0% found this document useful (0 votes)
29 views6 pages

Anomaly Detection On Industrial Electrical Systems Using Deep Learning

This document discusses using deep learning techniques for anomaly detection on industrial electrical systems. Specifically, it proposes using convolutional autoencoders to detect faults in industrial plants by analyzing current intensity data from one of the three phases, with the goal of minimizing false positives. The results showed this methodology was highly capable of detecting faults while generating few false alarms.

Uploaded by

davidtop666888
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
29 views6 pages

Anomaly Detection On Industrial Electrical Systems Using Deep Learning

This document discusses using deep learning techniques for anomaly detection on industrial electrical systems. Specifically, it proposes using convolutional autoencoders to detect faults in industrial plants by analyzing current intensity data from one of the three phases, with the goal of minimizing false positives. The results showed this methodology was highly capable of detecting faults while generating few false alarms.

Uploaded by

davidtop666888
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 6

This full text paper was peer-reviewed at the direction of IEEE Instrumentation and Measurement Society prior to the

acceptance and publication.

Anomaly Detection on Industrial Electrical Systems


using Deep Learning
Marco Carratù, Vincenzo Gallo, Antonio Pietrosanto, Paolo Gabriele Patrizi, Alessandro Bartolini, Lorenzo Ciani,
Sommella Marcantonio Catelani, Francesco Grasso
Department of Industrial Engineering Department of Information Engineering
University of Salerno University of Florence
Via Giovanni Paolo II, 132 Fisciano (SA), Italy Via di S. Marta, 3, 50139, Florence, Italy
{mcarratu, vgallo, apietrosanto, psommella}@unisa.it {gabriele.patrizi, a.bartolini, lorenzo.ciani marcantonio.catelani,
francesco.grasso}@unifi.it
2023 IEEE International Instrumentation and Measurement Technology Conference (I2MTC) | 978-1-6654-5383-7/23/$31.00 ©2023 IEEE | DOI: 10.1109/I2MTC53148.2023.10175908

Abstract— The recent development and spread of artificial variability of data that can be used for anomaly detection
intelligence-based techniques, particularly deep learning [8],[9]. In fact, anomaly detection can be applied to data with
algorithms, have made it possible to model phenomena that already extracted features, to univariate and multivariate time
were previously impossible to handle. Furthermore, the series, and images [10].
development of the Big Data paradigm is rapidly leading toward
The anomaly can be defined as an outlier, an observation
new research frontiers in predicting and classifying one-
dimensional signals. Anomaly detection plays a crucial role in that deviates greatly from other observations, leading to
the various areas that gain from the introduction of these suspicion about the mechanism that generated it.
methodologies. This extremely diverse field detects anomalies in There are mainly three types of anomalies: point
both time series and image data. Anomaly detection applications anomalies, contextual anomalies, and collective anomalies.
include the detection of failures of grid-connected machinery in Point anomalies are global outliers, isolated points that are far
industrial environments. The objective of this study was to removed from the surrounding trend. On the other hand,
propose a fault detection methodology based on deep learning, contextual anomalies are a series of non-continuous
specifically using convolutional autoencoders, using as few observations that are distant from the surrounding trend.
features as possible, specifically the current intensity in one of
Finally, collective anomalies are a series of continuous
the three phases of an industrial plant. The results showed a
high capability of the methodology to detect faults while observations that differ from the rest of the dataset. However,
generating a minimum number of false positives, paving the way individual points in this series do not constitute anomalies
for optimizations of the same and online deployment. [11]. An example of an anomaly is shown in Fig. 1
Anomaly detection in time series finds great applicability
Keywords— Anomaly detection; Current measurement; in many areas, including industrial power grid monitoring.
Deep learning; Fault detection; Industrial power systems. In this sector, anomaly detection is particularly important
since early identification of a malfunction in an electrical
I. INTRODUCTION system can be crucial to avoid economic and personal damage.
The increasingly strong development of low-cost sensors Anomaly detection is routinely employed in the
has enabled an increasingly widespread deployment of application domain of smart grids, particularly to monitor
measuring points over the past decade. anomalous behaviors, such as unusual consumption patterns,
This diffusion has had as an immediate consequence the faulty grid infrastructures, outages, cyber-attacks, and energy
generation of unprecedented amounts of data; this has led, in fraud [12]-[18].
addition to the problem of data management and storage, to Methodologies based on SVM (Support Vector Machine),
difficulty in interpreting the data, thus making it difficult to Bayesian Networks, KNN (k-nearest neighbors), and other
extract important features for analysis [1]. machine learning techniques have been employed to monitor
However, these large amounts of data, referred to as Big anomalies in electricity consumption in cloud computing, as
Data [2], also has great potential. These, in fact, can lead to described in [19].
identifying and describing known patterns with smaller
amounts of data. It isn’t easy to create a model with traditional
mathematical approaches using Big Data because of the
sometimes too complex relationships between the variables
involved.
In this context, new artificial intelligence techniques,
particularly Deep Learning algorithms, have recently made it
possible to develop extremely complex and accurate models
of the reality around us [3].
Deep Learning algorithms, particularly Neural Networks,
have recently had great use for data and image analysis,
performing feature extraction for classification, identification,
segmentation, and prediction tasks [4]-[7].
One of the areas where traditional algorithms have Fig. 1. Example of anomaly detection on Time Series [20]
encountered the most difficulty in use is anomaly detection.
This is due to the vastness of the field and, thus, the great

Authorized licensed use limited to: Northeastern University. Downloaded on August 10,2023 at 06:01:32 UTC from IEEE Xplore. Restrictions apply.
All of these methodologies are based on multivariate time A. Autoencoder
series analysis. Autoencoders are special types of neural networks capable
Recently, research has also focused on monitoring of encoding unlabeled datasets, making them belong to the
individual industrial plants, in the context of Industry 4.0, class of unsupervised learning. An example of Autoencoder
outside the dispatching logic inherent in smart grids, but solely architecture is shown in Fig.2, which highlights the three basic
aimed at identifying possible malfunctions in the machinery components of autoencoders:
and equipment used.
In this field, anomalies in High-Voltage Transformer 1. Encoder: A component that shrinks the input data
Bushings have been analyzed using Deep Learning-based from the train-validate-test set into an encoded form
methodologies [21], and the anomaly in the operation of an that is often many orders of magnitude less.
industrial gas turbine, again using the same family of
algorithms [22]. 2. Bottleneck: A module that is the most crucial network
Among the various deep learning-based techniques component since it includes the compressed
employed, those that have been most successful are those knowledge representations.
based on autoencoders. Specifically, in [23], the authors 3. Decoder: A component that aids the network in
applied autoencoders for feature extraction of anomalies in "decompressing" knowledge representations and
electricity usage behavior in residential homes. recovering the data from its encoded state. Next, the
On the other hand, in [24],[25], the authors proposed a output is contrasted with a source of truth.
smart metering system based again on autoencoders to
evaluate anomalous electricity consumption in order to reduce
costs for the company. Thus, autoencoders can represent data by reducing its
A similar goal has been pursued in [26], where the authors dimensionality and rejecting noise.
proposed an anomaly detection system based on multivariate Deep Autoencoders, Variational Autoencoders, and
time series analyzed with a sliding window. convolutional autoencoders are among the different types of
Another system has been applied in [27] to reduce autoencoders. Variational autoencoders perform a latent space
electricity consumption costs by analyzing the power profiles representation in order to generate data, following a
of an electrical system, including photovoltaic panels, thus probabilistic model. The purpose of these variational
trying to optimize self-generation and self-consumption. autoencoders, although similar in architecture to traditional
In contrast, this work aims to verify the performance of an autoencoders, is quite different from the aim of this paper.
Autoencoder to check for malfunctions in a three-phase Deep autoencoders are able to learn much more complex
industrial electrical system. In fact, in this field, it is optional features because of the two symmetric deep-belief networks.
to apply the methodologies used for smart power grids, mainly The main use of deep autoencoders is for data compression,
concerned with detecting collective anomalies, which are especially for images. Their use, therefore, is not within the
subgroups of contiguous observations that deviate from the
scope of this work. Convolutional autoencoders, on the other
rest of the dataset. In contrast, the anomalies under
hand, are characterized by a lighter architecture. Thanks to the
investigation concern industrial machinery malfunctions and
thus point to anomalies or contextual anomalies. development of one-dimensional CNN (Convolutional Neural
The innovative aspect of the work is the use of a single Network), they can be used for images and time series. Those
time sequence, thus performing a univariate anomaly are also used as signal reconstructors, making them
detection analysis on the intensity of the current flowing particularly suitable for identifying outliers or anomalies in
through one of the three phases of the electrical system under the signal, as these deviate from the normal data distribution
consideration. Moreover, the proposed method is totally that the autoencoder seeks to reproduce.
unsupervised, avoiding the need for time-consuming and The strength of Convolutional Autoencoders is their ease
of use: it is not necessary to know the anomaly pattern to be
costly data labeling for training.
investigated in advance, but it is only required training on an
The final aim is to detect possible malfunctions in
unfaulty time series.
industrial machinery connected to the power grid, mainly
composed of ohmic-inductive impedance loads. B. Proposed Algorithm
As stated in the previous section, this work has aimed to
identify malfunctions in the ohmic-inductive loads present in
II. PROPOSED METHODOLOGY
The proposed methodology is based on the use of an
Autoencoder based on one-dimensional convolutional layers.
The proposed approach aims to develop a completely
unsupervised method to detect anomalies in current intensity
in a three-phase industrial plant. The prompt identification of
such anomalies could prevent long abnormal behavior and
critical malfunctioning of machinery.
The following subsections will first introduce the
autoencoders, then the proposed algorithm will be explained,
and finally, the neural network architecture and the dataset
employed in this work will be described.

Fig. 2. Generic architecture of an Autoencoder

Authorized licensed use limited to: Northeastern University. Downloaded on August 10,2023 at 06:01:32 UTC from IEEE Xplore. Restrictions apply.
an industrial plant. These malfunctions are exhibited as more A stride size of 2 samples, a kernel size of 7 samples, and
or less sudden current spikes related to mild voltage surges. ReLu activation levels have been used for all levels.
The most prominent features in the problem under The optimizer employed for training the neural network
consideration have been then identified on the currents of the has been Adam, with a learning rate equal to 0.001 and the
individual phases of the general three-phase electrical system. loss function MAE.
These peculiarities made the problem an excellent candidate
D. Dataset Preparation
for anomaly detection on univariate time series.
In light of state of the art, it has been decided to employ a A customized dedicated acquisition unit has been used to
methodology based on Deep Learning. In particular, a system monitor the electrical parameters of a big industrial
based on convolutional Autoencoder has been chosen to manufacturing plant [28]. Current measurements on the three-
replicate the nominal root mean squared current trends during phase system have been acquired using three Rogowsky coils.
the daily arc. In particular, the current sensors provide a voltage output with
In particular, the autoencoder has been employed as the a sensitivity of 1 mV/A. Furthermore, the transducers are
nominal current intensity reference, trained under non-fault characterized by 20kHz bandwidth and a maximum range of
conditions. 3kA in terms of the Root Mean Square (RMS) value of each
The autoencoder acts as a filter on signal outliers and phase.
anomalies. This way, it is possible to generate an alert of The dataset consisted of a single time sequence reporting
equipment malfunction on the power grid by evaluating the the root mean squared current intensity flowing through one
deviation between the measured and expected current of the three phases of the industrial plant analyzed. The
intensity. dataset consisted of one-month measurements acquired
The alert threshold is specific to each dataset, so for this continuously with a sampling time of 120 seconds.
application, it has been decided to employ the maximum Mean To improve the performance of the autoencoder Adam
Absolute Error (MAE), calculated as in (1), produced by the optimizer, it was necessary to normalize the data against the
neural network training. mean and standard deviation of the training, as defined in (2).

(2)
∑ | | 1
Consequently, the data used for testing have also been
Where n is the number of samples in the training window, Imeas normalized using the same mean and standard deviation
is the current measured on the phase, and IDAE is the expected parameters used for training to keep the prediction unaltered.
measurement reported by the Convolutional Autoencoder. The network also required a subdivision of data into
C. Network Architecture windows to enable the convolutional layers to operate
properly. To perform this operation, the data have been split
As stated earlier, the network architecture used is the
with a sliding window of a step of one sample and length equal
classical architecture of a Convolutional Autoencoder,
to the variable "Steps," also shown in Table I.
consisting of a first-part encoder and a second-part decoder.
An example of the normalized data is shown in Fig. 3
For the application, which is characterized by particularly
large datasets, it has been decided to use convolutional layers, III. RESULTS
characterized by the absence of state memory, making the
training and especially inference process faster. A. Network Training
The autoencoder has been structured with two The dataset, consisting of more than one million samples,
convolutional encoding layers and two one-dimensional was divided into training and testing. Two hundred thousand
convolutional decoding layers. A summary of the architecture points corresponding to unfaulty measurements were selected
employed is shown in Table I. for training.
The size of the autoencoder analysis window, i.e., the
TABLE I
AUTOENCODER ARCHITECTURE "Steps," has been defined as 5000 samples.

Layer Output Shape

Input Batch, Steps


1D Conv Batch, Steps/2, 32
Dropout Batch, Steps/2, 16
1D Conv Batch, Seps/4, 16

1D Transposed Conv Batch, Steps/2, 16


Dropout Batch, Steps/2, 16
1D Transposed Conv Batch, Steps, 32
1D Transposed Conv Batch, Steps/2, 1
Fig. 3. Example of normalized data

Authorized licensed use limited to: Northeastern University. Downloaded on August 10,2023 at 06:01:32 UTC from IEEE Xplore. Restrictions apply.
A validation split of 30 percent of the training dataset, a
batch size of 32 samples, and a number of epochs equal to 50
have been used to train the network.
Training has been carried out using a graphics accelerator,
an Nvidia RTX Quadro 8000 GPU, with CUDA support. The
entire data preprocessing, training, and testing phase has been
performed using the Keras and TensorFlow libraries.
The best validation loss value that characterized the
model was 0.02.
Following the training, a fundamental analysis was
performed for the threshold value decision on reporting any
anomalies present in the analyzed signal.
Specifically, the Mean Absolute Error distribution of the Fig. 5. Distribution of Mean Absolute Error for the testing dataset
entire training dataset was reported. This distribution is
shown in Fig.4. operating situations and temporally distant from the training
The distribution shows a preponderance of deviation of unfaulty acquisitions.
the convolutional autoencoder output from the training values However, the fact that the maximum MAE obtained in the
equal to an MAE of 0.2. test distribution is smaller than that of the training suggests
Given the nature of the anomalies present on the analyzed that the unfaulty points in the test dataset are not marked as
power grid, which deviate widely from the main time series, false positives by the autoencoder.
and in order to make the entire methodology totally Therefore, to verify this claim, it was decided to analyze
unsupervised, it was decided to use the maximum MAE value the performance of the test time series and to see the response
obtained from training as a threshold for anomaly detection of the proposed algorithm during the fault and non-fault steps.
in inference, thus generating an alert. The result of this analysis is shown in Fig.6.
A maximum MAE of 1.6 has been found for the dataset One of the first results that can be seen from the figure is
under consideration, as visible in Fig.4. that the proposed algorithm has succeeded both in detecting
current spikes due to the malfunction of some three-phase
B. Anomaly Detection motors in the electrical system under consideration and in
Another portion of the dataset consisting of avoiding marking as abnormal some small spikes due to
approximately one hundred and sixty thousand points has normal switching on of loads. In particular, these mini spikes
been employed to test the network. This portion of the dataset are located between samples 110,000 and 120,000.
has been chosen as a test because it contained ten anomalous However, it is necessary to focus on the first part of the
RMS current flow data. signal, where there is a particular concentration of abnormal
Before the test, the data were normalized according to (2) events. The focus is shown in Fig.7, where it is also possible
and used the same mean and standard deviation values to see a false positive, which occurred at sample 5000.
obtained in the training phase. Network inference has then Table II summarizes the results obtained in terms of
been run, obtaining a time series of expected RMS current for Precision, Recall and F1-score.
each time instant. The results show that the proposed methodology can
The distribution of MAE loss values was again calculated identify outlier samples even before the value significantly
for the entire test dataset to check the prediction's validity. differs from the surrounding time series. This makes this
The results of this processing in the form of distribution system particularly suitable for detecting isolated point
density are shown in Fig.5. anomalies and contextual ones. The result is important since
As can be seen from the figure, the distribution of it is based on unsupervised learning and uses a single-variate
different MAE values is slightly different and reports a lower time sequence, avoiding the need to provide the system with
maximum value than that used in the training phase. This is other features necessary to improve identification
mainly because this dataset was acquired in different performance.

Fig. 6. Normalized RMS Current (blue line) and anomalous samples


Fig. 4. Distribution of Mean Absolute Error for the training dataset detected (Red dot)

Authorized licensed use limited to: Northeastern University. Downloaded on August 10,2023 at 06:01:32 UTC from IEEE Xplore. Restrictions apply.
TABLE III
ANOMALY DETECTION RESULTS ON 10 FAULT EVENTS WITH DIFFERENT
THRESHOLDS

Baseload + Baseload + Baseload +


Parameter
100% 150% 200%

Precision 0.09 0.21 0.85

Recall 0.98 0.85 0.75

F1-Score 0.16 0.34 0.79

As can be seen, the anomaly is not detected because it is


Fig. 7. Normalized RMS Current (blue line) and anomalous samples
detected (Red dot)
of a smaller amplitude than the next average current draw,
which worsens the Recall value.
The proposed methodology succeeded in classifying all Considering the presented results, the proposed
ten actual anomalous current flow samples as anomalous, methodology has been shown to exceed state of the art both
reporting at least two anomalous points per fault. Only one in performance and in not needing to select an alert threshold
sample was reported as a false positive for the provided for anomaly detection manually.
dataset.
IV. CONCLUSIONS
C. Comparison with state of the art
This paper proposed an autoencoder-based fault detection
To evaluate the performance of the proposed methodology on the industrial power grid, i.e., an
methodology, a comparison was made with a baseline unsupervised training approach on single-variate time series.
technique, fixed thresholding. This technique allows
The approach is characterized by ease of application, no need
reporting anomaly points in a time series after setting a
for prior knowledge of the model, or the type of anomaly
specific threshold. The threshold is calculated after a study
manifested by the fault to be searched for.
period of the time series, evaluating the current draw The method was applied to a dataset consisting of the
baseload, and setting an alert threshold higher by a certain RMS current intensity data of one of the three phases of an
percentage. The disadvantages of this methodology are two:
industrial plant. Specifically, the proposed methodology
correctly identified all faults recorded on the power grid,
i. Need for manual tuning of the alert threshold.
reporting only one false positive.
ii. Missed detection in case of variable baseload.
Future developments will concern optimization of the
choice of alert threshold for fault detection and comparison
The first issue, strongly related to the nature of the dataset, with other types of autoencoders and machine learning
was analyzed by testing different thresholds for the dataset
methodologies applied on the case under consideration. The
under consideration. The performances obtained using
on-line implementability of the methodology in the industrial
different threshold levels are presented in Table III.
plant will also be evaluated.
As expected, results show that as the threshold increases,
the value of Recall falls in favor of higher Precision.
However, Recall, which is more important than Precision, in REFERENCES
the application under consideration, is never equal to 1,
contrary to what has been achieved with the autoencoder. [1] H. E. Pence, “What is Big Data and Why is it Important?,” Journal of
The latter issue is related to the second disadvantage ii. Educational Technology Systems, vol. 43, no. 2, pp. 159–171, 2014,
described above: in this type of dataset, the baseload of the doi: 10.2190/ET.43.2.d.
current draw is not constant over time, but it is larger in the
daytime when all the industrial machinery is running. An
example of missed Detection due to this issue is shown in
Fig.8.

TABLE II
ANOMALY DETECTION RESULTS ON 10
FAULT EVENTS

Parameter Value

Precision 0.90

Recall 1.00

F1-Score 0.94
Fig. 8. Example of Missed Detection due to a change in current flow
baseload

Authorized licensed use limited to: Northeastern University. Downloaded on August 10,2023 at 06:01:32 UTC from IEEE Xplore. Restrictions apply.
[2] A. Gandomi and M. Haider, “Beyond the hype: Big data concepts, [16] Z. He, A. Raghavan, G. Hu, S. Chai, and R. Lee, “Power-Grid
methods, and analytics,” International Journal of Information Controller Anomaly Detection with Enhanced Temporal Deep
Management, vol. 35, no. 2, pp. 137–144, 2015, doi: Learning,” in 2019 18th IEEE International Conference On Trust,
https://doi.org/10.1016/j.ijinfomgt.2014.10.007. Security And Privacy In Computing And Communications/13th IEEE
[3] R. Baraniuk, D. Donoho, and Matan Gavish, “The science of deep International Conference On Big Data Science And Engineering
learning,” Proceedings of the National Academy of Sciences, vol. 117, (TrustCom/BigDataSE), 2019, pp. 160–167. doi:
no. 48, pp. 30029–30032, 2020, doi: 10.1073/pnas.2020596117. 10.1109/TrustCom/BigDataSE.2019.00030.
[4] M. Carratù, V. Gallo, V. Paciello and A. Pietrosanto, "A deep learning [17] P. K. Reddy Shabad, A. Alrashide, and O. Mohammed, “Anomaly
approach for the development of an Early Earthquake Warning Detection in Smart Grids using Machine Learning,” in IECON 2021 –
system," 2022 IEEE International Instrumentation and Measurement 47th Annual Conference of the IEEE Industrial Electronics Society,
Technology Conference (I2MTC), 2022, pp. 1-6, doi: 2021, pp. 1–8. doi: 10.1109/IECON48115.2021.9589851.
10.1109/I2MTC48687.2022.9806627. [18] B. Hooi et al., “GridWatch: Sensor Placement and Anomaly Detection
[5] G. Patrizi, A. Bartolini, L. Ciani, V. Gallo, P. Sommella and M. in the Electrical Grid,” 2018.
Carratù, "A Virtual Soil Moisture Sensor for Smart Farming Using [19] L. Feng et al., “Anomaly detection for electricity consumption in cloud
Deep Learning," in IEEE Transactions on Instrumentation and computing: framework, methods, applications, and challenges,”
Measurement, vol. 71, pp. 1-11, 2022, Art no. 2515411, doi: EURASIP Journal on Wireless Communications and Networking, vol.
10.1109/TIM.2022.3196446. 2020, no. 1, p. 194, Oct. 2020, doi: 10.1186/s13638-020-01807-0.
[6] D. Capriglione, M. Carratù, A. Pietrosanto and P. Sommella, "Soft [20] X. Liu and P. S. Nielsen, “Regression-based Online Anomaly
Sensors for Instrument Fault Accommodation in Semiactive Detection for Smart Grid Data,” Energy J., Jun. 2016, [Online].
Motorcycle Suspension Systems," in IEEE Transactions on Available: http://arxiv.org/abs/1606.05781.
Instrumentation and Measurement, vol. 69, no. 5, pp. 2367-2376, May [21] I. Mitiche, T. McGrail, P. Boreham, A. Nesbitt, and G. Morison, “Data-
2020, doi: 10.1109/TIM.2019.2963552. Driven Anomaly Detection in High-Voltage Transformer Bushings
[7] M. Carratù, V. Gallo, C. Liguori and V. Paciello, "Development of a with LSTM Auto-Encoder,” Sensors, vol. 21, no. 21, 2021, doi:
new speed measurement technique based on deep learning," 2022 IEEE 10.3390/s21217426.
International Instrumentation and Measurement Technology [22] G. Lee, M. Jung, M. Song, and J. Choo, “Unsupervised anomaly
Conference (I2MTC), 2022, pp. 1-6, doi: detection of the gas turbine operation via convolutional auto-encoder,”
10.1109/I2MTC48687.2022.9806625. in 2020 IEEE International Conference on Prognostics and Health
[8] S. Thudumu, P. Branch, J. Jin, and J. (Jack) Singh, “A comprehensive Management (ICPHM), 2020, pp. 1–6. doi:
survey of anomaly detection techniques for high dimensional big data,” 10.1109/ICPHM49022.2020.9187054.
Journal of Big Data, vol. 7, no. 1, p. 42, Jul. 2020, doi: 10.1186/s40537- [23] C.-W. Tsai, K.-C. Chiang, H.-Y. Hsieh, C.-W. Yang, J. Lin, and Y.-C.
020-00320-x. Chang, “Feature Extraction of Anomaly Electricity Usage Behavior in
[9] A. Bria, G. Cerro, M. Ferdinandi, C. Marrocco, and M. Molinara, “An Residence Using Autoencoder,” Electronics, vol. 11, no. 9, 2022, doi:
IoT-ready solution for automated recognition of water contaminants,” 10.3390/electronics11091450.
Pattern Recognition Letters, vol. 135, pp. 188–195, 2020, doi: [24] Y. Yuan and K. Jia, “A Distributed Anomaly Detection Method of
https://doi.org/10.1016/j.patrec.2020.04.019. Operation Energy Consumption Using Smart Meter Data,” in 2015
[10] M. Erz, J. F. Kielman, B. S. Uzun, and G. S. Gühring, “Anomaly International Conference on Intelligent Information Hiding and
detection in multidimensional time series—a graph-based approach,” Multimedia Signal Processing (IIH-MSP), 2015, pp. 310–313. doi:
Journal of Physics: Complexity, vol. 2, no. 4, p. 045018, Dec. 2021, 10.1109/IIH-MSP.2015.38.
doi: 10.1088/2632-072X/ac392c. [25] S. Lee et al., “Smart Metering System Capable of Anomaly Detection
[11] https://medium.com/datadailyread/types-of-data-anomalies- by Bi-directional LSTM Autoencoder,” in 2022 IEEE International
2f6fb1747eb1 Conference on Consumer Electronics (ICCE), 2022, pp. 1–6. doi:
[12] A. Takiddin, M. Ismail, U. Zafar, and E. Serpedin, “Deep Autoencoder- 10.1109/ICCE53296.2022.9730398.
based Detection of Electricity Stealth Cyberattacks in AMI Networks,” [26] H. Homayouni, S. Ghosh, I. Ray, S. Gondalia, J. Duggan, and M. G.
in 2021 International Symposium on Signals, Circuits and Systems Kahn, “An Autocorrelation-based LSTM-Autoencoder for Anomaly
(ISSCS), 2021, pp. 1–6. doi: 10.1109/ISSCS52333.2021.9497376. Detection on Time-Series Data,” in 2020 IEEE International
[13] J. Zhang, D. Wu, and B. Boulet, “Time Series Anomaly Detection for Conference on Big Data (Big Data), 2020, pp. 5068–5077. doi:
Smart Grids: A Survey,” 2021 IEEE Electrical Power and Energy 10.1109/BigData50022.2020.9378192.
Conference (EPEC), pp. 125–130, 2021. [27] C. Kaymakci, S. Wenninger, and A. Sauer, “Energy Anomaly
[14] L. Shouyu et al., “Anomaly detection of power grid dispatching Detection in Industrial Applications with Long Short-term Memory-
platform based on Isolation Forest and K-means fusion algorithm,” based Autoencoders,” Procedia CIRP, vol. 104, pp. 182–187, 2021,
Journal of Physics: Conference Series, vol. 1601, no. 2, p. 022010, Jul. doi: https://doi.org/10.1016/j.procir.2021.11.031.
2020, doi: 10.1088/1742-6596/1601/2/022010. [28] G. Patrizi et al., “Reliability Prediction of an innovative Power Quality
[15] A. Anwar and A. N. Mahmood, “Anomaly detection in electric network Meter,” in 2022 IEEE International Workshop on Metrology for
database of smart grid: Graph matching approach,” Electric Power Industry 4.0 & IoT (MetroInd4.0&IoT), Jun. 2022, pp. 195–200, doi:
Systems Research, vol. 133, pp. 51–62, 2016, doi: 10.1109/MetroInd4.0IoT54413.2022.9831774.
https://doi.org/10.1016/j.epsr.2015.12.006.

Authorized licensed use limited to: Northeastern University. Downloaded on August 10,2023 at 06:01:32 UTC from IEEE Xplore. Restrictions apply.

You might also like