Professional Documents
Culture Documents
This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TPWRD.2023.3248157
Abstract –This paper presents two approaches to power required [4]. It is desirable to utilize a non-invasive method to
circuit breakers condition assessment. The first one covers a obtain indices that will be used for the CB condition state
wide variety of machine learning classification algorithms where decision stage. Due to the mechanical nature of the CB, the
the input for the classification is a manually selected feature set. vibration signature recorded during the making and breaking
The second one utilizes deep learning classification based on the operation of the CB contains valuable information that can
convolutional neural network. Both approaches revolve around indicate a potential issue. The acquiring vibration signature
the idea behind spectral kurtosis, one of which exploits its visual does not require interfering with the normal operation thus it
representation in the form of kurtogram. The first approach uses represents a non-invasive measurement method.
a spectral kurtosis curve as the base for feature extraction while
the second approach uses a spectral kurtosis kurtogram as a Different approaches for analyzing the CB vibration
single input into the convolutional neural network. The fingerprint have been used in literature and engineering
validation is performed on a large set of vibration signatures practice. In general, most of the methods available in literature
and compared to competing state-of-the-art algorithms. The can be separated into two stages: the indices computation
results indicate promising features of the proposed approach. stage and the decision stage. For computing the suitable
indices, the authors have adopted the analysis in the time
Keywords—circuit breakers, vibration measurement, spectral domain, frequency domain, and time-frequency domain as
kurtosis, kurtogram, machine learning, deep learning
well as other approaches combining the previously stated
I. INTRODUCTION domains. In the early stages of this topic, researchers used
approaches based on indices derived from statistical analysis
More than two decades have passed since the first research in the time domain [5]. Furthermore, the approach called Short
results related to the vibration-based condition assessment of Time Energy (STE) uses windowing of the time domain signal
the mid and high-voltage circuit breakers (CB) were to obtain the energy of each segment [6]. On the other hand,
published. Ever since, the reliability of the CBs, as well as in the frequency domain, some of the authors approached the
condition assessment methods, has tremendously improved. analysis in a simple frequency spectrum of the vibration
However, the vibration-based methods until the present day, fingerprint [7].
have not been standardized and well accepted as the traditional
offline-based methods. Furthermore, the CB vibration However, due to the non-stationary nature of the signal,
fingerprint mirrors its mechanical nature. The CB is one of the most of the authors reject the frequency domain as the single
most important and common type of switch equipment in a tool for the condition assessment task. The time-frequency
power system. Since the CB is an electromechanical system, domain has been utilized in the form of Continues Wavelet
the energy stored in the operating mechanism is released and Transform (CWT), Discrete Wavelet Transform (DWT) [8],
transmitted to the moving contacts through the mechanical [9] and Hilbert Huang Transform (HHT) [10], [11]. Since the
elements and linkages in a very short time frame. The high non-stationary nature of the CB vibration fingerprint, some of
energy of the moving mechanical parts transfer results in the provided approaches in the literature are based on spectral
extrusion, friction and impact of components and decomposition of the signal based on Empirical Mode
subassemblies of the CB [1]. The mechanical nature of the Decomposition (EMD) [12], Ensemble Empirical Mode
CB over time results in a potential decrease in performance. Decomposition (EEMD) [13], Local Mean Decomposition
Even though the reliability of the CB is continuously (LMD) [14], ensemble empirical mode decomposition with
improved, failures still take place which sometimes implies adaptive noise (CEEMDAN) [15] and Variable Mode
costly consequences [2]. This calls for widely used time-based Decomposition (VMD) [1], [16]. Combining previously
condition assessment and maintenance of the CB. However, mentioned frequency and time-frequency methods with fields
the maintenance cost for aging CB is significant [3]. Thus, the such as chaos theory, some authors based their work on fractal
improvement of the CB reliability must be followed by the dimension [13], [17], ergodic quantities such as Largest
improvement of the condition assessment methods. Lyapunov exponent [18] or information entropy [19], [20].
Also, there is a large portion of available literature that utilizes
The time-based maintenance is supplemented and, in some variations of a well-known algorithm derived from the speech
cases, replaced by condition-based maintenance. The recognition niche, the Dynamic Time Warp (DTW) [21], [22],
excessive overhaul and expenses of time-based maintenance [23], [24].
are reduced by the utilization of condition-based maintenance.
However, to preserve the reliability of the power system grid The second stage – the decision stage – is directly
and its subassemblies including the CB, an accurate and dependent on the previous one. Thus, the authors choose
minimally invasive approach is a must. Most of the classical machine learning classification methods based on Support
CB testing methods and approaches are invasive, meaning the Vector Machine (SVM), [18], [19], [25], [26], Self-
partial or complete interruption of the CB normal operation is Organizing Maps (SOM) [13], [27], Convolutional Neural
__________________________________
TABLE I
SELECTED METHODS OVERVIEW
Deep Output Assessment
No. Reference Domain Isolated faults Validation object
learning (Classes) signal
1. Yang et.al. [47] Time-Freq No Multiclass (4) Vibration Not stated Not stated
2. Ye et.al. [28] Time Yes Binary Vibration - ECC VS1-ZN63A
3. Li et al. [25] Time-Freq No Multiclass (3) Vibration The loose screw, falling of the spring Not stated
Coil plunger jam, poor contact in coil
4. Wang et al. [48] Time-Freq Yes Multiclass (3) Vibration ECC VS1-ZN63A
circuit, jam in open latch
5. Mingliang et al. [49] Time-Freq No Multiclass (3) Vibration The loose screw, time delay Not stated
The jam between the coil and the
Vibration +
6. Ruan et al. [50] Time No Multiclass (4) plunger, poor contact, jam in the open ECC VS1 series
Coil current
latch
7. Method BT Freq. No Binary Vibration - ABB Emax SACE
8. Method CNNK Time-Freq Yes Binary Vibration - ABB Emax SACE
Networks (CNN) [10], [28], Extreme Learning Machine accuracy, sensitivity, specificity, F1 score and Matthew
(ELM) [29], Classification and regression tree (CART) [30], Correlation Coefficient.
etc. Table I provides a succinct qualitative comparison of the
proposed methods with the most relevant ones from the - Knowledge representation – The spectral kurtosis and the
literature. kurtogram preserve context-relevant features of the
vibration signal and represent discernible visual data
The methods presented in this paper provide the suitable for 2D input into the deep learning classification
noninvasive approach for detecting the potential issue of the network, which yields remarkably better classification
CB. The main idea elaborated in the paper is that most of the performance than a broad range of other vibration signal
anomalies and issues will reflect in the form of the prolonged representation modes. This knowledge represents the base
or delayed main contacts closing time of the CB. In for further investigation and expansion of the proposed
contemporary industrial applications, to obtain the main algorithm in terms of specific issue diagnostic and
contact closing time of the CB, two approaches can be localization.
utilized. The first one includes deenergizing CB and applying
constant test current throughout the contact measuring voltage The paper is organized as follows: Section II provides
drop. The second approach is to record the current change of insight into the origin of the CB vibrations as well as the nature
the current transformer secondary winding and compute the of the signal itself. Section III describes the theoretical
main contact closing time using the obtained signals. background, along with the description of two proposed
However, the first method requires downtime while the latter methods. To confirm the proposed algorithm's effectiveness
method depends on the availability of the current transformer and performance on the real object under test (OUT),
of the CB. Both methods are invasive and require interfering experimental results are provided in Section IV along with the
with the normal operation of the CB. To this end, the proposed details about the test setup as well as performance comparison
method does not interfere with the normal operation of the CB with the relevant concurrent methods given in section V. The
while it detects any time deviation of the CB timing or motion limitations of the proposed methods are presented in section
parameters. Furthermore, the proposed methods are not VI. The concluding remarks are given in section VII.
oriented toward a specific issue like most of the other methods II. THE ANATOMY OF THE CB VIBRATION SIGNATURE
available in the literature. The main goal of the proposed
method is primarily to detect the fault state of the CB without Due to the electromechanical nature of the CB and its
necessarily isolating the type of fault or its location. In other structure consisting of interrupting unit, operating mechanism
words, we are interested in the assessment of the manifestation and linkage, a series of superimposed percussion waves are
of the issue within CB, rather than in diagnosing or generated during the process of the energy transfer [31]. An
pinpointing its specific component or subsystem that has example of the vibration signature along with the time instant
undergone failure. of the main contact engagement for the closing operation and
the motion of the moving contacts of the de-energized CB is
The research highlights are listed as follows: given in Fig. 1.
- Practicality and application – By providing spectral
kurtosis and kurtogram as a foundation of the
classification features, they may be used as a vanguard for
training the classifier for the tested but possibly for other
types of CB. Furthermore, the proposed algorithm does not
address the origin of the issue but focuses on the outcome
resulting in the deviation of the CB normal operation
which yields generality. Such an approach substantially
facilitates noninvasive monitoring of the CB.
- High classification accuracy – The proposed methods are
capable of detecting the CB with a potential issue with an
accuracy of nearly 100%, as presented in the experimental
results section. The obtained models substantially Fig. 1 - (a) Main contact closing operation, motion of the main
outperform the state-of-the-art classifiers in terms of contacts and (b) vibration signature
Authorized licensed use limited to: UNIVERSITY OF AGDER. Downloaded on June 07,2023 at 08:35:13 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Power Delivery. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TPWRD.2023.3248157
Clearly, the vibration events are concentrated at the time represents the distance traveled by the contacts from the initial
of contact closing. However, the start of the operating starting position to the final resting position. The main contact
mechanism engagement as well as the linkage movement closing time parameters reflect the state of the control circuit
produces a portion of the vibration series superimposed before and the contact system while the motion stroke reflects the
the contacts closing event. Another series of vibration events condition of the operating mechanism and mechanical
are notable after the contacts are closed. Operation of the CB linkages. Since those parameters are defined by the IEC
(making and breaking main or arcing contacts) yields high- 62271-100:2021 standard and are well established in the
speed movement and high-intensity impacts resulting in the industry for detecting potential issues in offline testing
acceleration from zero to several meters per second in tens of methods, combined with the normal operation limits provided
milliseconds [1]. This makes the extraction of the exact by the manufacturer of the CB, the CB can be classified as a
closing time as defined by the standard, as well as changes in “Functional“ (F) or “Potentially defective“ (PD) in Stage 1.3
the total motion of the moving contacts (stroke), almost [32]. Stage 2.1 and Stage 2.2 differ depending on the method.
impossible.
To address the frequency content of the vibration signal,
Fig. 2 shows the power spectrum of the example of the healthy
CB where the closing time and stroke parameters are within
the limits defined by the manufacturer, and the CB with the
potential issue where the closing time parameter was outside
of the defined limits. It is evident that neither in the time nor
frequency domain there is no visible characteristic or pattern
that might differentiate the vibration fingerprint of healthy CB
from the one with a potential issue. The graphs presented in
Fig. 2 serve as a simple illustration that justifies our
motivation for digging deeper within different domains and
choosing more subtle classification features.
Authorized licensed use limited to: UNIVERSITY OF AGDER. Downloaded on June 07,2023 at 08:35:13 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Power Delivery. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TPWRD.2023.3248157
Here, ℎ(𝑡) is the impulse response resulting from a single classification, the Minimum Redundancy Maximum
impact where {𝑿𝑘 }𝑘∈ℤ and {𝝉𝑘 }𝑘∈ℤ are sequences of a Relevance (MRMR) algorithm has been used.
random variable that account for possibly random amplitudes
The MRMR algorithm finds an optimal set of features that
and random occurrences of the impact, respectively. The
is mutually and maximally dissimilar and can represent the
𝑁(𝑡) represents stationary additive noise. The stochastic
response variable effectively. The core of the algorithm is
process 𝑋(𝑡) can be decomposed into:
minimizing the redundancy of a feature set and maximizing
1 the relevance of a feature set to the response variable.
2
𝑋(𝑡) = ∫ 𝐻(𝑡, 𝑓)𝑒 𝑗2𝜋𝑓𝑡 𝑑𝑍𝑥 (𝑓), (2) The goal of the MRMR algorithm is to find an optimal set
1
S of features that maximizes 𝑉𝑠 , the relevance of S with respect
−
2 to a response variable y, and minimizes 𝑊𝑠 , the redundancy of
where 𝑑𝑍𝑥 (𝑓) is spectral increment and 𝐻(𝑡, 𝑓) is the S, where 𝑉𝑠 and 𝑊𝑠 are defined with mutual information I:
complex envelope of 𝑋(𝑡) at frequency f. This representation
is also known as the Wold-Cramer representation [34].
1 1
According to this representation (Eq.2), the SK is defined as a 𝑉𝑠 = ∑ 𝐼(𝑥, 𝑦), 𝑊𝑠 = 2 ∑ 𝐼(𝑥, 𝑧). (5)
fourth-order cumulant: |𝑆| |𝑆|
𝑥∈𝑆 𝑥,𝑧∈𝑆
Authorized licensed use limited to: UNIVERSITY OF AGDER. Downloaded on June 07,2023 at 08:35:13 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Power Delivery. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TPWRD.2023.3248157
SK as a function of frequency f and STFT window length 𝑁𝑤 preprocessing stage. Finally, the preprocessed kurtogram
is called kurtogram. However, to yield the “true“ central images are resized to 256x256 pixels using the lanczos3
frequency and bandwidth, all possible widths should be interpolation kernel. The information loss during the
enumerated which makes the method computationally very resampling process is inevitable, however, we have not
intense. As an improvement, the Fast kurtogram (FK) of the observed any substantial performance degradation of the
complexity 𝑁𝑙𝑜𝑔𝑁 based on an arborescent multi-rate filter classifier when resized kurtograms are used as inputs. The
bank structure has been proposed [41]. CNN has been fed with pre-processed kurtograms along with
the corresponding labeled “class“, thus following a typical
workflow of supervised learning.
Fig. 6 - The kurtogram example for the (a) “Functional” and (b)
“Potentially defective” real CB vibration fingerprint
Authorized licensed use limited to: UNIVERSITY OF AGDER. Downloaded on June 07,2023 at 08:35:13 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Power Delivery. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TPWRD.2023.3248157
(FC) layers. The flattened and FC layers represent the second calculated parameters and corresponding vibration
stage of the CNN. The flattened layer converts feature maps fingerprints. For experimental OUT, statistically obtained
to a one-dimensional (1D) vector. 1D feature map is acceptable limit, using 6% deviation from the average main
forwarded to an FC layer with 128 units, then transferred to contact closing time and 4% of the average motion stroke for
the second FC layer that produces one binary output (1 or 0) the first 100 operations of the CB before it became worn out,
since the Sigmoid activation function is used in the last layer. are [58, 61] ms and [48, 51] °, respectively. The results within
Other layers use Rectified Linear Unit (ReLu) activation the defined limits are labeled with class F. Any deviation
function since it is commonly used in CNN due to its faster outside of this parameter range yields labeling with class PD.
operation. It is worth noting that the dropout layer is used The obtained result set is imbalanced with the higher number
between the flattened and FC layer to prevent overfitting. As of results classified as F (91.69 %) compared to the ones
a loss function, binary cross-entropy is chosen. In essence, it labeled as PD (8.31 %). As expected, most of the PD labels
is suitable due to discrete classifier output. Furthermore, since come toward the end of the test before the CB failure.
the loss is logarithmic in nature, it yields a large loss for large
differences close to 1 and a small loss for small differences B. Results – Method BT
tending to 0. Moreover, it can compensate for any exponential In order to obtain the best possible performance with the
behavior that is often given through activation units like selected indices as well as to yield the most suitable machine
Sigmoid or SoftMax [43], [44]. Finally, the Adam optimizer learning classification algorithm for this task, a number of
is used in this network due to its efficient way of changing the learning attempts (different classifiers and different subsets of
learning rate during the training. All computations, design and features) have been conducted using the full set of 16 features.
training have been performed using MathWorks MATLAB The trained model has been validated using 5-fold cross-
2022a, specifically Neural Network Toolbox and Deep validation and tested with the 5% of dedicated data.
Learning Toolbox for the Method BT. For the Method CNNK, TABLE II
the Python libraries Keras 2.4.0, scikit-learn 1.0.0., NumPy, RESULTS OF CLASSIFICATION WITH FULL 16-FEATURES DATASET
Pandas and matplotlib.pyplot 3.5.0. have been used. Acc Sens Spec F1
Algorithm MCC
[%] [%] [%] score
IV. EXPERIMENTAL VALIDATION
Fine Tree 92.72 68.56 93.35 33.66 36.61
To validate the proposed methods, the test has been Medium Tree 91.93 65.22 92.06 10.67 18.75
conducted on real CB and vibration fingerprints have been Coarse Tree 91.61 33.33 91.66 01.53 5.43
recorded during its operation. The experiment consisted of the Gaussian Naive Bayes 72.50 18.40 96.00 01.50 7.81
brand new CB which has been put under test with a large
Kernel Naive Bayes 80.09 21.15 94.86 28.45 24.36
number of operations during which parameters such as main
Linear SVM 91.63 50.00 91.64 - -
contact closing and opening time, coil current, main contact
motion, auxiliary channels as well as vibration fingerprint in Quadratic SVM 91.62 50.00 91.64 - -
1 position on the chassis of the CB have been recorded. Cubic SVM 74.12 10.54 92.26 01.40 7.41
Fine Gaussian SVM 94.80 87.71 95.07 63.20 64.25
A. Test setup and Data Acquisition Medium Gaussian SVM 91.66 50.00 91.64 - -
The OUT used for the experimental validation of the Coarse gaussian SVM 91.69 50.00 91.64 - -
proposed algorithm has been de-energized and not grounded Fine KNN 95.95 74.17 98.00 76.53 74.36
ABB SACE 2 EMAC CB family member CB. One PR MEMS Medium KNN 93.25 61.05 95.60 54.08 50.48
(piezoresistive micro electromechanical system) mono-axial Coarse KNN 91.64 13.64 91.64 02.29 7.44
accelerometer is used for recording the vibration fingerprint –
Cosine KNN 93.13 60.53 95.61 53.81 50.12
mounted on the CB chassis. The used Brüel & Kjær type 4393
Cubic KNN 93.10 60.45 95.57 53.45 49.84
accelerometer has a frequency range of [0.1 16500] 𝐻𝑧 and
Weighted KNN 96.77 81.70 98.00 79.89 78.14
sensitivity 0.316 ± 2% 𝑝𝐶/(𝑚/𝑠 2 ). The vibration signal as
well as main contact voltage and motion trajectory has been Boosted Trees 91.75 92.30 91.70 03.20 11.23
acquired using Brüel & Kjær data acquisition system and Bagged Trees 97.80 96.38 97.87 85.50 85.10
Pulse LABSHOP software suite. In the available literature, Subspace KNN 97.43 90.55 97.96 83.02 81.85
there are several approaches available for mounting the RUSBoosted Trees 67.25 18.44 98.00 30.50 29.22
vibration sensor and the most common are screws [24], glue Linear Discriminant 91.22 28.75 91.38 23.42 18.45
[45], wax and magnets [46]. Since the intention of this Quadratic Discriminant 81.42 21.21 94.42 - -
experimental setup is a large number of operations spread Logistic Regression 91.69 50.00 91.64 - -
throughout several weeks, mounting with the beeswax is not
appropriate. Mounting using strong magnets reduces the
frequency bandwidth of the recorded vibration pattern. The ratio of datasets for training/validation/test has been
Accordingly, the mounting has been performed using strong designated to 76% / 19% / 5% of the complete dataset,
rigid glue. The experiment was initiated on the brand new respectively. Using this specific data allocation for training,
OUT until its failure. The signals required for the calculation validation and testing, we have preserved the ratio of the initial
of the main contact closing time, motion stroke, rebound and dataset of 91.69% for F and 8.31% for the PD class. Most of
overtravel are acquired at the rate of 20 kHz while vibration the tested machine learning classification algorithms had a
signals are sampled at the rate of 33 kHz. The selected promising performance. Performance of the classifiers
sampling rates are in accordance with the ones used in the presented with F1 and MCC in the following tables are
available research literature as well as in industrial practice. A multiplied by 100 for a sake of clear presentation. However,
total of 35473 complete cycles of the de-energized CB the best results in terms of accuracy have been achieved with
consisted of open operation, charge spring, and close Ensemble Bagged Trees (97.80%) as well as the highest F1
operation. All closing operations are extracted along with the score (85.50) and Mathew Correlation Coefficient (85.10).
Authorized licensed use limited to: UNIVERSITY OF AGDER. Downloaded on June 07,2023 at 08:35:13 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Power Delivery. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TPWRD.2023.3248157
TABLE V
THE CONFUSION MATRIX FOR THE TRAINED CNN MODEL
False 998 2
Actual value
True 0 1000
False True
Predicted value
Authorized licensed use limited to: UNIVERSITY OF AGDER. Downloaded on June 07,2023 at 08:35:13 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Power Delivery. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TPWRD.2023.3248157
V. COMPARISON WITH OTHER METHODS algorithms [17], [13]. Along with these differences, the
An additional step in the proposed method validation is the concept can be found in a vast range of available research
comparison with the selected methods provided in the papers with the difference in the decomposition approach in
literature. Since there is a wide variety of methods and form of LMD, EMD, EEMD, VMD, etc. Since the method
approaches described in the published papers, the benchmark [47] assumes multiclass output, the comparison is performed
comparison will be focused on the 2 most related methods. such that all three faulty classes are treated as a PD class. The
The training of the machine learning classification algorithms second method (denoted as Reference method 2, Table I, No.
proposed in both methods is performed using the same 2) used for a comparison with our method is based on the one-
vibration signatures data set as in Section IV containing 1000 dimensional convolution neural network (1D-CNN) enriched
examples of PD and 1000 examples of F class. Additionally, with attention mechanism and capsule layer [28]. The original
the proposed methods are tested using the variable number of unprocessed vibration signal is used as the input signal into
instances from each class. The comparison of the obtained the network. The authors state, considering the network
results is presented in Table VI. The first method (denoted as consisted of two convolution layers and one capsule layer, that
Reference method 1, Table I, No. 1) to be compared is the one the main benefit is the independence of a number of samples
based on the EEMD, where the first 4 IMFs are used to obtain while keeping the network architecture simple. The
the correlation dimension using G-P algorithm [47]. The performances of the training of the machine learning
obtained 4 correlation dimensions are used as a feature vector classification algorithms based on 1000 F and 1000 PD,
for the SVM machine learning algorithm training. A similar provides the following results. For Reference method 1.,
approach regarding the feature extraction stage is persuaded obtained an accuracy of 95.49 %, a sensitivity of 95.49 % and
by other authors with the main difference in classification a specificity of 91.40 % of the model using the proposed
feature vector underperforms the classification performance
TABLE VI
EVALUATION OF PROPOSED MODELS AND COMPARISON WITH CONCURENT ALGORITHMS
Class Method BT Method CNNK Reference Method 1 [47] Reference Method 2 [28]
Sens F1 Sens F1 Sens F1 Sens F1
Acc [%] Score Acc [%] Score Acc [%] Score Acc [%] Score
F PD
[%] Spec [%] Spec [%] Spec [%] Spec
MCC MCC MCC MCC
[%] [%] [%] [%]
100 88.20 99.72 99.86 97.11 85.38 95.79 94.43
351 351 98.61 99.83 97.33 97.91
98.22 88.17 100 99.57 96.01 82.96 92.22 88.66
100 87.76 99.72 99.86 96.99 87.31 95.73 95.71
175 98.55 99.81 96.40 94.30
97.77 87.06 100 99.57 96.11 86.21 91.43 87.21
100 87.76 99.72 99.86 93.67 81.23 95.73 96.75
87 98.63 99.77 95.90 94.75
98.32 87.06 100 99.30 96.38 86.23 90.8 84.18
100 85.22 99.72 99.86 87.50 74.43 95.73 97.33
43 99.00 99.75 96.70 95.17
98.87 85.11 100 98.72 97.74 84.47 90.69 78.40
100 85.21 99.72 99.86 78.26 79.91 95.73 97.53
21 99.51 99.73 97.80 95.43
99.43 85.11 100 97.56 99.14 81.11 90.48 69.00
100 80.01 99.72 99.86 61.54 79.91 95.73 97.71
351
Authorized licensed use limited to: UNIVERSITY OF AGDER. Downloaded on June 07,2023 at 08:35:13 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Power Delivery. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TPWRD.2023.3248157
obtained using Method CNNK. Furthermore, Reference real-world field environment considering multiple various
method 2 performance accuracy of 99.67 %, underperforms CBs. The second open problem is to test the algorithm
the accuracy of 99.98% of our proposed algorithm as well. dependency on the vibration sensor mounting approach. The
Additionally, the results of the performances for Reference third aspect worth investigating is whether kurtogram captures
methods 1 and 2 with the different levels of imbalance are the information on specific issues (loosed bolt, jammed or
given in Table VI as well. partially jammed dumper, etc.). This approach may provide
the means to finer-grained diagnostics, rather than a binary-
classification-based one. Finally, this approach should be
considered for other types of CBs covering different voltage
levels, different contact system constructions and as well as
operating mechanisms.
REFERENCES
Authorized licensed use limited to: UNIVERSITY OF AGDER. Downloaded on June 07,2023 at 08:35:13 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Power Delivery. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TPWRD.2023.3248157
10
[16] Q. Yang, J. Ruan, Z. Zhuang, D. Huang and Z. Qiu, "A New Vibration [35] M. L. Bouhalais and M. Nouioua, "The analysis of tool vibration
Analysis Approach for Detecting Mechanical Anomalies on Power signals by spectral kurtosis and ICEEMDAN modes energy for insert
Circuit Breakers," IEEE Access, vol. 7, pp. 14070 - 14080, 2019. wear monitoring in turning operation," International Journal of
[17] Z. Jianfeng, L. Mingliang, W. Keqi, X. Jingyan and S. Shuli, "Feature Advanced Manufacturing Technology, vol. 115, no. 9-10, pp. 2989-
extraction of HV circuit breaker based on ensemble empirical mode 3001, 2021.
decomposition and correlation dimension," in Electronic [36] X. Ni, J. Zhao, Q. Hu, X. Zhang and H. Li, "A new improved
Measurement & Instruments (ICEMI), 2015 12th IEEE International Kurtogram and its application to planetary gearbox degradation
Conference on, 2015. feature analysis," Journal of Vibroengineering, vol. 19, no. 5, pp.
[18] Q. Yang, J. Ruan, Z. Zhuang and D. Huang, "Chaotic Analysis and 3413-3428, 2017.
Feature Extraction of Vibration Signals from Power Circuit Breakers," [37] F. Combet and L. Gelman, "Optimal filtering of gear signals for early
IEEE Transactions on Power Delivery, vol. 35, no. 3, pp. 1124-1135, damage detection based on the spectral kurtosis," Mechanical Systems
2019. and Signal Processing, vol. 23, no. 3, pp. 652-668, 2009.
[19] N. Huang, H. Chen, S. Zhang, H. Cai, W. Li and D. Xu, "Mechanical [38] T. Barszcz and R. B. Randall, "Application of spectral kurtosis for
fault diagnosis of high voltage circuit breakers based on wavelet time- detection of a tooth crack in the planetary gear of a wind turbine,"
frequency entropy and one-class support vector machine," Entropy, Mechanical Systems and Signal Processing, vol. 23, no. 4, pp. 1352-
vol. 18, no. 1, pp. 1-17, 2016. 1365, 2009.
[20] Y. Zhang, Y. Shao, W. Gong, Z. Zhang and J. Wu, "Research on [39] C. Liu, X. Gao, D. Chi, Y. He, M. Liang and H. Wang, "On-line
combined diagnosis of mechanical fault vibration-sound signal of high chatter detection in milling using fast kurtogram and frequency band
voltage circuit breaker based on EEMD-energy entropy feature," in power," European Journal of Mechanics, A/Solids, vol. 90, 2021.
International Conference on Sensors and Instruments 2021, Qingdao, [40] J. Antoni and R. Randall, "The spectral kurtosis: Application to the
China, 2021. vibratory surveillance and diagnostics of rotating machines,"
[21] M. Runde, T. Aurud, L. Lungaard, G. Ottesen and K. Faugstad, Mechanical Systems and Signal Processing, vol. 20, no. 2, pp. 308-
"Acoustic diagnosis of high voltage circuit-breakers," IEEE 331, 2006.
Transactions on Power Delivery, vol. 7, no. 3, pp. 1306-1315, 1992. [41] J. Antoni, "Fast computation of the kurtogram for the detection of
[22] M. Ermidoro and F. Previdi, "A condition assessment algorithm based transient faults," Mechanical Systems and Signal Processing, vol. 21,
on dynamic time warping for high-voltage circuit breaker," in IEEE no. 1, pp. 108-124, 2007.
International Conference on Control and Automation (ICCA), [42] P. H. Nguyen and J.-M. Kim, "Multifault diagnosis of rolling element
Katmandu, Nepal, 2016. bearings using a wavelet kurtogram and vector median-based feature
[23] L. Fasanotti, S. Cavalieri, M. Tomasini, F. Floreani and S. Ierace, analysis," Shock and Vibration, vol. 2015, no. 320508, pp. 0-14, 2015.
"Prognostics algorithms for circuit breaker application: A benchmark [43] A. Ruby, P. Theerthagiri, J. Jacob and Y. Vamsidhar, "Binary cross
analysis," in International Conference on Industrial Informatics, entropy with deep learning technique for Image classification,"
INDIN 2014, 2014. International Journal of Advanced Trends in Computer Science and
[24] M. Landry, F. Léonard, C. Landry, R. Beauchemin, O. Turcotte and Engineering, vol. 9, no. 4, pp. 5393-5397, 2020.
F. Brikci, "An improved vibration analysis algorithm as a diagnostic [44] I. Goodfellow, Y. Bengio and A. Courville, Deep learning, Illustrated
tool for detecting mechanical anomalies on power circuit breakers," edition ed., London: The MIT Press, 2016.
IEEE Transactions on Power Delivery, vol. 23, no. 4, pp. 1986-1994,
2008. [45] N. Charbkeaw, T. Suwanasri and T. Bunyagul, "Mechanical Defect
Detection of SF6 High Voltage Circuit Breaker Using Wavelet Based
[25] L. Li, J. Xiao, B. Wu, M. Zhou and Q. Wang, "Online monitoring and Vibration Signal Analysis," in 5th International Conference on
diagnosis of high voltage circuit breaker faults : feature extraction Electrical EngineeringElectronics Computer Telecommunications
analysis of vibration signals," International Journal of Metrology and and Information Technology, 2008.
Quality Engineering, vol. 10, no. 13, 2019.
[46] A. Brandt, Noise and Vibration Analysis - Signal Analysis and
[26] Y. Zhang, Y. Jiang, Y. Chen and Y. Zhang, "Fault Diagnosis of High Experimental Procedures, West Sussex, United Kingdom: John Wiley
Voltage Circuit Breaker Based on Multi-classification Relevance & Sons Ltd, 2011.
Vector Machine," Journal of Electrical Engineering and Technology,
vol. 15, no. 1, pp. 413-420, 2019. [47] P. Yang, M. Liu, Q. Peng, Y. Xing and B. Li, "A improved way for
fault diagnose based on EEMD and SVM," in 29th Chinese Control
[27] M. Wong, L. Jack and A. Nandi, "Modified self-organising map for And Decision Conference (CCDC), Chongqing, China, 2017.
automated novelty detection applied to vibration signal monitoring,"
Mechanical Systems and Signal Processing, vol. 20, no. 3, pp. 593- [48] X. Wang, J. Ruan, T. Zhou, X. Peng, Y. Deng and Q. Yang, "Data
610, 2006. Mining in the Vibration Signal of the Trip Mechanism in Circuit
Breakers Based on VMD-PSR," Electronics, vol. 11, no. 22, 2022.
[28] X. Ye, J. Yan, Y. Wang, L. Lu and R. He, "A Novel Capsule
Convolutional Neural Network with Attention Mechanism for High- [49] L. Mingliang, W. Keqi, S. Laijun and Z. Jianfeng, "A New Fault
Voltage Circuit Breaker Fault Diagnosis," Electric Power Systems Diagnosis Method for High Voltage Circuit Breakers Based on
Research, vol. 209, no. 108003, 2022. Wavelet Packet and Radical Basis Function Neural Network," Open
Cybernetics & Systemics Journal, vol. 8, pp. 410-417, 2014.
[29] D. Jiang, S. Zhao, J. Ren and Y. Xu, "Extreme learning machine in the
breaker fault diagnosis," Advanced Materials Research, Vols. 960- [50] J. Ruan, X. Wang, T. Zhou, X. Peng, Y. Deng and Q. Yang, "Fault
961, pp. 896-899, 2014. identification of high voltage circuit breaker trip mechanism based on
PSR and SVM," IET Gener. Transm. Distrib., pp. 1-11, 2022.
[30] S. Ma, B. Jia, J. Wu, Y. Yuan, Y. Jiang and W. Li, "Multi-vibration
information fusion for detection of HVCB faults using CART and D–
S evidence theory," ISA Transactions, vol. 113, pp. 210-221, 2021.
[31] M. Kapetanovic, High Voltage Circuit Breakers, Sarajevo: Faculty of
Electrical Engineering Sarajevo, 2011.
[32] IEC, "IEC 62271-100 High-voltage switchgear and controlgear – Part
100: Alternating - current circuit-breakers," International
Electrotechnical Comission (IEC), 2021.
[33] D. Wang, P. W. Tse and K. L. Tsui, "An enhanced Kurtogram method
for fault diagnosis of rolling element bearings," Mechanical Systems
and Signal Processing, vol. 35, no. 1-2, pp. 176-199, 2013.
[34] J. Antoni, "The spectral kurtosis: A useful tool for characterising non-
stationary signals," Mechanical Systems and Signal Processing, vol.
20, no. 2, pp. 282-307, 2006.
Authorized licensed use limited to: UNIVERSITY OF AGDER. Downloaded on June 07,2023 at 08:35:13 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.