RUL Prediction of Electro-mechanical Equipment

under Diverse Cyber Attacks

1ST AZKA ALI 2ND MUHAMMAD AHMAD

Artificial Intelligence Artificial Intelligence
Technology Center (AITeC) Technology Center (AITeC)
National Center for Physics National Center for Physics
Islamabad, Pakistan Islamabad, Pakistan
[email protected] [email protected]

Abstract—The integrity of Predictive Maintenance (PdM) pressure ratio, fan inlet pressure, and rotational speeds to
systems is increasingly threatened by cyber-attacks. This study ensure timely maintenance and safety. Several IoT (Internet
investigates the impact of cyber-attacks on the performance of of Things) sensors are mounted both inside and outside of an
Deep Learning (DL) based Predictive Maintenance (PdM) engine as part of an EHM system to monitor different
systems using NASA’s Commercial Modular Aero-Propulsion parameters. All IoT sensors are wirelessly connected [3]. IoT
System Simulation (C-MAPSS) dataset. Long-Short Term sensors monitor aircraft engine statistics and notify
Memory (LSTM) and 1D Convolutional Neural Network (1D manufacturers when an engine's Remaining Useful Life
CNN) models are designed and deployed to predict the
(RUL) is approaching its end [4]. This is accomplished by
Remaining Useful Life (RUL) of the turbofan engines.
utilizing Predictive Maintenance (PdM) systems within EHM
Specifically, two distinct attack scenarios i.e. Episodic Data
Injection Attack (EDIA) and Continuous Data Injection Attack
systems to predict RUL. EHM systems are essential for
(CDIA) are generated with different variations including biased cutting costs associated with labour, fuel, hardware, and
and random scenarios on the sensors data and their influence on logistics, enabling early identification of potential problems
the RUL prediction of the equipment is analyzed. Using various and reducing maintenance and operating expenses.
Machine Learning (ML) algorithms including Decision Tree, PdM systems are vulnerable to cyber threats, including
Random Forest Classifier, K-Nearest Neighbors (KNN) and Data manipulation Spoofing attacks, which compromise
Logistic Regression, two different types of Data manipulation
physical sensors, communication networks, and data
Spoofing attacks: Random spoofing and Constant Spoofing are
processing systems. Due to unforeseen malfunctions, these
classified. This study demonstrates how attacks can significantly
damage the accuracy of predictive maintenance systems, attacks can lead to delayed asset failures, costly maintenance
emphasizing the need for strong cybersecurity safeguards in periods, and even fatalities [5]. IoT devices, which play a
Internet of Things (IoT) contexts to secure essential predictive significant role in predictive maintenance, are also vulnerable
maintenance in addition to the classification of cyber-attacks to attacks such as botnet assaults, denial-of-service attacks,
and healthy data. and data breaches. False data injection attacks can threaten the
integrity of EHM systems, resulting in inaccurate diagnoses
Keywords— Deep Learning (DL), Predictive maintenance and maintenance decisions. Data manipulation Spoofing
(PdM), Commercial Modular Aero-Propulsion System Simulation attacks manipulate IoT sensor measurements, often targeting
(CMAPSS), Long-Short Term Memory (LSTM), 1d Convolutional predictive maintenance systems. These cyber-attacks can
Neural Network (1D CNN), remaining useful life (RUL), Episodic cause catastrophic failures or delays, emphasizing the
Data Injection Attack (EDIA), Continuous Data Injection Attack vulnerability of IoT systems to data manipulation.
(CDIA), Machine Learning (ML), Data manipulation Spoofing
attacks, in Internet of Things (IoT). Data manipulation Spoofing attacks have caused
catastrophic events like the 2003 USA Northeast blackout and
I. INTRODUCTION the Ukrainian power grid attack. On August 14, 2003, the US
The rise of Internet of Things (IoT) technology has andCanada experienced the largest electrical power blackout
sparked a surge in AI interest, leading to increased use of AI in North American history, affecting 40 million people in
techniques in IoT cybersecurity applications to detect threats eight states and 10 million in Ontario. The outage caused
and potential attacks, including decision trees, neural financial losses of $4-10 billion and was caused by monitoring
networks, and support vector machines [1]. The latest and diagnostic system failure, communication issues, and lack
industrial revolution, Industry 4.0, combines automation with of system understanding [6]. Whereas Ukrainian power grid
modern production to reduce human labour and resources. attack left over 230,000 people without electricity for several
Predictive maintenance (PdM) is an industry 4.0 solution that hours [7]. These events are directly attributable to Data
uses advanced machine learning (ML) techniques, including manipulation Spoofing attacks, necessitating extensive
deep learning, and IoT sensors to detect failures in research to detect and mitigate these attacks. These real-life
components or systems. scenarios highlight how Data manipulation Spoofing attacks
can exploit vulnerabilities, leading to costly failures and
Engine health monitoring (EHM) [2] systems monitor potential disastrous consequences, especially in safety-critical
engine parameters like exhaust gas temperatures, engine applications like power grids and industrial plants.

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

 Artificial intelligence is being utilized to enhance IoT A. Dataset
security in predictive maintenance and diagnosis. Deep This study utilizes NASA's C-MAPSS turbofan engine
Learning (DL) approaches, such as Convolutional Neural degradation simulation dataset [10] to test CNN and LSTM
Network (CNN) and Long-Short Term Memory (LSTM), deep learning methods. The data was obtained from a high-
have demonstrated rapid development and outperformed fidelity system-level engine simulation designed to simulate
standard prognostic algorithms in RUL forecasts for turbofan nominal and fault engine degradation over a series of flights.
engine degradation [8]. AI systems analyse large-scale data to There are four sub-datasets (FD001-04) in this dataset. Each
identify abnormal patterns and trigger prompt intervention [9]. subset contains training and test data. The recordings were
By learning device behavior, AI can identify security issues sampled at 1 Hz and consisted of 30 engine and flight
by learning the behavior of devices. Automated responses to condition parameters. Faults/anomalies were identified from
security issues have been enabled, lowering the impact of the inlet engine fan, low-pressure compressor, high-pressure
cyber-attacks. AI-driven security solutions strengthen compressor, high-pressure turbine, and low-pressure turbine.
cybersecurity posture, enabling prompt and efficient The dataset that we used for our study is FD001 that contains
responses, and protecting IoT devices and data from malicious 21 sensors with varying operating and fault conditions, for
actors. Physical sensor compromise requires physical access, training and test data. Each row represents a time cycle, with
while hacking data transmission lines and processing 26 columns representing engine ID and current operational
programs is a simpler method [7]. Successful Data cycle number. Table I provides the details for CMAPSS
manipulation Spoofing attacks can cause engine sensor dataset.
inaccurate numbers, affecting predictive maintenance
TABLE I
This study investigates the impact of Data manipulation FEATURES OF COMMERCIAL MODULAR AERO-PROPULSION SYSTEM
Spoofing attack on aircraft engine PdM systems, highlighting SIMULATION (CMAPSS) DATASET.
potential maintenance delays and catastrophic failures. In this CMAPSS FD001
study, we focused on three main contributions: Sr. Number Column Name Description
1. ID Identifier for the engine unit
2. Cycles Time of operation (in cycles)
1. The study uses Deep Learning algorithms such as LSTM 3. Operational condition 1 Engine operational setting 1
and 1D CNN to analyse the RUL of a turbofan engine, 4. Operational condition 2 Engine operational setting 2
utilizing the NASA’s Commercial Modular Aero- 5. Operational condition 3 Engine operational setting 3
Propulsion System Simulation (CMAPSS) dataset for an 6. Sensor 01 Sensor measurement 1
aircraft predictive maintenance system. 7. Sensor 02 Sensor measurement 2
2. The study analysed two attack scenarios generated on 8. ... Additional sensor
sensor data to evaluate their impact on the Remaining measurements
Useful Life (RUL) of a turbofan engine. 9. Sensor 21 Sensor measurement 21
10. RUL Remaining useful life
a) Episodic Data Injection Attack (EDIA) and (engine health)
Continuous Data Injection Attack (CDIA) were
simulated to represent actual cyberattacks on IoT
devices. B. Compute Resources:
b) Deep Learning algorithms were used to predict the A desktop computer equipped with a 13th generation Intel
engine's RUL, assessing their response to data core i7 CPU, 16 GB RAM, and a 64-bit operating system was
manipulation spoofing attacks. used in the experiment for data processing and machine
learning tasks. This hardware capacity ensured accurate and
3. Furthermore, the study generated and classified Random quick results, highlighting the necessity of sufficient
and Constant Spoofing attacks. For the classification of computational resources for advanced research in predictive
these attacks five Machine Learning methods including maintenance and diagnostics.
Decision Tree, Random Forest, KNN, Logistic
Regression, and Support Vector Machine were
evaluated. Figure 1 provides an illustration of the II. METHODOLOGY
attacks.
This research is divided into three sections. Firstly, this
study evaluated the Remaining Useful Life (RUL) using
LSTM and 1D CNN models trained on the C-MAPSS dataset.
Secondly, the impact of cyber-attacks by simulating episodic
and continuous attacks was analyzed. The study investigated
the impact of these various attack scenarios on the accuracy
and resilience of our RUL predictions. Finally, two attacks
(Random and Constant) were generated and classified using
Fig. 1. Illustrations for Constant Spoofing and Random Spoofing
five ML algorithms.
respectively
A. Section I
Dataset Preprocessing:
Materials & Methods To extract important features and labels for RUL
estimation, the CMAPSS dataset was loaded and pre-
processed. Sensor readings and operational conditions were
retrieved, and RUL values were calculated depending on the
number of cycles remaining until failure. The training set has
data until the engine achieves a predetermined degradation
threshold, whereas the test set contains time series data that B. Section II
ends before complete degradation, with RUL values available In this section episodic and continuous cyber attacks with
for evaluation. different variations were developed to assess their impact on
the accuracy of RUL forecasts. The goal was to determine how
Data Splitting: various attacks and their variations affect the performance of
Then we divided the pre-processed dataset into training deep learning models used in predictive maintenance.
and test sets. Most of the data is usually used to train the Attack Simulation:
models, with a little amount set aside to evaluate the model's
functionality. We generated various attack scenarios on sensor data to
evaluate their impact on the remaining useful life (RUL) of a
Model Selection: turbofan engine. The data was analysed by deleting
We performed RUL estimation by using two deep unnecessary columns and focusing on key sensor readings.
learning architectures: 1D CNN and LSTM networks. The CDIA attack was begun for a specified range of engines,
LSTMs [11] are Recursive Neural Networks (RNNs) that can beginning with engine ID 1 and ending with engine ID 37. It
was programmed to start on the first cycle of each engine and
learn long-term dependencies. Unlike RNN, which continue through all subsequent cycles, targeting three
frequently has long-term dependency issues, LSTM is sensors. The original dataset was changed with the
specifically built to prevent these issues. It has received high manipulated sensor readings, which replaced valid signals
appreciation for its achievements in machine learning and with biased values for the selected engine and cycle ranges.
speech recognition. CNN, a deep learning algorithm, is more For the CDIA biased scenario, a biased noise was introduced
effective than traditional machine learning methods. It uses to the chosen sensors, with a fixed value shown in Table 01.
feed-forward artificial neural networks and has various This bias was used to model a continual attack on sensor
readings. For the CDIA random scenario, we defined the
applications, including a CNN-based PdM system [12]. [13]
bounds for the noise to be added i.e. with values between a
provides a detailed overview of CNN construction and specific bound (high to low) was applied to the selected
network design. This paper uses LSTM and 1D CNN models sensors. Using the formula 𝑟 = 𝑎 + (𝑏 − 𝑎) ⋅ 𝑟𝑎𝑛𝑑(1,1), a
to estimate the engine's RUL. These models were chosen for random noise value within the specified bounds was produced
their capacity to detect temporal connections in sequential for each sensor. This noise was then applied to the selected
sensor readings.
data, which is critical for time-series forecasting tasks such as
RUL estimation. 𝑛𝑜𝑖𝑠𝑒 = 𝑏𝑜𝑢𝑛𝑑1 + (𝑏𝑜𝑢𝑛𝑑2 − 𝑏𝑜𝑢𝑛𝑑1) ∗ 𝑟𝑎𝑛𝑑(1,1)
Model Training: For the EDIA attack, the attack was limited to a specified
The 1D CNN and LSTM models were trained using the cycle period, for example ranging from cycle start=130 to
training dataset. During training, the models learn to map cycle end=150, which represented a periodic attack. Engines
input sensor data and operating parameters to their respective and sensors were defined for the attack same as we did for
RULs. CDIA. The EDIA biased and random attacks were
Model evaluation: implemented with the same logic as discussed earlier for the
It involves assessing the performance of trained models CDIA.
against a testing dataset. Evaluation metric like Root Mean C. Section III
Square Error (RMSE) is used to assess the difference between
We generated two different Data manipulation Spoofing
anticipated RUL values and ground truth. The performance
attacks and classified the data by using different ML
of 1D CNN and LSTM models is compared using RMSE
algorithms. The attacks we generated involved Random
scores. spoofing and Constant spoofing.
Hyperparameter Tuning:
Hyperparameter tuning techniques can be used to Attacks:
improve the performance of the models even further. To Random Spoofing is a sort of attack in which interference
increase the accuracy of the models, parameters such as or noise is introduced in a random manner. The attacker's
learning rate, batch size, and network architecture were transmissions are intermittent and unpredictable, making it
adjusted. Table II shows the hyperparameters used for impossible for the target system to detect the attack. The
training the deep learning algorithms. attack cycles between active mode and non-active i.e. sleep
TABLE II
mode. The timing and duration of these phases are random.
HYPERPARAMETERS FOR 1D CNN AND LSTM. Constant spoofing emphasizes on how consistent and
ongoing the attack is. This makes it easier to detect and
1D CNN address than random spoofing. Constant spoofing attacks
Sequence RMSE Epochs Batch Validation- involve continuously injecting random garbage data into the
length size split
100 8.4139 100 200 0.05 sensor stream without following any pattern. Unlike random
spoofing, which may have occasional variations, constant
LSTM spoofing transmits a consistent stream of incorrect data,
Sequence RMSE Epochs Batch Validation- resulting in permanent noise in sensor readings.
length size split
80 7.7389 100 200 0.09
 III. RESULTS 1D CNN
0.010 34.5812 33.4721 0.005- 34.1126 34.3341
A. Section I: 0.01
0.015 35.5310 36.6834 0.01- 35.1872 42.2844
This section shows the applicability of many models for 0.02
predicting the Remaining Useful Life (RUL) of turbofan 0.020 36.6025 41.7755 0.02- 37.5124 58.5698
engines. The RMSE was used to assess each model's 0.03
performance. The table below summarizes the findings and 0.025 37.9206 48.6859 0.03- 43.1100 79.8149
displays the RMSE values for both the training and testing 0.04
phases. Training and testing RMSEs for the 1D Convolutional 0.030 39.7521 57.1600 0.04- 45.3468 103.638
0.05
Neural Network (1D CNN) model were 8.4139 and 9.8918,
0.05 49.1911 97.0199
respectively. On the other hand, the Long Short-Term
Memory (LSTM) model performed well, with a training
RMSE of 7.7389 and a testing RMSE of 9.1998. Table III C. Section III:
provides the training and testing RMSEs of the models
This section focuses on the classification of Data
utilized.
manipulation Spoofing attacks on turbofan engine sensor
TABLE III values. Different machine learning algorithms have been used
COMPARISON OF RMSE SCORES FOR 1D CNN AND LSTM MODELS IN to classify the attacks to evaluate their detection. We
RUL PREDICTION
generated three different forms of attacks: Random Spoofing,
RUL Prediction without Cyber Constant Spoofing and Data Manipulation Spoofing. We
generated these attacks, and then classified the modified
MODELS Root Mean Square Error sensor inputs using different machine learning techniques.
Table V and Table VI shows the train and test accuracies for
1D CNN TRAIN RMSE TEST RMSE various ML algorithms including Decision Tree, Random
Forest Classifier, K-Nearest Neighbours (KNN) and Logistic
8.4139 9.8918 regression for Random and Constant spoofing attacks
respectively.
LSTM TRAIN RMSE TEST RMSE
TABLE V
7.7389 9.1998 TRAINING AND TESTING ACCURACY OF ML MODELS UNDER RANDOM
SPOOFING ATTACK SCENARIO

B. Section II: RANDOM SPOOFING

This section shows how different Data manipulation
Spoofing attacks affect the precision of RUL estimates for Algorithms Train Accuracy Test Accuracy
turbofan engines using LSTM and 1D CNN models. The Decision Tree 1.00 1.00
models were evaluated using the RMSE. The results show the
RMSE fluctuations resulting from both EDIA and CDIA Random Forest 1.00 1.00
across a variety of attack ranges. Table IV shows how much Classifier
the RUL predictions for both models are impacted by Data
manipulation Spoofing attacks. The RMSE values increase KNN 0.9897 0.9848
with the severity of the attacks.
Logistic Regression 0.8043 0.7990
TABLE IV
IMPACT OF EDIA (BIASED AND RANDOM) AND CDIA (BIASED AND SVM 0.8305 0.8277
RANDOM) ATTACK SCENARIOS ON REMAINING USEFUL LIFE (RUL) FOR
DIFFERENT RANGES.
TABLE VI
TRAINING AND TESTING ACCURACY OF ML MODELS UNDER
RUL Prediction with Cyber Attack CONSTANT SPOOFING ATTACK SCENARIO
Episodic Continuous Episodic Continuous
Data Data Data Data
Injection Injection Injection Injection CONSTANT SPOOFING
RANGE Attack Attack RANGE Attack Attack
(EDIA) (CDIA) (EDIA) (CDIA) Algorithms Train Accuracy Test Accuracy
Biased Biased Random Random
RMSE RMSE RMSE RMSE Decision Tree 1.00 1.00
LSTM
0.010 41.7048 41.1284 0.005- 41.5872 40.7602 Random Forest 1.00 1.00
0.01 Classifier
0.015 42.1503 40.5768 0.01- 42.1156 40.0923
0.02 KNN 0.9950 0.9921
0.020 42.6067 40.0433 0.02- 42.8755 42.4744
0.03 Logistic 0.8698 0.8719
0.025 42.806 40.4796 0.03- 43.0978 43.7817 Regression
0.04
0.030 42.9348 40.6672 0.04- 43.4871 44.5032 SVM 0.8838 0.8863
0.05
0.05 43.5976 42.2641
 Impact of Cyber Attacks on 1D CNN RUL Predictions
D. Discussion: 1. EDIA Biased:
A. Section I The RMSE increases from 34.5812 at the lower range
In this study, we used RMSE to assess the prediction (0.010) to 49.1911 at the higher range (0.05). An abrupt
model’s performance in predicting turbofan engine’s RUL. increase in RMSE suggests that 1D CNN RUL prediction
The 1D CNN model achieved an RMSE of 8.4139 on the accuracy has been significantly affected by EDIA biased
training data and 9.8918 on the test data. Whereas LSTM attacks.
model achieved an RMSE of 7.7389 on the training data and 2. CDIA Biased:
9.1998 on the test data, demonstrating its ability to effectively
capture the engine's complex temporal dependencies. These The RMSE starts at 33.4721 for the lower range (0.010)
RMSE values indicate the models' relative performance and and skyrockets to 97.0199 at the higher range (0.05). This
applicability for Remaining Useful Life (RUL) prediction attack type results in the greatest rise in RMSE, indicating that
tasks. Fig. 2 clearly shows that the LSTM model is better CDIA biased attacks have the greatest detrimental influence
suited to accurate Remaining Useful Life (RUL) prediction in on 1D CNN RUL predictions.
this circumstance. 3. EDIA Random:
The RMSE increases from 34.1126 at the lower range
(0.005-0.01) to 45.3468 at the higher range (0.04-0.05).
Although the RMSE increases, the effect is less severe than
biased attacks.
4. CDIA Random:
The RMSE starts at 34.3341 for the lower range (0.005-
0.01) and dramatically increases to 103.638 at the higher range
(0.04-0.05). This attack type results in the largest RMSE of
any attack type, showing that random CDIA attacks have the
most impact on 1D CNN RUL predictions.
From the analysis, it is evident that CDIA random have the
Fig. 2. Deep learning algorithms training and testing RMSE for greatest effect on RUL prediction accuracy for both LSTM
Remaining Useful Life prediction. and 1D CNN models, with the highest RMSE values found in
this category. Fig. 4 shows the comparison of these attacks.
B. Section II: CDIA Biased also show a considerable impact, particularly on
the 1D CNN model. EDIA biased and random have an impact
Impact of Cyber Attacks on LSTM RUL Predictions on models, but they are less severe than CDIA. Among the
1. EDIA Biased: models, 1D CNN is more sensitive to biased and random
CDIA, with larger RMSE values than LSTM under similar
The RMSE increases from 41.7048 at the lower range attack conditions. These results demonstrate how susceptible
(0.010) to 43.5976 at the higher range (0.05). There is a RUL prediction models are to various cyberattacks, with
gradual increase in RMSE, demonstrating that EDIA biased biased and random spoofing attacks posing the biggest risk to
attacks have a notable impact on the accuracy of RUL predict RUL of industrial equipment and effect their accuracy.
predictions, with higher ranges resulting in greater deviance. This highlights the importance of effective detection and
mitigation measures to safeguard such models from cyber
2. CDIA Biased:
attacks.
The RMSE starts at 41.1284 for the lower range (0.010)
and rises to 42.2641 at the higher range (0.05). While there is
an increase in RMSE, it is not as much as in the case of EDIA
biased attacks, indicating that CDIA biased attacks have
marginally less of an effect on RUL prediction accuracy.
3. EDIA Random:
The RMSE increases from 41.5872 at the lower range
(0.005-0.01) to 43.4871 at the higher range (0.04-0.05). EDIA
random attacks exhibit a steady rise in RMSE, like EDIA
biased attacks, suggesting a noteworthy influence on RUL
predictions. Fig. 4. Comparison of Different EDIA and CDIA Attacks and Their
Impact on RUL
4. CDIA Random:
The RMSE starts at 40.7602 for the lower range (0.005- C. Section III
0.01) and rises to 44.5032 at the higher range (0.04-0.05). This
form of attack exhibits the biggest RMSE increase of any In this section, we classified two types of data spoofing
attack type, indicating that CDIA random attacks have the attacks: Random Spoofing and Constant Spoofing. We used a
most impact on LSTM RUL predictions. variety of machine learning approaches, including Decision
Tree, Random Forest Classifier, KNN, Logistic Regression
and Support Vector Machine to detect these attacks. The References
performance of these classifiers was evaluated based on their
accuracy and capacity to recognize altered data. Our study
[1]. Kuzlu, M., C. Fair, and O. Guler, Role of artificial intelligence in
showed that three classifiers were quite good at detecting the Internet of Things (IoT) cybersecurity. Discover Internet of
attacks. The Decision Tree, Random Forest Classifier, and things, 2021. 1(1): p. 7.
KNN performed well at differentiating between normal and [2]. Tumer, I. and A. Bajwa. A survey of aircraft engine health
attacked sensor data. Fig. 3 demonstrates how effective monitoring systems. in 35th joint propulsion conference and
machine learning models are in detecting and addressing exhibit. 1999.
[3]. Bai, H., M. Atiquzzaman, and D. Lilja. Wireless sensor network
possible data spoofing vulnerabilities in IoT devices. for aircraft health monitoring. in First International Conference
on Broadband Networks. 2004. IEEE.
[4]. Si, X.-S., et al., Remaining useful life estimation–a review on the
statistical data driven approaches. European journal of
operational research, 2011. 213(1): p. 1-14.
[5]. Mode, G.R., P. Calyam, and K.A. Hoque. Impact of false data
injection attacks on deep learning enabled predictive analytics. in
NOMS 2020-2020 IEEE/IFIP Network Operations and
Management Symposium. 2020. IEEE.
[6]. Luke, T.W., Power loss or blackout: The electricity network
collapse of August 2003 in North America. Disrupted cities:
When infrastructure fails, 2010: p. 55-68.
[7]. Mode, G.R., P. Calyam, and K.A. Hoque, False data injection
attacks in internet of things and deep learning enabled predictive
analytics. arXiv preprint arXiv:1910.01716, 2019.
[8]. Ellefsen, A.L., et al., Remaining useful life predictions for
turbofan engine degradation using semi-supervised deep
Fig. 3. ML algorithm’s training and testing accuracies for two different architecture. Reliability Engineering & System Safety, 2019.
attack scenarios. 183: p. 240-251.
[9]. Powrie, H. and C. Fisher. Engine health monitoring: towards total
Conclusion prognostics. in 1999 IEEE Aerospace Conference. Proceedings
RUL prediction using AI models can be severely (Cat. No. 99TH8403). 1999. IEEE.
[10]. Saxena, A. and K. Goebel, Turbofan engine degradation
compromised and miscalculated due to cyber attacks on simulation data set. NASA Ames Prognostics Data repository,
sensory data. This analysis is based upon ID CNN and LSTM NASA Ames Research Center, Moffett Field. 2008, ed.
models trained for RUL predictions. In second part of this [11]. Zheng, S., et al. Long short-term memory network for remaining
study, a new model has been developed to classify data as useful life estimation. in 2017 IEEE international conference on
either under cyber attack or healthy, thereby improving the prognostics and health management (ICPHM). 2017. IEEE.
[12]. Silva, W., Cnn-pdm: A convolutional neural network framework
reliability of RUL predictions. The RMSE values increase for assets predictive maintenance. 2019, The University of
with the severity of the attacks, particularly for higher ranges Western Ontario (Canada).
of CDIA random attacks. This demonstrates how attacks can [13]. Ince, T., et al., Real-time motor fault detection by 1-D
significantly damage the accuracy of PdM systems, convolutional neural networks. IEEE Transactions on Industrial
emphasizing the need for strong cybersecurity safeguards in Electronics, 2016. 63(11): p. 7067-7075.
IoT context to secure essential predictive maintenance and
classification of these attacks.