Hindawi

Mathematical Problems in Engineering

Volume 2020, Article ID 2420456, 10 pages
https://doi.org/10.1155/2020/2420456

Research Article
Dissolved Gas Analysis of Insulating Oil in Electric Power
Transformers: A Case Study Using SDAE-LSTM

Zhao Luo ,1 Zhiyuan Zhang,1 Xu Yan,2 Jinghui Qin,1 Zhendong Zhu,1 Hao Wang,1
and Zeyong Gao1
1
Faculty of Electric Power Engineering, Kunming University of Science and Technology, Kunming 650500, China
2
Jiangsu Frontier Electric Technology Co., Ltd., Nanjing 211102, China

Correspondence should be addressed to Zhao Luo; [email protected]

Received 8 February 2020; Revised 9 September 2020; Accepted 24 October 2020; Published 11 November 2020

Academic Editor: Ricardo Aguilar-Lopez

Copyright © 2020 Zhao Luo et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Dissolved gas analysis (DGA) is the most important tool for fault diagnosis in electric power transformers. To improve
accuracy of diagnosis, this paper proposed a new model (SDAE-LSTM) to identify the dissolved gases in the insulating oil of
power transformers and perform parameter analysis. The performance evaluation is attained by the case studies in terms of
recognition accuracy, precision ratio, and recall ratio. Experiment results show that the SDAE-LSTM model performs better
than other models under different input conditions. As evidenced from the analyses, the proposed model achieves con-
siderable results of recognition accuracy (95.86%), precision ratio (95.79%), and recall ratio (97.51%). It can be confirmed that
the SDAE-LSTM model using the dissolved gas in the power transformer for fault diagnosis and analysis has great
research prospect.

1. Introduction Over the past several decades, two different types of

approaches for obtaining dissolved gas analysis results have
Power transformers are considered as the core of the been proposed. The first type belongs to mathematical ap-
electric power systems, and its running state determines proaches, which includes International Electrochemical
whether the power network is controllable or not. Power Commission (IEC) ratio, Doerneburg ratios, David triangle
transformers tend to decompose and produce some gases method, and Rogers ratios. However, these approaches have
during operation. These gases are dissolved in insulating oil some limitations, such as incompletely encoding and the
due to stress (electrical stress, mechanical stress, and absolution of fault boundary distinction. These approaches
thermal stress, respectively) [1–3]. Generally, there are are often difficult to exactly describe the running state and
many kinds of gases dissolved in insulating oil, such as trend of power transformers due to its limitations, even
hydrogen (H2), methane (CH4), acetylene (C2H2), ethylene though they are simple to operate. Meanwhile, the mathe-
(C2H4), ethane (C2H6), carbon monoxide (CO), and carbo matical approaches cannot be able to provide an interpre-
dioxide (CO2) [4]. In fact, the dissolved concentrations of tation for faults or possible trends [9].
these gases are closely associated with the running state of The second type of the approaches is applying the
the power transformers. It is common that Dissolved Gas machine learning method that stems from artificial intelli-
Analysis (DGA) is considered as a potential way to diag- gence to overcome these issues. These approaches applied
nose the essential fault both in interior and exterior of the machine learning method to diagnose the running state of
power transformer and predict how the power transformer the power transformers and revealed the trends of the power
will turn out [5–8]. Therefore, the result of DGA directly transformers. Compared with mathematical approaches,
provides enough effective information to diagnose the machine learning-based approaches performed better di-
running states of the power transformers. agnostic accuracy due to its excellent capabilities of
2 Mathematical Problems in Engineering

classification and predication of machine learning. There- And SDAE contains noise reduction factor, which can avoid
fore, these are potential means of fault diagnosis in power noise interference. In order to enhance the ability of model
transformers. to extract fault data features, an improved SDAE-LSTM
Typical machine learning methods had been imple- transformer fault diagnosis method is proposed. LSTM
mented for DGA, including artificial neural networks overcomes the “gradient” existing in traditional neural
(ANNs), support vector machine (SVM), relevance vector network in information processing. Therefore, SDAE-LSTM
machine (RVM), and fuzzy theory [10, 11]. To some extent, transformer fault diagnosis method can not only effectively
these approaches improve the accuracy of fault diagnosis, enhance the model’s ability to extract fault features but also
whereas they are usually some specific weaknesses, too. For can track the variation of gas concentration in oil with time
instance, the diagnostic accuracy of an ANN is closely re- better, improve the diagnostic accuracy, and provide a
lated to the choice of lots of parameters. Failure to find the strong guarantee for the safe and stable operation of power
most appropriate parameters for the model has a vast in- transformers.
fluence on the diagnostic accuracy. Another problem of the In this study, a LSTM-based model is developed to di-
ANN is overfitting. The diagnostic accuracy of SVM depends agnose the essential fault by the DGA. As the traditional fault
on the choice of the kernel function and corresponding diagnosis method of transformer difficult to deal with the
parameters (e.g., cost parameter, slack variables, and the fault feature information effectively, using Sade to fully mine
margin of the hyper plane), and that is not an easy thing. An the fault feature information of transformer is helpful to
RVM model reduces the computation of the kernel function, improve the accuracy of transformer fault diagnosis. Due to
so it costs more training time than SVM. the interference and influence of monitoring device, ambient
In recent years, deep learning known as deep neural temperature, personnel operation, and other aspects of
network (DNN) has attracted considerable attentions in dissolved gas content in transformer oil, it is necessary to
machine learning field. On the one hand, recurrent neural extract features from the original data. Feature extraction
network (RNN) is a type of ANNs, which especially contain a can reduce the impact of data on the model performance and
loop structure inside. This loop structure makes the input get improve the training speed and diagnostic accuracy of the
more complete information at each point, which not only model. The processed data are used as the input of LSTM.
contains the current information but also includes the pre- Then, the LSTM is performed for further analysis. To testify
vious information of the hidden layer. The hidden layers reveal the proposed model, its diagnostic accuracy is obtained from
a function of memory block, so that preserves historical in- case studies. Furthermore, three other approaches, namely,
formation and constantly renews the information with the BPANN, SVM, and RF, respectively, are compared with the
updated input. Although the RNN model uses historical in- proposed model.
formation to predict the current state, it improves the limi- The rest of this study is organized as follows. Sections 2.1
tations of remote dependence in the field of time series. But, in and 2.2 briefly demonstrates the needed methods including
practical applications, it will bring problems such as vanishing SDAE and LSTM. Section 2.3 describes the proposed model.
gradient and explosive gradient. To solve the issue, Hochreiter The experiment results of case study are presented in Section
proposed the long short-term memory (LSTM), a special RNN 3. Section 4 gives conclusions and extensive discussions.
method. LSTM is an RNN architecture published by Sepp
Hochreiter and Jürgen Schmidhuber. Like most RNNs, a 2. Proposed Model
LSTM network is universal in the sense that, given enough
network units, it can compute anything a conventional 2.1. Stacked Denoising Autoencoder. For a given data, DNN
computer can compute, provided it has the proper weight stems from ANN is a classifier consisting of a set of models
matrix, which may be viewed as its program. Unlike traditional attempts to learn high-level representations with very deep
RNNs, an LSTM network is well-suited to learn from expe- neural networks, which is usually deeper than three layers.
rience to classify, process, and predict time series when there There are three parts of a typical DNN: an input layer, an
are very long time lags of unknown size between important output layer, and a lot of hidden layers stacked between the
events [12, 13]. This is one of the main reasons why LSTM out former two layers. The network is firstly layerwise initialized
performs alternative RNNs and Hidden Markov Models and by a type of unsupervised training and then it turns into a
other sequence learning methods in numerous applications. supervised way. A large number of layers consisting of a
At present, the LSTM has been widely used in some fields, like finite number of nodes are introduced to realize highly
machine translation, automatic speech recognition, and nat- nonlinear functions in DNNs, which can grasp the pre-
ural language processing. Meanwhile, autoencoder (AE), sentation of statistical regularity from the data. And the
denoising autoencoder (DAE), and stacking denoising representation features have many virtues for classification
autoencoder (SDAE) stem from DNNs have been applied to as a common way in deep learning [15]. Figure 1 presents the
machine health monitoring and fault diagnosis. Especially, difference between typical machine learning and deep
DAE and SDAE can learn feature representations from input learning. Where, SVM (support vector machine), RVM
data by denoising, such as reconstructing input from artificial (relevance vector machine), RF (random forest), and PCA
corruption [14]. Inspired by that, we try to consider using (principal component analysis) are few of the most widely
SDAE-LSTM in fault diagnosis of power transformers. used data dimensionality reduction algorithms. The main
SDAE adopts unsupervised learning method, which can idea of PCA is to map the n-dimensional features to the k-
learn the internal characteristics of fault data more deeply. dimension, which is a new orthogonal feature, also known as
Mathematical Problems in Engineering 3

Regular machine learning

SVM ANN
PCA
Data Output
acquisition RVM RF
Feature extraction
and feature selection Model training

Softmax/
regression
Data
Output
acquisition

Feature representation

Deep learning

Figure 1: Comparison between regular machining learning and deep learning.

the principal component. It is a k-dimensional feature the noise in the original input. In the other words, it is not
reconstructed from the original n-dimensional features. helpful to extract the useful features from the original input.
An AE is a three-layer network, which consist of an Vincent first proposed a SDAE model to escape an
encoder and a decoder. On the one hand, the encoder acts on equivalent representation of the original input. A SDAE is a
the input data from a high-dimensional space to a low- stack of single-layer DAE, which takes a corrupted version of
dimensional space. On the other hand, the decoder re- data as input to reconstruct or denoise original input.
constructs the input data from the corresponding codes. On Figure 2 presents the construction of the SDAE [16]. Above
the basis of the training samples xi ∈ Z, the encoder all, a DAE randomly breaks the original input based on qD ,
transforms the input vector x into a hidden representation so that the original input turns to x ̃. To make y become the
hi ∈ Z through a nonlinear mapping as in the following hidden representation, take the encoding function f that
equation: acts on x􏽥 . And, the decoding function g finally reconstructs y
to z. Here, qD is one of the kinds of functions, such as
h � α W1 x + b1 􏼁, (1)
Gaussian noise, Masking noise, and Salt-and-pepper noise.
where α(x) � 1/[1 + exp(−x)] is a nonlinear activation Likewise, the encoding and decoding function both are
function, which is used for nonlinear deterministic mapping. active function (i.e., sigmoid function, tanh function, and
Then, decoder maps hidden representation back to the ReLU function). Generally, the reconstruction error is
original representation in a similar way in the following minimized to acquire the appropriate parameters of DAE.
equation: The well-trained encoding function acts on the original
input to generate the feature representation, which is served
g � α W2 x + b2 􏼁. (2) as the input at the back [17]. The SDAE could be built quite
mechanically by repeating the certain straightforward pro-
The training process of an AE is conducted to optimize cesses over and over again as shown in Figure 2. By adding
the parameter set, θ � 􏼈W1 , b1 , W2 , b2 􏼉, that is evaluated by noise in training data, the AE will reconstruct the original
minimizing the reconstruction error between g and x. input data, so that the DAE not only overcomes the in-
Generally, the measurement of the average reconstruction terference of the noise but also captures more effective
error is mean square error (MSE). The calculation method of features and improves the generalization ability of the
MSE is shown in model.
1 m 1��� ��2 In general, the training process of SDAE can be divided
JMSE (θ) � 􏽘􏼒 �z − xi �� 􏼓 into two stages: the unsupervised pretraining and the su-
m i�1 2 i
pervised finetune. Unsupervised pretraining is conducted
(3)
by training the parameters of initialization network, so that
1 m 1�� ��2
� 􏽘􏼒 ��fW,b xi 􏼁 − xi �� 􏼓. each layer is trained as a DAE by minimizing the error in
m i�1 2 reconstructing its input. After that, the network goes
An AE attempts to learn an identity map for the approx- through a supervised finetune by adding a logistic re-
imation between output and input in the training process. gression layer on top of network. As a result, the gener-
However, the retaining information of the original input does alization performance of the SDAE is improved
not guarantee to separate effectively the useful information from dramatically.
4 Mathematical Problems in Engineering

g2

g f2

f qD f2

qD f f

Feature extraction of
Single-layer DAE Double-layer DAE SDAE

Figure 2: The structure of SDAE.

xt ht–1 xt ht–1
2.2. LSTM. LSTM is an excellent variant model of RNN, which
inherits the characteristics of most RNN models, and at the
same time solves the Vanishing Gradient problem caused by ht
the gradual reduction in the gradient back propagation process.
When it comes to information processing tasks, LSTM is very
suitable for processing problems that are highly related to time tanh
series. LSTM has a very good performance when dealing with Ct–1
Ct Ct
time-series data classification problems. Considering that
LSTM has solved the “gradient dissipation” problem of tra-
ditional neural networks in terms of information processing, it
works better than RNN when dealing with some problems that
require long-term storage. Therefore, this article will use the
it Ct
LSTM model for data processing.
LSTM is a collection of basic unit, which is the memory
block consisting of memory cell and three gates of control
memory cell state, namely, forget gate, input gate, and output xt ht–1 xt ht–1
gate, respectively. The forget gate plays an important role on Figure 3: The structure of memory block.
removing some unnecessary information, and the input gate
determines the effect of input on the state of memory units.
Likewise, the output gate selects and sends out the output.
Finally, the output gate gives the current information
Figure 3 presents the structure of the memory block. Given the
as follows:
input vector xt at the time t and the output ht − 1 at the time
t − 1, we define the W, U, and b as weight matrix and bias ot � σ WO xt + UO ht−1 + bo 􏼁,
vector. The memory block refreshes the state and decides the (6)
ht � ot ∗ tanh Ct 􏼁,
output.
Firstly, the forget gate removes the historical and futile where σ represents the logistic sigmoid function and Ct
information, as can be shown in represents the state of the memory cell at the time t
Meanwhile, ft , it , and ot represent the output state of forget
ft � σ 􏼐Wf xt + Ui ht−1 + bf 􏼑. (4) gate, input gate, and output gate at the time t, respectively.
LSTM improves the traditional memory block of RNN
Then, on the basis of the input and historical infor- model (single tanh layer or single sigmoid layer), that
mation, the input gate refreshes the state. It is given by overcome the limitation of long-term preservation for in-
formation [18]. By the design of gate structure and memory
it � σ Wi xt + Ui ht−1 + bi 􏼁, cell, LSTM can effectively update and transmit the key in-
􏽥 t � tanh WC xt + Uc ht−1 + bc 􏼁,
C (5) formation in time series. Compared with the traditional
RNN model, LSTM performs better on information selec-
􏽥 t.
Ct � ft−1 ∗ Ct−1 + it ∗ C tion and learning characteristics of time series, which can
Mathematical Problems in Engineering 5

1
t–
T=
t
Output: ht T=
1
Hidden layer: ct t+
T=
Input: xt

Figure 4: The internal loop structure of the LSTM.

apply the long-distance information to the current predic- dispersed that reflect enormous discrepancy among the
tion. The internal loop structure of the LSTM neural network orders of magnitudes. On the other hand, the DGA data may
is shown in Figure 4 [19]. As can be seen from it, the input of have great uncertainty for the same type of fault due to the
each time step is a vector, and the state of the hidden layer diversity. In that, various ratio method are proposed to solve
gradually moves backward according to the time series. these drawbacks, such as composition ratio (H2%, CH4%,
As a special structure, it is necessary to input a sequence C2H2%, C2H4%, C2H6%), IEC three-ratio codes (CH4/H2,
of feature vectors, which consists of continuous feature C2H2/C2H4, C2H4/C2H6), Rogers four-ratio codes (CH4/H2,
vector from M time steps. In that, it is necessary to construct C2H2/C2H4, C2H4/C2H6, C2H6/CH4), and noncode ratio
the input sequence of LSTM model before training. Given (CH4/H2, C2H2/C2H4, C2H4/C2H6, C2H2/(C1 + C2), H2/
the input vector xt at the t time step, the constructed se- (H2 + C1 + C2), C2H4/(C1 + C2), CH4/(C1 + C2), C2H6/
quence is 􏼈xt−M+1 , xt−M+2 , . . . , xt 􏼉. More specifically, the first (C1 + C2), (CH4 + C2H4)/(C1 + C2)).
sequence is x1 , x2 , . . . , xM and the second sequence is In this study, we have proposed a SDAE-LSTM model
x2 , x3 , . . . , xM+1 . Likewise, the other sequences are acquired. for transformer incipient fault diagnosis. As a typical un-
supervised deep learning network, SDAE can dig the deep-
level features of unlabeled data. Without adding any labels to
2.3. SDAE-LSTM Model. In this section, we apply a SDAE- the input data, training can derive the intrinsic feature of the
LSTM model consisting of deep neural network and LSTM input data, so that the input intrinsic can extract high-di-
to diagnose the faults of the power transform. mensional features; tap the potential value within the data.
The framework of SDAE-LSTM model is shown in Through the monitoring of relevant data during transformer
Figure 5. As seen in Figure 5, the first step is to extract the failure, it can be found that there are large numerical dif-
features from the original data vector by SDAE. Then, the ferences between different parameters. If the original gas
last hidden layer of the SADE is regarded as the input of the concentration data are directly placed in the model for
LSTM, which is an effective representation of the original training operations, the weight setting may be affected due to
data. the large difference in data. Due to different orders of
According to the common faults in daily monitoring, magnitude, a larger value will affect the training results,
this paper establishes the following data tags of power resulting in distortion of the training results of the entire
transformer status: classified into normal pattern (F1), network. In order to avoid the influence of different original
discharge with high energy pattern (F2), discharge with low data orders on training weights, it is necessary to preprocess
energy pattern (F3), partial discharge pattern (F4), high- the gas data in transformer oil. Because the data range is
temperature thermal pattern (F5), middle-temperature relatively concentrated, this article uses a standardized
thermal pattern (F6), and low-temperature thermal pattern method. A normalization method has been developed to
(F7). normalize the input vector as follows:
x − xmin
2.4. Input Vector. Based on a large number of researches on x∗ � , (7)
xmax − xmin
transformer, the gases selected usually contains five hy-
drogenous gases, such as hydrogen (H2), methane (CH4), where x presents the original vector, x∗ means standardized
acetylene (C2H2), ethylene (C2H4), and ethane (C2H6). vector, and xmax and xmin represent the maximum and
However, it is not uncommon that some limitations are minimum separately.
revealed for diagnosing faults on the basis of these five gases. After the LSTM training, the output results are the final
On the one hand, the values of these gases largely are classification of the original input features, and all the
6 Mathematical Problems in Engineering

LSTM ht–1 ht ht+1

tanh tanh tanh

σ σ tanh σ σ σ tanh σ σ σ tanh σ

Xt–1 Xt Xt+1
SDAE

Figure 5: The framework of the SDAE-LSTM transformer fault diagnosis system.

experimental results are obtained by 10-fold cross- percentage of particular samples recognized correctly in the
validation. entire same samples. This criterion reflects the performance
on recognizing correctly the particular patterns:
3. Case Studies bf1
PR � × 100%, (9)
In this section, case experiments confirmed that the proposed bf
model has good classification performance. All experiments are
coded in PYTHON 3.7 and implemented on Intel Core i7- where bf means the number of samples, which come from
6200CPU, 2.30 GHz computer with an 8 GB memory. The the same pattern. bf1 gives the information about the
SDAE and LSTM are established by TensorFlow of Python. number of samples recognized correctly in the whole same
In this case, we collect the practical data consisting of samples.
normal and abnormal state from China national grid and In addition, the recall ratio (RR) as an effective index is
China southern power grid, which is called the original utilized for measuring the performance of the proposed
dataset. The dataset was composed of 1723 samples and model in fault diagnosing. It demonstrates the percentage of
randomly divided into a training set (1378 samples, 80 a particular pattern recognized in the whole number of the
percentage) and a testing set (345 samples, 20 percentage). same patterns. Moreover, recall ratio and prediction ration
On basis of that, we have conducted some experiments to are regarded as two interconstrain measures:
validate the efficiency of the proposed model in this section.
bf1
RR � × 100%, (10)
df
3.1. Evaluation of Fault Diagnosis Accuracy. The fault di-
agnosis accuracy is quantified by recognition accuracy (RA), where df means the whole number of samples.
which has been widely used in machine learning models. The In this study, RA not only is an indicator in measuring,
criterion means the proportion of samples recognized but beyond that, PR and RR are effective indexes for eval-
correctly in the whole samples recognized, which represents uating the fault diagnosis. In that, performance of the
the overall performance of the SDAE-LSTM model: proposed model can be quantified objectively and overall.
af
RA � × 100%, (8)
af + cf
3.2. Comparison with Different Inputs. The RAs, PRs, and
where af means the number of samples recognized cor- RRs are attained from the same model (LSTM model) with
rectly and cf gives the number of the whole samples different inputs in this section, classified into DGA, com-
recognized. position ratio, IEC three ratio codes, Rogers four-ratio codes,
The fault diagnosis accuracy also can be evaluated by noncode ratio, and SDAE. That is to say, the experiments are
precision ratio (PR), which gives the information on the conducted to verify the importance of SDAE for LSTM
Mathematical Problems in Engineering 7

Table 1: RAs’ comparison for different inputs.

Inputs
Patterns
DGA (%) Composition ratio (%) IEC three-ratio codes (%) Rogers four-ratio codes (%) Noncode ratio (%) SDAE (%)
F1 93.13 96.62 93.37 97.04 97.12 98.34
F2 89.74 96.67 90.94 95.59 95.58 97.32
F3 87.54 94.01 91.83 92.97 93.35 94.57
F4 89.96 91.25 89.44 93.21 93.28 95.24
F5 89.68 90.43 88.09 94.80 92.39 94.21
F6 92.80 90.53 95.20 92.77 92.45 98.15
F7 85.34 94.91 94.65 90.56 92.93 93.16
Average 89.74 93.49 91.93 93.85 93.87 95.86

model in Section 4.2. Table 1 gives the information about recall ratio via the model, after inputting the former five
RAs, after inputting the different vectors to the same model. inputs separately, 91.92% in noncode ratio, 89.58% in
As is evident from the Table 1, the model gains the best Rogers four-ratio codes, 83.44% in IEC three-ratio codes,
recognition accuracy when it comes to SDAE, which 76.88% in composition ratio, and 82.41% in DGA. Most
attained 95.86% on average alone. Four recognition accu- recall ratios of seven patterns are less than 80% for the
racies in seven patterns are more than 95%, especially two of former five inputs. Considering that, according to the RRs’
them are more than 98% dramatically. Moreover, the model comparison for different input, the model is more sensitive
performs well in terms of recognition accuracy on average, to recognize seven patterns of electric power transformers
which are moderately worse than SDAE-LSTM model, for SDAE.
93.49% in composition ration, 93.85% in Rogers four ratio
codes, and 93.87% noncode ratio. Every RA is no less than
90%, whilst no more than two patterns’ recognition accu- 3.3. Comparison with Other Methods. Backpropagation
racies surpass to 95%. In a stark contrast, negative phe- neural network (BPANN), support vector machine (SVM),
nomena are found in results of the model, which are and random forest (RF) are the most traditional machine
inputted by DGA. The average recognition accuracy is learning methods to recognize patterns in this interest. In
merely to 89.74%, and most of patterns are recognized to less this section, four comparative experiments are conducted to
than 90%. After all, it is not uncommon due to its originality illustrate the potential merits of the SDAE-LSTM model
in DGA. In that, the SDAE attaches its importance on the which compare the RAs, PRs, and RRs with these three
recognition ability of the model. common methods. The parameters of these three models are
Table 2 presents the precision ratios for different inputs shown in Table 4.
by the proposed model. A thorough analysis from the table is Table 5 gives the information about recognition accu-
that the precision ratios of normal pattern are acquired to racies of different models for dissolved gas analysis process
more than 95% for all different inputs. For other patterns, of insulating oil in electric power transformers. As shown
the model has poor performances for the former five inputs, from it, the SDAE-LSTM model achieves the best average
where one of patterns acquires the precision ratio less than recognition accuracy to 95.86%. The recognition accuracies
90% at least. Especially, most of precision ratios are less than of patterns surpass 95% except three of them. In particular,
90% when it comes to DGA as the input of the proposed the recognition accuracies of normal discharge with high
model. Interestingly, the most outstanding performance is energy and middle temperature thermal pattern are more
shown for SDAE input in terms of the average precision than 97%. The next one is RF model, which is moderately less
ratio, which is to considerable 95.79%. The next one is by 3.6% on average RA. Interestingly, two of seven perform
slightly less by 2% around on average, 94.84% in noncode well, 95.89% in normal pattern, and 95.13% in partial dis-
ratio, 92.59% in Rogers four-ratio codes, 92.67% in IEC charge pattern. The rest of them have middle positions, even
three-ratio codes, and 93.46% in composition ratio, re- the RA of low temperature thermal pattern is merely to
spectively. There is no doubt it is not a good result for DGA 88.54%. The BPANN and SVM have the worst average RA,
by the model, which is merely to 89.83% on average. As a which are 82.90% and 86.98%, respectively. It is obvious that
result, the proposed model reflects the most potential per- the RAs of patterns are less than 90% except only one
formance for SDAE when it comes to precision ratio. pattern, sometimes even less than 80%. Therefore, the
Table 3 illustrates the recall ratios for different inputs. proposed model performs better than other three common
As shown in Table 3, the LSTM model for SDAE reflects approaches on the recognition accuracies.
better recall ratios than other inputs, which attained 97.51% Table 6 represents detailed results of different models in
on average alone, and even recall ratio and prediction terms of precision ratio. As is the most evident from it, the
ration are regarded as two interconstrain measures. More SDAE-LSTM model reflects the largest average precision
specifically, every pattern’s recall ratio almost is more than ratio, which attained 95.79%. More specifically, the precision
95%, which are considerable. On the contrary, there are not ratios are more than 95% considerately except the precision
great generalization capacity in the terms of the average ratios of partial discharge pattern (93.86%) and low
8 Mathematical Problems in Engineering

Table 2: PRs’ comparison for different inputs.

Inputs
Patterns
DGA (%) Composition ratio (%) IEC three-ratio codes (%) Rogers four-ratio codes (%) Noncode ratio (%) SDAE (%)
F1 97.49 96.11 95.91 95.65 97.21 98.52
F2 87.69 93.87 89.75 89.34 94.46 95.53
F3 87.37 91.96 91.64 92.49 93.91 95.19
F4 88.53 89.54 92.31 93.56 97.05 93.86
F5 86.42 94.25 94.97 90.73 89.57 95.85
F6 92.98 96.08 92.85 94.77 96.73 97.62
F7 88.34 92.40 91.26 91.58 94.94 93.97
Average 89.83 93.46 92.67 92.59 94.84 95.79

Table 3: RRs’ comparison for different inputs.

Inputs
Patterns
DGA (%) Composition ratio (%) IEC three-ratio codes (%) Rogers four-ratio codes (%) Noncode ratio (%) SDAE (%)
F1 94.10 89.93 90.48 93.24 94.37 94.48
F2 90.11 90.01 90.35 93.13 95.61 97.44
F3 56.14 65.79 56.91 73.58 90.76 97.79
F4 92.51 97.11 82.38 91.14 88.57 99.73
F5 77.95 85.74 82.15 90.79 94.11 97.66
F6 87.36 89.24 93.40 92.79 90.91 95.89
F7 78.75 79.55 88.40 92.42 89.15 99.61
Average 82.41 76.88 83.44 89.58 91.92 97.51

Table 4: RAs’ comparison for different methods.

Models Parameters Values
Hidden unit number 16
Activation function Tansig
Training method Traingd
BPANN
Learning rate 0.05
Training iterations 0.001
Objective value 1000
Kernel function RBF
SVM RBF sigma σ 8.1493
Box constraint C 1
The number of trees (ntree) 500
RF
The capacity of feature space (mtry) 4

temperature thermal pattern (93.97%). Moreover, the av- LSTM model reflects better recall ratio performance than
erage precision ratio of RF model is attained to 91.04%, the other three traditional machine learning models, while
which is less than 2.93%. Most of the seven patterns’ pre- the single recall rate is greater than 94.4%, and the average
cision ratios are more than 90%. In addition, the SVM model recall rate is reaching 97.51%. The following one is the RF
gets a worse average precision ratio (85.83%), since four model, which is 91.64% in average when it comes to recall
patterns’ precision ratios are less than 85%. In particular, the ratio. Interestingly, although one of patterns is merely
SVM model gains the smallest value in discharge with low 87.50% (discharge with low energy pattern), the rest are
patterns (70%) among twenty-eight data when it comes to more than 90%. Comparing with RF and SDAE-LSTM
precision ratio. Furthermore, the average precision ratio of models, the average recall ratios are in stark contrast both
BPANN model is the smallest among the figures of other in BPANN model and SVM model, 78.57% and 84.24%,
three models, which is merely to 81.34%. As a consequence, respectively. It should be clear that sharp declines are found
these experiment results reveal that the SDAE-LSTM model in these fourteen datums, such as discharge with low energy
performs best among four models in terms of precision ratio. patterns in BPANN model (46.88%) and SVM model
Recall ratios’ comparison of different models is shown (65.63%). Arising from analyses represented, performance
in the Table 7. To compare rationally the SDAE-LSTM of the SDAE-LSTM model is the best when it comes to
model with other approaches, recall ratios are conducted as recall ratio.
well due to interconstrain relationship between precision In conclusion, merits of the SDAE-LSTM model are
ratio and recall ratio. As can be seen in Table 7, the SDAE- revealed in three aspects in comparison with other three
Mathematical Problems in Engineering 9

Table 5: RAs’ comparison of different models. accuracy, precision ratio, and recall ratio [20, 21]. Next,
Models
another comparison experiment result demonstrated that
Patterns the proposed model achieves better performance than three
BPANN SVM RF SDAE-LSTM
traditional machine learning models (BPANN, SVM, and RF
F1 86.41 91.77 95.89 98.34 models). When it comes to recognition accuracy, the SDAE-
F2 82.84 85.53 91.12 97.32
LSTM attains 95.86%, which is obviously higher than
F3 79.23 84.29 92.67 94.57
F4 80.17 88.69 95.13 95.24
BPANN (82.90%), SVM (86.96%), and RF (92.17%). Simi-
F5 82.33 86.49 90.28 94.21 larly, the SDAE-LSTM model gains better precision ratio
F6 89.45 86.16 91.55 98.15 (95.79%) than BPANN (81.34%), SVM (85.83%), and RF
F7 79.86 85.78 88.54 93.16 (91.04%). Obviously, a larger advantage is found in terms of
Average 82.90 86.96 92.17 95.86 recall ratio for the proposed model than of BPANN
(78.57%), SVM (84.24%), and RF (91.64%) as well, which is
97.51% alone. That means the SDAE-LSTM model plays
Table 6: PRs’ comparison of different models. more promising role for recognizing patterns in dissolved
Models gas analysis of insulating oil in electric power transformers.
Patterns Considering that the SDAE-LSTM model is a promising
BPANN SVM RF SDAE-LSTM
model for analyzing dissolved gas in the insulating oil of
F1 86.11 85.37 91.89 98.52
F2 83.00 87.91 92.39 95.53 power transformers by identifying patterns, we can further
F3 78.95 70.00 87.50 95.19 study its application in online monitoring of power trans-
F4 80.00 84.21 85.71 93.86 formers. In addition, the comparisons with different ca-
F5 83.61 84.62 91.67 95.85 pacities of samples are also interesting.
F6 85.71 93.67 96.15 97.62
F7 72.00 95.00 92.00 93.97 Data Availability
Average 81.34 85.83 91.04 95.79
The data used to support the findings of this study are
available from the corresponding author upon request.
Table 7: RRs’ comparison of different models.
Models
Conflicts of Interest
Patterns
BPANN SVM RF SDAE-LSTM The authors declare that they have no conflicts of interest.
F1 86.11 97.22 94.44 94.48
F2 92.22 88.89 94.44 97.44 Acknowledgments
F3 46.88 65.63 87.50 97.79
F4 80.00 80.00 90.00 99.73 This work was supported in part by the National Natural
F5 85.00 91.67 91.67 97.66 Science Foundation of China (grant no. 51907084), Yunnan
F6 87.80 90.24 91.46 95.89 Provincial Talents Training Program (grant no.
F7 72.00 76.00 92.00 99.61
KKSY201704027), and Scientific Research Foundation of
Average 78.57 84.24 91.64 97.51 Yunnan Provincial Department of Education (grant no.
2018JS032).
traditional machine learning models (BPANN, SVM, and RF
models), classified into recognition accuracy, precision ratio, References
and recall ratio. To be more specific, the recognition ac-
curacies of seven patterns all reach high level, especially in [1] N. Lelekakis, D. Martin, W. Guo, and J. Wijaya, “Comparison
normal and middle temperature thermal patterns. Similarly, of dissolved gas-in-oil analysis methods using a dissolved gas-
in-oil standard,” IEEE Electrical Insulation Magazine, vol. 27,
higher precision ratios and recall ratios are found in the
no. 5, pp. 29–35, 2011.
SDAE-LSTM model than in the other three models. To put it [2] J. I. Aizpurua, V. M. Catterson, B. G. Stewart et al., “Power
in a nutshell, the SDAE-LSTM model is more effective model transformer dissolved gas analysis through Bayesian networks
for pattern recognition in the dissolved gas analysis process and hypothesis testing,” IEEE Transactions on Dielectrics and
of insulating oil in electric power transformers. Electrical Insulation, vol. 25, no. 2, pp. 494–506, 2018.
[3] J. Fan, Z. Liu, A. Meng et al., “Characteristics of tin oxide
chromatographic detector for dissolved gases analysis of
4. Conclusions transformer oil,” IEEE Access, vol. 7, pp. 94012–94020, 2019.
[4] J. Jiang, Z. Wang, X. Han et al., “Multi-gas detection in power
This study proposes a new model based on LSTM to analyze
transformer oil based on tunable diode laser absorption
the dissolved gas of insulating oil in electric power trans-
spectrum,” IEEE Transactions on Dielectrics and Electrical
formers. Two parts’ experiments are conducted to evaluate Insulation, vol. 26, no. 1, pp. 153–161, 2019.
the proposed model. As the comparison experiments for [5] J. Faiz and M. Soleimani, “Dissolved gas analysis evaluation in
different inputs (DGA, composition ratio, IEC three-ratio electric power transformers using conventional methods a
codes, Rogers four-ratio codes, noncode ratio, and SDAE) review,” IEEE Transactions on Dielectrics and Electrical
show, the SDAE-LSTM model is superior on recognition Insulation, vol. 24, no. 2, pp. 1239–1248, 2017.
10 Mathematical Problems in Engineering

[6] J. Faiz and M. Soleimani, “Assessment of computational in- [20] I. B. M. Taha, S. S. M. Ghoneim, and H. G. Zaini, “Im-
telligence and conventional dissolved gas analysis methods for provement of rogers four ratios and IEC code methods for
transformer fault diagnosis,” IEEE Transactions on Dielectrics transformer fault diagnosis based on dissolved gas analysis,”
and Electrical Insulation, vol. 25, no. 5, pp. 1798–1806, 2018. in Proceedings of the 2015 North American Power Symposium
[7] S. Eeckhoudt, S. Autru, and L. Lerat, “Stray gassing of (NAPS), Charlotte, NC, USA, October 2015.
transformer insulating oils: impact of materials, oxygen [21] X. F. Wang, Z. D. Wang, Q. Liu, G. Wilson, D. Walker, and
content, additives, incubation time and temperature, and its P. W. R. Smith, “Dissolved gas analysis (DGA) of mineral oil
relationship to oxidation stability,” IEEE Electrical Insulation under thermal faults with tube heating method,” in Pro-
Magazine, vol. 33, no. 6, pp. 27–32, 2017. ceedings of the 2017 IEEE 19th International Conference on
[8] J. Li, Q. Zhang, K. Wang, J. Wang, T. Zhou, and Y. Zhang, Dielectric Liquids (ICDL), Aberdeen City, UK, June 2017.
“Optimal dissolved gas ratios selected by genetic algorithm for
power transformer fault diagnosis based on support vector
machine,” IEEE Transactions on Dielectrics and Electrical
Insulation, vol. 23, no. 2, pp. 1198–1206, 2016.
[9] J. Jiang, R. Chen, M. Chen, W. Wang, and C. Zhang, “Dy-
namic fault prediction of power transformers based on hidden
markov model of dissolved gases analysis,” IEEE Transactions
on Power Delivery, vol. 34, no. 4, pp. 1393–1400, 2019.
[10] S. S. M. Ghoneim, I. B. M. Taha, and N. I. Elkalashy, “In-
tegrated ANN-based proactive fault diagnostic scheme for
power transformers using dissolved gas analysis,” IEEE
Transactions on Dielectrics and Electrical Insulation, vol. 23,
no. 3, pp. 1838–1845, 2016.
[11] S. Li, G. Wu, B. Gao, C. Hao, D. Xin, and X. Yin, “Inter-
pretation of DGA for transformer fault diagnosis with
complementary SaE-ELM and arctangent transform,” IEEE
Transactions on Dielectrics and Electrical Insulation, vol. 23,
no. 1, pp. 586–595, 2016.
[12] J. Maumela, F. Nelwamondo, and T. Marwala, “Condition
monitoring of transformer bushings using rough sets, prin-
cipal component analysis and granular computation as pre-
processors,” in Proceedings of the 2013 International
Conference on System Science and Engineering (ICSSE),
Budapest, Hungary, July 2013.
[13] L. Kong, L. Luan, K. Zhou, C. Chen, J. Chen, and Z. Wang,
“Running state prediction and evaluation of power trans-
formers,” in Proceedings of the 2019 IEEE 4th International
Conference on Advanced Robotics and Mechatronics (ICARM),
Toyonaka, Japan, July 2019.
[14] J. Dai, H. Song, G. Sheng, and X. Jiang, “Cleaning method for
status monitoring data of power equipment based on stacked
denoising autoencoders,” IEEE Access, vol. 5, pp. 22863–
22870, 2017.
[15] B. Tan, J. Yang, Y. Tang, S. Jiang, P. Xie, and W. Yuan, “A deep
imbalanced learning framework for transient stability as-
sessment of power system,” IEEE Access, vol. 7, pp. 81759–
81769, 2019.
[16] E. K. Wang, X. Zhang, and L. Pan, “Automatic classification of
CAD ECG signals with SDAE and bidirectional long short-
term network,” IEEE Access, vol. 7, pp. 182873–182880, 2019.
[17] W. Mao, J. Chen, X. Liang, and X. Zhang, “A new online
detection approach for rolling bearing incipient fault via self-
adaptive deep feature matching,” IEEE Transactions on In-
strumentation and Measurement, vol. 69, no. 2, pp. 443–456,
2020.
[18] H. Shi, L. Guo, S. Tan, and X. Bai, “Rolling bearing initial fault
detection using long short-term memory recurrent network,”
IEEE Access, vol. 7, pp. 171559–171569, 2019.
[19] J. Dai, H. Song, G. Sheng, and X. Jiang, “LSTM networks for
the trend prediction of gases dissolved in power transformer
insulation oil,” in Proceedings of the 2018 12th International
Conference on the Properties and Applications of Dielectric
Materials (ICPADM), Xi’an, China, May 2018.