Deep Learning Fault Detection Algorithms in WSNs

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 09 | Sep 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 264
Deep Learning Fault Detection Algorithms in WSNs
Prof. Dr. Gouri Patil1, S. Kaneez Rabiya Quadri
1,2Dept. Of CSE Engineering, Guru Nanak Dev Engineering College, Karnataka, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Networks of wireless sensors are deployed in
harsh environments. Their main advantages are flexibility
and low cost. But they may face many shortcomings that
lead to the need to improve data accuracy. Many artificial
intelligence techniques have displayed outstanding
performance in error detection and diagnosis. Recently,
machine learning has grown into a potent method based
on artificial intelligence to solve the failure difficulty with
WSN. Deep learning approach is been introduce for fault
awareness. Deep learning neural networks (artificial
neural networks) use a combination of data inputs,
weights, and biases to try to replicate the human brain.
These components cooperate to correctly identify,
categorise, and describe items in your data.
Key Words: WSN, Sensor, Deep Learning, CNN, ANN,
LSTM
1. INTRODUCTION
The term "wireless sensor network" (WSN) extends to a
group of unconnected sensor devices connected by a
wireless channel. These are structures of understanding
that work closely with the environment. They are
designed for very limited tasks. Basically, the sensor is
real-world apparatus that records information on a real-
world thing, process, or change in temperature or
pressure. WSN has real-time monitoring potential and is
already implemented in military applications, health
monitoring, industrial applications, environmental
monitoring, etc. WSN limits include node power and disk
space limits.
Wireless sensor networks can now support a range of
identification applications because to recent
developments in wireless communication and embedded
computing.. Utilizing wireless sensor networks to support
a variety of monitoring and control applications such as
environmental monitoring, industrial sensing, and traffic
control. Environmental monitoring, industrial sensors,
traffic sensors, and other small, low-power radio devices
are all included in a WSN. Small, low-power wireless
devices are frequently used in crowded or isolated areas
and make up a significant portion of WSNs. Various mobile
and inescapable applications constantly collect and
process data from the physical world, providing data on
detected situations or opportunities in great detail. In
particular, The benefits of information sparsity, global
optimality, and broad applicability make SVM a desirable
classification approach.
WSNs are prone to failure because they are routinely
installed in high-risk, unmonitored, inaccessible
environments. These conditions can be further categorised
into several groups.
• software errors;
• Hardware failure and
• Communication failure.
In conjunction with the information gathered, The
following descriptions perhaps utilised to classify errors.
• Gain Error: when the rate of change of the acquired
data is different from the expected value. • Stuck on
Error: When there is no change in the set of collected
data.
• Out of Range: When an observation falls outside the
expected range.
• Peak error: when the estimated time series's excess
beyond the time series' projection is greater than the
predicted trend of change.
• Noise error: when randomly distributed numbers
are added to the expected value.
• Data Loss Error: When a certain amount of data is
missing from the collected values during a certain
time interval. • Random error: This is an error in
which the observed data is unbalanced.
To identify various problems that can occur in wireless
sensor networks, we employ a variety of algorithms and
deep learning techniques in our study.
1.1 Motivation
There have been significant advancements regarding
wireless sensor networks and technologies recently. They
are mainly used for communication. Communication
between different devices is wired or wireless, so the risk
of fire and explosion over the network is increasing every
day. Identification and reduction of fraud are the main
priorities when it comes to secure communication. As a
result, testing of exposure and penetration prevention
techniques has become an important part of the
engineering field. Using an exposure and intrusion
prevention system, we can identify and then report
normal and unusual user activity. Therefore, For wireless
sensor networks, it is essential to use deep learning and
machine learning to create an efficient intrusion detection
and mitigation system. In this piece, a trial involved and
evaluating the effectiveness of several deep learning and

machine learning for malware detection and mitigation
systems. The performance estimation of these techniques
was performed by experiments performed on the WSN-DS
dataset. Comparative evaluation shown by deep learning
classifications is better penetration exposure results than
machine learning techniques.
1.2 Problem Definition
Using deep learning to detect Wireless sensor network
faults can arise for a number of reasons. WSNs are used in
risky, unattended and inaccessible environments, making
them more susceptible to power outages. These errors fall
into three groups: software errors, hardware failures, and
communication errors.
1.3 Challenges
Error detection in WSN faces many challenges due to the
following reasons:
• The facilities and resources at the node level are very
limited, which forces the nodes to use classifiers because
they do not require complex computations.
• Sensor nodes are installed in hazardous environments,
for example. at home, indoors, in war zones, in hurricanes,
earthquakes, etc.
• The error detection process must be accurate and fast to
avoid any loss, for example, the process must determine
the difference between the abnormal and the normal, so
that it can be lost in the event of a collection. Collecting
wrong data can lead to erroneous results.
2. EXISTING SYSTEM
WSNs are built with many sensor nodes connected and
sharing the information collected by them. Administering
a network that is so vast and intricate requires scalable
and efficient algorithms. Also, for some reason, WSNs can
change dynamically and require a redesign of the entire
network architecture. This may sometimes require
changes to routing strategy, location of specific nodes,
interlayer design, etc. Algorithms for machine learning are
required to deal with such situations. With ML, machines
learn on their own without human intervention or any
kind of reprogramming. ML algorithms can accurately test
complex data at node speed. The foundation of WSN is
constituted of ML algorithms, which have the capacity to
deliver optimum solutions through self-learning.
2.1 Disadvantages of Existing System
• Power outages are reported in the WSN for a number of
reasons. One of the reasons could be the location where
the WSNs are deployed. Reliance on sensor nodes'
batteries, hardware and software failures, as well as
required topology changes can be other reasons. The
multi-error detection classifier might not be able to be
found by the existing system.
• Due to resources being scarce, it can be challenging to
identify these errors., the harsh environment in which the
WSN is deployed, or the separation of failed and non-faulty
nodes
2.3 Datasets:
Data is collected from two outdoor multi-step sensors. It is
temperature and humidity data detected. Each vector
consisted of data collected at three consecutive cases t0, t1
and t2, and each case was constructed from two
temperature measurements and two humidity
measurements T1, T2 and H1, H2. Then, different types of
errors (lag, boost, freeze, out of bounds, spike and data
loss) are randomly caused at different rates (10%, 20%,
30%, 40%) and 50%). A total of 40 datasets were
assembled with a set of 9566 tests (vectors) and 12
dimensions for each set. Data sets labeled with a target
column are marked as one for normal testing and -1 for
outliers.
2.4 Proposed System
Networks of wireless sensors are deployed in harsh
environments. Their main advantages are flexibility and
low cost. But they may face many shortcomings that lead
to the need to improve data accuracy. Many artificial
intelligence techniques have displayed outstanding
performance in error detection and diagnosis. Recently,
machine learning has grown into a potent method based
on artificial intelligence to solve the failure difficulty with
WSN. In this essay, For defect awareness, a deep learning
technique was introduced.
3 DESIGN
3.1 Architecture
Error detection mechanisms are considered to be of great
importance to ensure the normal operation of WSNs. To
prevent loss and clearly indicate the condition of the data,
they must be precise and quick. However, due to the
sensor's restricted properties, faults are challenging to
detect. The mechanism of anomaly detection has been the
subject of numerous studies from various perspectives.
Few approaches are distributed, centralized, or hybrid.
They are based on dynamics, auto-detection, and machine
learning. Artificial intelligence is implemented through
machine learning, which gives systems the capacity to
autonomously learn from the past and get better.
Classification is a frequently often used strategies. to data
mining, It is a part of artificial intelligence. It clearly
divides data into different categories and helps in decision

making. According to data details, there are three classes
of machine learning techniques:
Supervised Learning: Data mining strategies are applied to
data labeled with predefined classes.
Unsupervised learning: Approaches applied to unlabeled
data. Data is classified without prior knowledge.
Semi-supervised learning: Here, both unsupervised and
supervised learning are combined.
For fault detection, Convolutional Neural Network,
Artificial Neural Network (ANN), Decision Tree (DT),
LSTM and Random Forest (RF) classifiers, are used to
classify the sensed data into two cases, i.e. normal cases or
abnormal cases.
Fig. 3.1 Types of classifiers
3.2 Fault detection Algorithms
3.2.1 Decision Tree
Classification is a two-step process, a learning step and a
prediction step, in machine learning. During the training
phase, the model develops based on the given training
data. In the prediction step, the model is used to predict
the response to the given data. Decision trees are
straightforward and often used classifications algorithms
to understand and interpret.
Fig. 3.2 Decision Tree
3.2.2 Random Forest
Individual decision trees are combined to create a random
forest, and finally, each decision tree votes in making the
correct prediction for the concerned problem.
3.2.3 CNN
A deep learning method called Integrated Neural Network
(ConvNet/CNN) can process photos as input, assign
importance (weights and assimilable biases) to other
aspects/objects images and can distinguish between them.
Preprocessing requirements in ConvNet are much lower
than in other classification algorithms.
Fig. 3.4 Convolutional neural network architecture
3.2.4 Artificial Neural Network
A good way to think of a NN is as an aggregate function.
You give it an input and it gives you an output.
Three parts make up the architecture of the basic NN.
These are:
 Units / Neurons.
 Connections / Weights / Parameters.
 Prejudices.
To create a basic NN architecture, you require all of the
aforementioned components.

Fig. 3.5 Artificial Neural Network Architecture
3.2.5 LSTM
Neural networks are a set of algorithms that closely
resemble the human brain and are designed to recognize
patterns. Through automated perception, categorization,
or grouping of unprocessed inputs, they evaluate sensory
data. They are able to identify the digital patterns found in
the vectors that must be used to transform all real data
(such as pictures, sounds, texts, or time series).
Fig. 3.5 LSTM Architecture
5 RESULT
Fig. 5.1 Snapshot of providing input for fault detection
Fig. 5.2 Snapshot of getting output for fault detection
Fig. 5.3 Snapshot of providing input for fault detection
Fig. 5.4 Snapshot of getting output for fault detection
4. CONCLUSIONS
The research work for this project is preceded by a dataset
preparation block. The insight makes up the dataset vector
𝑉𝑡, which consists of her two water measurements H1 and
H2 and two temperature measurements, T1 and T2, are

present three times in succession. Two errors with
different error rates (10%, 20%, 30%, 40%, and 50%) are
then inserted into the record: a spike error and a data loss
error. Next, six classifiers, ANN, CNN, KNN, DT, RF, and
LSTM applied to outdoor data collected from multi-hop
WSNs. Classifiers are scored based on four different
performance metrics.
In future work, we will use the same classifier to predict
the next error in the data and develop an error avoidance
mechanism. In addition, we will work on WSN failure
detection to accurately identify and subsequently detect
failures at the sensor (node) level. The sturdiness of the
algorithm can also be confirmed by expanding the sensor
count. This helps us understand her WSN's resilience to
attacks.
5 REFERENCES
1. "Wireless sensor network survey," B. Mukherjee,
D. Ghosal, and J. Yick, Computer Networks, volume.
52, 2008, pp. 2292-2330.
2. R. A. Shaikh and M. Thaha, "An examination of fault
detecting algorithms in wsns," Journal of Network
and Computer Applications, vol.78, no.3,2017, pp.
267–287.
3. "Fault detection in wsns by SVM classifier," IEEE
Sens. J., no.18, pp.340-347,2017. Moulahi, S. Zidi
and B. Alaya
4. "A Data_Driven_Design for Fault Detection of
Wind_Turbines Using Random Forests and
XGboost," IEEE Access, vol. 6, pp. 21020-21031,
2018. D. Zhang, L. Qian, B. Mao, C. Huang, B.
Huang, and Y. Si
5. Probabilistic neural networks: a quick introduction
of theory implementation and application, Elsevier,
2020, pp. 347–367. M. Behshad, T. Amirhessam, M.
B. Anke, and G. H. Amir.
6. "A Review of ML Based Fault Detection Algorithms
in WSNs," Suman Avdhesh Yadav.
7. Agarwal, A., Sinha, V. K., & Palisetty, R. (2019).
Performance analysis and FPGA prototype of
variable rate GO-OFDMA baseband transmission
scheme. Wireless Personal
Communications, 108, 785–809.

Deep Learning Fault Detection Algorithms in WSNs

Recommended

More Related Content

Similar to Deep Learning Fault Detection Algorithms in WSNs (20)

More from IRJET Journal (20)

Recently uploaded (20)

Deep Learning Fault Detection Algorithms in WSNs