facing an increasingly
algorithms have been applied for diagnosing transformer incip-
aged population of power transformers in their grids. To ensure
ient faults using DGA measurement data [11]–[13]. Recently,
reliable operation of a transformer, its health condition must
intelligent algorithms have also been applied to compute health
be continuously monitored and evaluated for appropriate asset
index for a transformer to provide an indication of the sta-
management decisions [1]–[4].
tus of its insulation system [14]–[16]. Nevertheless, the above
A variety of measurements have been adopted in util-
techniques scattered over different aspects of interpreting mea-
ities for condition monitoring and assessment of power
surement data and making faults diagnosis; there is still lack
transformers [5]–[10]. With the oil samples collected from
of a common framework, within which different intelligent
transformers, utilities conduct dissolved gas analysis (DGA)
algorithms can be effectively deployed and executed to ful-
to identify thermal and discharge faults in transformers and
fill the tasks of performing sensor-based transformer condition
Manuscript received February 14, 2014; accepted December 16, 2014. monitoring and assessment.
This work was supported in part by the Australian Research Council, in part This paper presents an intelligent framework for transformer
by Powerlink Queensland, in part by Energex, in part by Ergon Energy, and condition monitoring and assessment. Within this framework,
in part by TransGrid. Paper no. TSG-00102-2014.
The authors are with the School of Information Technology and Electrical a number of intelligent algorithms have been applied includ-
Engineering, University of Queensland, St. Lucia, QLD 4072, Australia ing: 1) digital signal processing algorithms, which concern
(e-mail: [email protected]). with denoising and extracting signals acquired by sensors
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. during online sensor-based transformer condition monitor-
Digital Object Identifier 10.1109/TSG.2014.2384501 ing; and 2) pattern recognition algorithms, which focus on
1949-3053 c 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON SMART GRID

With the modern sensing and data acquisition technologies,

it is relatively easier for gathering, storing, and accessing raw
data in a more straightforward manner. However, the huge
volume of raw data may consist of erroneous records, redun-
dant information, and even incomplete records. Therefore, it is
necessary to perform feature extraction, which aims to select
representative characteristics from raw data. The selected fea-
tures are embedded with more condensed and smaller amount
information, which will be more suitable and manageable for
further information processing to support transformer condi-
tion assessment. Feature extraction forms the third component
of the proposed framework.
After feature extraction, pattern recognition algorithms are
adopted as the fourth component for classifying faults types
and evaluating health conditions of transformers. The merit of
Fig. 1. Intelligent framework for power transformer asset management. pattern recognition algorithms is the learning from data, which
can make use of a historic database to obtain the knowledge
regarding the correlation between the measurement data and
transformer fault types classification and health condition eval- transformers’ conditions. Subsequent condition assessment on
uation. To verify the applicability of these algorithms, case the transformer of interest can be performed based on such
studies and results of applying these algorithms to interpret knowledge.
online and offline measurements of transformers are pro- By combining the results of various measurements, the
vided in this paper. The case studies include: 1) denoising overall dynamic condition of a transformer can be evaluated.
signals obtained from online PD measurement; 2) recogniz- Within the proposed framework, a data and information fusion
ing types of discharge sources (insulation defects), which component will be implemented, which will integrate pattern
cause PDs in transformers; and 3) recognizing transformer recognition results of individual measurement to estimate the
faults by interpreting DGA measurement data. The hardware probability of different types of faults that might occur in a
and software platform for implementing the proposed intel- transformer.
ligent framework and deploying the above signal processing The following sections will detail the above key components
and pattern recognition algorithms will also be presented in of the proposed frameworks, i.e., signal processing algorithms,
this paper. feature extraction, and pattern recognition algorithms for trans-
former condition monitoring and assessment. Case studies and
results will also be provided in each section to demonstrate the
II. I NTELLIGENT F RAMEWORK FOR P OWER applicability of these algorithms.
T RANSFORMER A SSET M ANAGEMENT
Fig. 1 presents the proposed intelligent framework for
III. S IGNAL P ROCESSING FOR O NLINE S ENSOR -BASED
transformer asset management with the focus on transformer
T RANSFORMER C ONDITION M ONITORING
condition monitoring and diagnosis. The key components
included in the framework will be discussed in this section. This section presents signal processing algorithms for
The first component of the proposed framework is an inte- denoising sensor acquired signals (i.e., original signals), which
gration of various online sensor measurements to provide are prone to environmental interference and noise. These
dynamic information of a transformer’s condition. These mea- algorithms are applied for denoising PD signals, which are
surements may include current and voltage measurements, acquired from PD measurements performed on experimental
temperature measurement, water content in oil measurement, PD models and transformers.
PD detection, and acoustic/vibration monitoring. Necessary PD is the localized breakdown in insulation systems and
data acquisition hardware, which can acquire sensor mea- its measurement can provide a means of online monitoring
sured signals, and then transmit these signals through com- and diagnosis of a transformer’s insulation system [9], [17].
munication peripherals for further data processing will be Fig. 2 depicts two types of PD measurement set-ups: 1) RLC
incorporated. impedance in series with a coupling capacitor for acquiring
For practical transformer condition monitoring in a sub- PD signals, the PD signals are recorded by a commer-
station environment, various interference, and noise can be cial measurement system complying with IEC60270 [17]
imposed on the sensor acquired signals. The interference and and 2) high frequency current transformer (HFCT), which
noise can significantly affect the sensitivity and reliability of clamps on a grounding wire of the transformer for acquir-
sensor measurement and consequently has reverse impacts ing PD signals, the PD signals are recorded by using a digital
on correctly identifying the faults occurring in transformers. oscilloscope.
Therefore, the second component of the proposed framework During online PD measurement, different types of noise
is a suite of signal processing algorithms, which can remove may exist including white noise, periodic noise, and stochas-
noise from original measured signals. tic noise. A number of techniques for removing noise from
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

MA et al.: SMART TRANSFORMER FOR SMART GRID—INTELLIGENT FRAMEWORK AND TECHNIQUES 3

4) if N = 1, setting REC = IMF (N) and N = N + 1, and

going to Step 5;
5) if kurtosis of IMF, K (N) = K (N − 1)/2, then setting
REC = REC + IMF (N) and N = N + 1, going to
Step 5; otherwise, going to Step 6;
6) executing AMT on REC to remove noise;
7) reconstructing PD signals (denoised signals).
As shown above, the hybrid algorithm firstly decomposing
sensor acquired signals into IMFs. After signal decomposition,
kurtosis of each IMF is calculated as

 L
k= (Ii − I) 4
(L − 1)σ 4 . (1)
i=1
Fig. 2. Set-ups of PD measurements of power transformer. (a) Capacitive
PD measurement (IEC60270). (b) HFCT-based PD measurement.
In the above equation, the mean, length, and standard devia-
tion of IMF are denoted as I, L, and σ , respectively. The larger
value of kurtosis implies the signals consisting of abruptly
measured PD signals have been proposed in [9] and [18]–[22]. changed impulses while smaller value of kurtosis refers to
The following Sections III-A and III-B will review two novel the slowly fluctuated signals or signals with evenly distributed
denoising algorithms proposed by Chan et al. [21], [22]. amplitudes. As kurtosis is sensitive to transient signals, it is
Section III-C will provide results of denoising PD signals col- used to select IMFs for reconstructing PD signals with noise
lected from experimental PD models and transformers. Due removed.
to the space limitation, mathematic derivations of these algo- The kurtosis-based selection process starts from the first
rithms will be omitted and readers can refer to [21] and [22] IMF, which embeds the highest frequency component of the
for the details. measured signals. If an IMFs kurtosis value decreases to half
compared to the previous IMF, it implies that this IMF consists
of slow fluctuations due to periodic noise and/or white noise.
A. Hybrid Algorithm for Adaptive Signal Denoising
As a result, this IMF is considered as noise without containing
Discrete wavelet transform (DWT) has been widely adopted any PD signals and thus can be discarded in the reconstruc-
for PD signal denoising [20], [21]. Based on the convolution tion. Otherwise, it is regarded as PD signal and added to the
between original signals and a selected mother wavelet, DWT previous IMF for PD signals reconstruction. Such IMF selec-
decomposes the original signals into many coefficients resid- tion process will be continued on the remaining IMFs. With
ing at different frequency bands. One of major challenges the completion of the above selection process, most noise can
of DWT-based signal denoising is the proper selection of be removed and the PD signals are reconstructed.
mother wavelet. DWT can attain desirable denoising perfor- However, it is not uncommon that the reconstructed signals
mance given the mother wavelet has high correlation with pure may still consist of some noise, which has the same frequency
PD signals. However, for the signals with low signal-to-noise scales of the selected IMFs. To deal with this problem, AMT
ratios, a mother wavelet which correlates to the noise rather is proposed, which can create thresholds for removing noise
than the pure PD signals may be chosen. Also, for practical originated signals having similar frequency scales as pure PD
online measurement, a fixed mother wavelet may not be feasi- signals [21]. Therefore, the hybrid algorithm based on EEMD
ble as it cannot match different types of PD signals generated and morphological morphology is an automatic signal denois-
by different insulation defects [21]. ing method, which is suitable for online PD measurement of
To provide an adaptive and automatic PD denoising method, transformer.
Chan et al. [21] proposed a hybrid algorithm, which is
based on ensemble empirical mode decomposition (EEMD)
and morphological morphology. In this hybrid algorithm, B. Algorithm for Removing Stochastic Noise
the sensor measured signals are decomposed into a series The frequency characteristic of stochastic noise has con-
of mono-component signals, which are named as intrinsic siderable similarity with that of pure PD signal. As such,
mode functions (IMFs), for PD signals extraction from noise. stochastic noise is relatively more difficult to remove compared
Preselection of mother wavelet is not required. Moreover, to other types of noise such as periodical noise and Gaussian
an automatic morphological thresholding (AMT) is used to noise. Chan et al. [22] developed an algorithm based on fractal
create threshold for further separating noise from PD sig- dimension and entropy analyses for removing stochastic noise.
nals. The procedure of this hybrid algorithm is listed as This algorithm starts with transforming the original sensor
follows [21]: measured signals, i.e., stochastic noise corrupted PD signals,
1) obtaining sensor measured signals, which consists of PD into 2-D signals. Then, on the transformed signals, a fractal
signal, white noise, and periodic noise; dimension is computed for each power cycle. For a power
2) setting reconstructed signal, REC = 0, and IMF number, cycle with severe noise, the value of its fractal dimension
N = 1; will be significantly different from that of other power cycles.
3) performing EEMD on the measured signals; Thus, fractal dimension can be used to locate and accordingly
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON SMART GRID

Fig. 4. Stochastic noise removal results of PD signals collected from PD mod-

els of internal discharge. (a) Measured signals (noise corrupted PD signals) in
Fig. 3. Denoising results of PD experimental model of internal dis- time domain over ten power cycles. (b) Measured signals with respect to phase
charge. (a) Measured signals (noise corrupted PD signals) in time domain. angles of power cycle. (c) Denoised PD signals with respect to phase angles of
(b) Measured signals with respect to the phase angles of power cycle. power cycle. Brightly colored pulses are denoised PD signals whereas pulses
(c) Denoised PD signals in time domain. (d) Denoised PD signals with respect in the background are noise.
to the phase angles of power cycle.

remove the noise. However, in PD measurement some stochas- Fig. 4 presents stochastic noise removal results by using
tic noise may appear at most of power cycles and have similar fractal dimension and entropy analyses. In the figure, the PD
fractal dimension values compared to PD signals. For such sit- signals were also acquired by using PD measurement set-up
uation, it may not be able to completely identify and remove one (Fig. 2) with the experimental model of internal discharge
noise from PD signals by only using fractal dimension. over ten power cycles. It can be observed from the figure
For a PD signal it has high chance to appear at the same that PD signals locate periodically in most of power cycles.
phase location of different power cycles. In contrast, for Severe stochastic noise occurs at the first and second power
stochastic noise, if it appears in a particular phase location cycles where relatively insignificant noise can be found at
of a power cycle, it will have low chance to appear at that other power cycles.
phase location in the subsequent power cycles [22]. Therefore, Fig. 4(b) is the original signals with respect to phase angles
at each phase location, the entropy values of noise is sig- [all signals in ten power frequency cycles were mapped into
nificantly different from the entropy values of PD signals. 0◦ –360◦ phase angles, which is referred to as phase resolved
Based on this observation, entropy value is used with frac- PD (PRPD)] without applying the denoising algorithm. It indi-
tal dimension to distinguish and remove stochastic noise from cates that PRPD before signal denoising is contaminated by
PD signals. stochastic noise. Fig. 4(c) is the noise removed PRPD, which
proves that fractal dimension and entropy analyses are capable
of identifying and removing stochastic noise.
C. Case Studies of PD Signal Denoising Fig. 5 presents the results of using fractal dimension
This section presents the results of denoising measured and entropy analyses to denoising PD signals, which were
PD signals using the above signal processing algorithms. The collected from PD measurement performed on a 5 MVA trans-
PD signals were collected from experimental PD models and former (using PD measurement set-up one, Fig. 2). It can be
transformers [21]. observed from Fig. 5(a) and (b) that the measured PD signals
Fig. 3 presents the denoising results of measured PD signals are embedded with heavy noise, which causes considerable
by using hybrid algorithm based on EEMD and morphological difficulties for the identification of PD signals. However, as
morphology. The original signals were collected from PD mea- shown in Fig. 5(c), the proposed fractal dimension and entropy
surement on the experimental PD model of internal discharge analyses method can effectively reveal PD signals. In Fig. 5(c),
using PD measurement set-up one (Fig. 2). the PD signals appear at the negative power cycle (180◦ –360◦ )
It can be seen from Fig. 3 that the hybrid algorithm can may be generated by corona and/or surface discharge; the
effectively extract PD signals from noise and also preserve PD signals appear at both positive (0◦ –180◦ ) and negative
the phase locations of these extracted PD signals. This can (180◦ –360◦ ) power cycles may be due to internal discharge.
assist pattern recognition algorithm to recognize different types A pattern recognition algorithm can be trained to classify the
of PD sources, which caused by different types of insulation above types of PD pattern. This will be discussed in the next
defects in transformers [23]. section.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

MA et al.: SMART TRANSFORMER FOR SMART GRID—INTELLIGENT FRAMEWORK AND TECHNIQUES 5

Fig. 6. PCA visualizations of PD sources dataset constructed using EEMD

feature extraction method.

records (i.e., features in dataset Y) and the corresponding types

Fig. 5. Stochastic noise removal results of PD signals collected from a
of faults occurring in transformers. It involves exploring the
transformer. (a) Measured signals (noise corrupted PD signals) in time domain function of f : yj → Ck , j = 1, . . . , N, Ck ∈ {1, . . . , K}. The
over ten power cycles. (b) Measured signals with respect to the phase angles algorithm can be subsequently used to identify the fault types
of power cycle. (c) Denoised PD signals with respect to the phase angles of
power cycle. Brightly colored pulses are denoised PD signals whereas pulses
of transformers, which are not included in the above training
in the background are noise generated signals. dataset X or Y.

B. Classification of PD Sources in Transformer

Though the above signal processing algorithms were applied
to PD monitoring and diagnosis of transformer, they can also be PD source classification is to distinguish the types of defects
applied to other types of sensor-based online measurement for that cause discharge in transformer insulation system. The
transformer condition monitoring. An example is the acoustic conventional approach for PD source classification involves
measurement for transformer OLTC fault identification. several steps:
1) collecting PD signals from PD measurements on differ-
ent types of experimental PD models;
IV. F EATURE E XTRACTION AND PATTERN R ECOGNITION 2) constructing PRPD pattern, i.e., PD pulses distribution
FOR T RANSFORMER FAULT T YPES C LASSIFICATION of magnitude and number count with respect to phase
After signal denoising, the next step within the pro- angles of power cycle for all acquired PD signals;
posed intelligent framework for transformer asset management 3) using statistic operators to characterize the PD pat-
involves feature extraction and pattern recognition, which aim terns and forming a training database of different PD
to identify types of faults occurring in transformers. sources;
Starting with the mathematic formulation of feature extrac- 4) training pattern recognition algorithms, which are used
tion and pattern recognition, this section will present several to identify the PD sources in transformers [23].
techniques and their applications on PD sources classification Instead of using the above PRPD pattern, this paper adopts
and transformer faults diagnosis. EEMD for feature extraction because it can perform two tasks,
PD signal denoising and feature extraction, in one step. Firstly,
data are extracted from IMFs with nine-level decomposition.
A. Mathematic Formulation of Feature Extraction and To reduce the high dimensionality introduced by decomposi-
Pattern Recognition tion (a power cycle may have several hundred discharge pulses,
After signals acquisition and denoising, the resultant mea- and each discharge pulses is represented by nine coefficients),
surement records
 form a N  × D dataset of X = [x1 , . . . , xN ], the first four moment statistics including mean, variance, skew-
where xi = xi1 , . . . , xiD is a D dimensional vector with ness, and kurtosis are computed for the distribution formed by
each dimension is regarded as a feature. Feature extraction the nine coefficients of discharge pulses. As a result, 36 fea-
reduces the dimensionality
 of dataset X and transforms
 it to a tures (4 × 9 = 36) are used to represent each record of PD
N × d dataset of Y = y1 , . . . , yN , where yi = y1i , . . . , ydi measurement.
and d  D. Though substantial features are removed from A training database having 1000 records for five PD sources
the dataset X in feature extraction, the remaining features (internal discharge, surface discharge, discharge in oil, dis-
in dataset Y can still provide essential characteristics for charge in air, and discharge due to floating particles) has been
discriminating different types of faults [23]. constructed [23]. Each record consists of the above 36 features
Each record in dataset Y can be categorized into one of K and the corresponding types of PD sources. Therefore, the Y
independent classes, Ck ∈ {1, . . . , K} with each class corre- dataset for training is with the dimension of 1000 × 36. To
sponds to one type of transformer faults. A pattern recognition visualize data distribution in this dataset, principle component
algorithm can be evoked to learn the correlation between analysis (PCA) is adopted as shown in Fig. 6.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON SMART GRID

TABLE I
S AMPLES D ISTRIBUTION IN T RAINING DATABASE
(B EFORE AND A FTER SMOTE)

In utilities, transformers’ faulty rate is generally low. Thus,

the records belonging to some faults types can be insufficient,
which leads to an imbalanced DGA training database, i.e., the
number of records of some faults types (majority classes) are
quite larger than that of other faults types (minority classes).
An algorithm trained by such an unbalanced database will be
biased in that it can lead to higher misclassification rate on
the minority classes [25].
Fig. 7. Multiple PD sources obtained from internal discharge and discharge To tackle the above class imbalanced problem, this paper
due to floating particles. (a) PD pattern of multiple PD sources of internal applies synthetic over-sampling technique (SMOTE) to reduce
discharge and discharge due to floating particle. (b) PD pattern of internal
discharge. (c) PD pattern of discharge due to floating particles. the degree of imbalance among different classes (i.e., faults
types) and subsequently facilitate pattern recognition algo-
rithms consistently achieving desirable classification accu-
From Fig. 6, it can be seen that the records of differ- racy [26]. SMOTE is implemented in three steps: 1) computing
ent PD sources exhibit a certain degree of overlapping. This the imbalance ratio between majority and minority classes in
implies the necessities of adopting algorithms with nonlinear the original database, which is then used to decide the number
discriminability, i.e., ANNs or support vector machine (SVM) of synthesized samples; 2) computing the Euclidean distance
algorithm, for PD sources classification. The validation results between a sample and its closest neighbors in a minority class;
show that SVM (with EEMD feature extraction) can achieve and 3) generating the synthesized samples according to
97% classification accuracy on this training dataset (70%  
mi = x + rand (0, 1) × d x, x∗i . (2)
records in the dataset are used for training SVM while the
remaining 30% records are used for validating the trained In the above equation, mi denotes the synthesized sam-
SVM algorithm) [23]. ple, rand (0, 1) is referred to as a random
 number within the
Different types PD sources can be co-existence in a trans- range from zero to one, and d x, x∗i is the Euclidean distance
former. SVM algorithm can estimate the probability of exis- between sample x and its neighbor x∗i .
tence of each type of PD sources. Fig. 7 depicts PRPD pattern To further improve the performance of pattern recognition
of an experimental model of multiple PD sources, which is algorithms, a boosting approach is adopted. Boosting is a
due to simultaneous occurrence of internal discharge and dis- meta-learning technique, which constructs a high-performance
charge due to floating particles. A SVM algorithm trained by ensemble of algorithms to attain desirable performance [27].
the above dataset shown in Fig. 6 indicates the probability of Boosting is implemented as follows: 1) performing ini-
discharge due to floating discharge and internal discharge are tial training, in which all samples are with equal weights;
67% and 17%, respectively. 2) assigning higher weights to the samples, which are incor-
The above trained SVM is also applied to make PD sources rectly classified in the previous step; and 3) the training
classification for the PD measurement on a 5 MVA power iterations continue until the ensemble attains the desirable
transformer, which was shown in Fig. 5 in Section III. It indi- classification accuracy. A case study of applying SMOTE and
cates that the internal discharge accounts for 54%, surface boosting to transformer fault diagnosis is as follows.
discharge accounts for 34%, and corona accounts for 14% of A DGA database provided by a utility company is pre-
discharges in this transformer. sented in Table I. The first and second rows of the table are
records distribution before and after performing SMOTE oper-
ation, respectively. It is apparently the original database shown
C. Transformer Faults Diagnosis Using DGA Data in Table I is class imbalanced. After SMOTE process, the
DGA has been in use in utilities for diagnosing transformer database achieves approximate equal distribution of records
faults such as overheating, discharge fault, and PD. As men- throughout different faults types.
tioned in Section I, pattern recognition techniques have also The SVM and radial basis function network (RBF) algo-
been applied for DGA data interpretation and transformer rithms were trained using the above database. The inputs
faults diagnosis [11]–[13]. A pattern recognition algorithm is (features) to the algorithms are concentrations of dissolved
trained by a historic DGA database, which consists of DGA gases C2 H4 , C2 H2 , C2 H6 , CH4 , and H2 [15]. Descriptions
measurement records and the fault types of corresponding of SVM and RBF algorithms are given in [28]. After the
transformers. After training the algorithm is able to identify algorithms were trained they were used to identify the faults
the fault type of a transformer with its DGA data [24]. types of 206 transformers, of which DGA data are provided
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

MA et al.: SMART TRANSFORMER FOR SMART GRID—INTELLIGENT FRAMEWORK AND TECHNIQUES 7

TABLE II
SVM A SSESSMENT R ESULTS ON E IGHT T RANSFORMERS
OF T ESTING DATABASE

TABLE III
SVM AND RBF C LASSIFICATION ACCURACY Fig. 8. Software platform based on a MAS for transformer condition
ON T ESTING DATABASE monitoring and assessment.

from sensor during online condition monitoring of transformer.

The smartbox can also execute signal processing algorithms
and pattern recognition algorithms for signal denoising, feature
extraction, and fault pattern recognition.
To provide a platform for deploying various algorithms,
multiagent system (MAS) will be adopted as shown in Fig. 8.
by another utility. The records of these 206 transformers form MAS can provide system extensibility to include newly
a testing dataset and were not included in the training database installed sensors and the latest signal processing and pattern
(Table I). Table II shows the fault types of eight transformers recognition algorithms as they become available.
amongst these 206 transformers classified by the trained SVM MAS systems encompass attributes required to automate
and RBF algorithms. condition monitoring and diagnosis applications [29], [30].
A comparison between the algorithm classified transform- With the integration and interaction with pattern recognition
ers’ fault types and the utility expert’s manually determined algorithms, each intelligent agent in a MAS system can learn
fault types is shown in Table III. The classification accuracy from data for making reasoning. Also, through historical data
is the ratio between the number of transformers with cor- analysis using pattern recognition algorithms, these agents can
rectly classified faults types and the total number of the testing enhance their capabilities of handling uncertainties, which are
transformers (i.e., 206). important for practical intelligent condition monitoring and
It can be seen from Table III that original RBF and SVM assessment of power transformers in substation environment.
algorithms (i.e., without integrating SMOTE and boosting)
almost failed to make classification (classification accuracy is
merely above 50%). This is not surprising since the train- VI. C ONCLUSION
ing data and testing data are collected from different utilities.
This paper proposed an intelligent framework for condi-
However, with the integration of SMOTE and boosting, SVM
tion monitoring and assessment of power transformer using
and RBF algorithms can attain higher classification accuracy data obtained from online sensor measurement and historic
on the testing dataset. This demonstrates that SMOTE and
database. Within the proposed intelligent framework, a variety
boosting can assist pattern recognition algorithms in achiev-
of signal processing and pattern recognition techniques were
ing desirable generalization capability when they are applied applied for data and information processing and transformer
for classifying transformer faults.
faults identification. Case studies were provided to verify the
effectiveness of the proposed framework and applied tech-
V. S MARTBOX —H ARDWARE AND S OFTWARE P LATFORM niques. This paper is on the implementation and deployment of
FOR I MPLEMENTING P ROPOSED I NTELLIGENT the proposed intelligent framework on the smartbox hardware
F RAMEWORK and software platform.
To implement the proposed intelligent framework and
deploy signal processing and pattern recognition algorithms
ACKNOWLEDGMENT
for transformer condition monitoring and assessment, a config-
urable architecture (“smartbox”) has been adopted as hardware The authors would like to thank J. C. Chan, Y. Cui,
development platform. The smartbox consists of signal con- and J. Seo for the implementation and verification of the
ditioning and high speed digitizer, which can acquire signals algorithms.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON SMART GRID

R EFERENCES [22] J. C. Chan, H. Ma, T. K. Saha, and C. Ekanayake, “Stochastic noise

removal on partial discharge measurement for transformer insulation
[1] M. Wang, A. J. Vandermaar, and K. D. Srivastava, “Review of condition diagnosis,” in Proc. 2014 IEEE Power Energy Soc. Gen. Meeting,
assessment of power transformers in service,” IEEE Elect. Insul. Mag., National Harbor, MD, USA, pp. 1–5.
vol. 18, no. 6, pp. 12–25, Nov./Dec. 2002. [23] H. Ma, J. C. Chan, T. K. Saha, and C. Ekanayake, “Pattern recogni-
[2] Z. Xiang and E. Gockenbach, “Asset management of transformers based tion techniques and their applications for automatic partial discharge
on condition monitoring and standard diagnosis,” IEEE Elect. Insul. source classification,” IEEE Trans. Dielect. Elect. Insul., vol. 20, no. 2,
Mag., vol. 24, no. 4, pp. 26–40, Jul./Aug. 2008. pp. 468–478, Apr. 2013.
[3] A. E. B. Abu-Elanien and M. M. A. Salama, “Asset management tech- [24] H. Ma, T. K. Saha, and C. Ekanayake, “Statistical learning tech-
niques for transformers,” Elect. Power Syst. Res., vol. 80, pp. 456–464, niques and their applications for condition assessment of power trans-
Apr. 2010. former,” IEEE Trans. Dielect. Elect. Insul., vol. 19, no. 2, pp. 481–489,
[4] T. K. Saha, “Review of modern diagnostic techniques for assessing insu- Apr. 2012.
lation condition in aged transformers,” IEEE Trans. Dielect. Elect. Insul., [25] H. Ma, C. Ekanayake, and T. K. Saha, “Power transformer fault diagno-
vol. 10, no. 5, pp. 903–917, Oct. 2003. sis under measurement originated uncertainties,” IEEE Trans. Dielect.
[5] M. Duval, “Dissolved gas analysis: It can save your transformer,” Elect. Insul., vol. 19, no. 6, pp. 1982–1990, Dec. 2012.
IEEE Elect. Insul. Mag., vol. 5, no. 6, pp. 22–27, Nov./Dec. 1989. [26] N. V. Chawla, F. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer,
[6] S. R. Blue, D. Uttamchandani, and O. Farish, “A novel solid-state “SMOTE: Synthetic minority over-sampling technique,” J. Artif. Intell.
material for furfuraldehyde detection,” IEEE Trans. Dielect. Elect. Insul., Res., vol. 16, pp. 321–357, Jan. 2002.
vol. 4, no. 3, pp. 341–343, Jun. 1997. [27] G. Ratsch, T. Onoda, and K. R. Muller, “Soft margins for AdaBoost,”
[7] W. S. Zaengl, “Dielectric spectroscopy in time and frequency Mach. Learn., vol. 42, no. 3, pp. 287–320, 2001.
domains for HV power equipment, Part I: Theoretical considerations,” [28] C. M. Bishop, Pattern Recognition and Machine Learning.
IEEE Elect. Insul. Mag., vol. 19, no. 5, pp. 5–19, Sep./Oct. 2003. New York, NY, USA: Springer, 2006.
[8] J. A. S. B. Jayasinghe, Z. D. Wang, P. N. Jarman, and A. W. Darwin, [29] D. Weyns and A. M. Uhrmacher, Eds., Multi-Agent Systems Simulation
“Winding movement in power transformers: A comparison of FRA and Applications. Boca Raton, FL, USA: CRC Press, 2009.
measurement connection methods,” IEEE Trans. Dielect. Elect. Insul., [30] L. B. Cao, Ed., Data Mining and Multi-agent Integration. New York, NY,
vol. 13, no. 6, pp. 1342–1349, Dec. 2006. USA: Springer, 2009.
[9] R. Bartnikas, “Partial discharges. Their mechanism, detection and mea-
surement,” IEEE Trans. Dielect. Elect. Insul., vol. 9, no. 5, pp. 763–808,
Oct. 2002.
[10] P. Kang and D. Birtwhistle, “Condition monitoring of power trans-
former on-load tap-changers. Part I: Automatic condition diagnostics,”
IEE Proc. Gener. Transmiss. Distrib., vol. 148, no. 4, pp. 301–306,
Jul. 2001. Hui Ma (M’95) received the B.E. and M.E. degrees
[11] V. Miranda and A. R. G. Castro, “Improving the IEC table for from Xi’an Jiaotong University, Xi’an, China, in
transformer failure diagnosis with knowledge extraction from neural 1991 and 1994; the M.E. (by research) degree from
network,” IEEE Trans. Power Del., vol. 20, no. 4, pp. 2509–2516, the Nanyang Technological University, Singapore, in
Oct. 2005. 1998; and the Ph.D. degree from the University of
[12] R. Naresh, V. Sharma, and M. Vashisth, “An integrated neural fuzzy Adelaide, Adelaide, SA, Australia, in 2008, all in
approach for fault diagnosis of transformers,” IEEE Trans. Power Del., electrical engineering.
vol. 23, no. 4, pp. 2017–2024, Oct. 2008. He is currently a Research Fellow with the
[13] W. H. Tang, J. Y. Goulermas, Q. H. Wu, and Z. J. Richardson, School of Information Technology and Electrical
“A probabilistic classifier for transformer dissolved gas analysis with Engineering, University of Queensland, St. Lucia,
a particle swarm optimizer,” IEEE Trans. Power Del., vol. 23, no. 2, QLD, Australia. He has several years of experience
pp. 751–759, Apr. 2008. in research and development. From 1994 to 1995, he was a Researcher at Xi’an
[14] A. E. B. Abu-Elanien, M. M. A. Salama, and M. Ibrahim, “Calculation Jiaotong University. From 1997 to 1999, he was a Firmware Development
of a health index for oil-immersed transformers rated under 69 kV using Engineer at CET Technologies Private Ltd., Singapore. He was also a
fuzzy logic,” IEEE Trans. Power Del., vol. 27, no. 4, pp. 2029–2036, Research Engineer with the Singapore Institute of Manufacturing Technology,
Oct. 2012. Singapore, from 1999 to 2003. His current research interests include indus-
[15] A. Ashkezari, H. Ma, T. K. Saha, and C. Ekanayake, “Application of trial informatics, condition monitoring and diagnosis, power systems, wireless
fuzzy support vector machine for determining the health index of insu- sensor networks, and sensor signal processing.
lation system of in-service power transformer,” IEEE Trans. Dielect.
Elect. Insul., vol. 20, no. 3, pp. 965–973, Jun. 2013.
[16] H. Ma, T. Saha, A. Thomas, and C. Ekanayake, “Intelligent framework
and techniques for power transformer insulation diagnosis,” in Proc.
IEEE Power Energy Soc. Gen. Meeting, Calgary, AB, Canada, 2009,
pp. 1–7. Tapan K. Saha (M’93–SM’97) was born in
[17] High-Voltage Test Techniques—Partial Discharge Measurements, Bangladesh, in 1959. He received the B.Sc.
International Electrotechnical Commission (IEC), IEC Standard 60270, degree in electrical and electronic engineering from
2000. the Bangladesh University of Engineering and
[18] T. Abdel-Galil, A. Ei-Hag, A. Gaouda, M. Salama, and R. Bartnikas, Technology, Dhaka, Bangladesh; the M.Tech. degree
“De-noising of partial discharge signal using eigen-decomposition tech- in electrical engineering from the Indian Institute of
nique,” IEEE Trans. Dielect. Elect. Insul., vol. 15, no. 6, pp. 1657–1662, Technology Delhi, New Delhi, India; and the Ph.D.
Dec. 2008. degree in electrical engineering from the University
[19] S. Sriram, S. Nitin, K. Prabhu, and M. Bastiaans, “Signal denoising of Queensland, Brisbane, QLD, Australia, in 1982,
techniques for partial discharge measurements,” IEEE Trans. Dielect. 1985, and 1994, respectively.
Elect. Insul., vol. 12, no. 6, pp. 1182–1191, Dec. 2005. He has had visiting appointments for a semester
[20] X. Zhou, C. Zhou, and I. Kemp, “An improved methodology for applica- at both the Royal Institute of Technology, Stockholm, Sweden, and the
tion of wavelet transform to partial discharge measurement denoising,” University of Newcastle, Callaghan, NSW, Australia. He is currently a
IEEE Trans. Dielect. Elect. Insul., vol. 12, no. 3, pp. 586–594, Jun. 2005. Professor of Electrical Engineering with the School of Information Technology
[21] J. C. Chan, H. Ma, T. K. Saha, and C. Ekanayake, “Self-adaptive partial and Electrical Engineering, University of Queensland. His current research
discharge signal extraction based on ensemble empirical mode decom- interests include condition monitoring of electrical plants, power systems,
position and optimal morphological thresholding,” IEEE Trans. Dielect. and power quality.
Elect. Insul., vol. 21, no. 1, pp. 294–303, Feb. 2014. Prof. Saha is a Fellow of the Institution of Engineers, Australia.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

MA et al.: SMART TRANSFORMER FOR SMART GRID—INTELLIGENT FRAMEWORK AND TECHNIQUES 9

Chandima Ekanayake (M’00) received the B.Sc. Daniel Martin (M’12) received the B.E. degree
degree (Hons.) from the University of Peradeniya, in electrical and electronic engineering from the
Peradeniya, Sri Lanka, in 1999, and the Tech. Lic. University of Brighton, Brighton, U.K., in 2000, and
and Ph.D. degrees from the Chalmers University of the Ph.D. degree in electrical insulation from the
Technology, Gothenburg, Sweden, in 2003 and 2006, University of Manchester, Manchester, U.K., in
respectively, all in electrical engineering. 2008, where he investigated the suitability of using
He was a Senior Lecturer with the University vegetable oils and synthetic esters as substitutes for
of Peradeniya. He is currently a Lecturer with the mineral oil within large power transformers.
School of Information Technology and Electrical He joined Racal Electronics, Weybridge, U.K.,
Engineering, University of Queensland, Brisbane, where he served on communication and aircraft sys-
QLD, Australia. He was serving for a European tems. He was with Monash University, Melbourne,
Union Project called REDIATOOL, where he was engaged in research VIC, Australia, where he was engaged in transformer condition monitor-
related to diagnostics of transformer insulation from dielectric response mea- ing, quickly assuming the directorship of the Centre for Power Transformer
surements. Since 2001, he has been involved in condition monitoring of Monitoring. In 2013, he joined the University of Queensland, St. Lucia, QLD,
transformers installed at the Ceylon Electricity Board, Colombo, Sri Lanka. Australia, where he served on a transformer condition monitoring project
His current research interests include condition monitoring of power appa- funded by the local utilities.
ratus, alternatives for insulating oil, transient studies on power systems, and
energy related studies.
Dr. Ekanayake was the Chair of the IEEE Sri Lanka Section in
2006 and 2007.