ISSN: 3005-9852 Journal of Robotics Spectrum 1(2023)

Deep Learning and Machine Learning Algorithms

for Enhanced Aircraft Maintenance and Flight
Data Analysis
Malene Helgo
Federal University of Maranhão, Vila Bacanga, Sao Luís - MA, 65080-805, Brazil.
[email protected]

Correspondence should be addressed to Malene Helgo : [email protected]

Article Info
Journal of Robotics Spectrum (https://anapub.co.ke/journals/jrs/jrs.html)
Doi: https://doi.org/10.53759/9852/JRS202301009
Received 05 March 2023; Revised from 20 April 2023; Accepted 12 May 2023.
Available online 15 May 2023.
©2023 Published by AnaPub Publications.
This is an open access article under the CC BY-NC-ND license. (http://creativecommons.org/licenses/by-nc-nd/4.0/)

Abstract – This paper examines the use of machine learning and deep learning algorithms in the aviation industry, with a
specific emphasis on aircraft diagnosis/prognosis, predictive maintenance, feature selection, and flight data monitoring
(FDM). This study highlights the potential use of these algorithms in enhancing the efficacy and effectiveness of various
aircraft operations. In the field of aviation prognosis and diagnosis, many designs have been acknowledged as efficient for
defect identification, calculation of remaining usable life, and prediction of excessive vibration in aero-engines. The
architectural models discussed in this paper include deep autoencoders, deep belief networks, long short-term memory
networks, and convolutional neural networks. The use of feature selection and scalar feature selection methodologies has
been seen to augment the efficacy of FDM (Feature Detection and Matching) algorithms by means of identifying
noteworthy features and detecting highly linked features. The application of machine learning algorithms in the domain of
predictive maintenance enables real-time assessment of equipment health, hence reducing possible hazards and improving
overall equipment performance. The research results emphasize the importance of flight data monitoring in improving
safety and operational efficiency in the field of civil aviation. The application of machine learning approaches, namely
classification algorithms, facilitates the analysis of flight data for the aim of identifying unsafe behaviors or violations from
established operational standards.

Keywords – Artificial Intelligence, Machine Learning, Deep Learning, Aircraft Prognosis Diagnosis, Feature Selection,
Predictive Maintenance, Flight Data Monitoring.
I. INTRODUCTION
The aviation industry has seen notable technological breakthroughs that have had a deep influence on its significance within
the global economy and its capacity to facilitate other sectors. These inventions have played a vital role in enhancing
efficiency and operating capacities, tackling air traffic control obstacles, advancing materials technology, advocating for
sustainable fuels, establishing digital systems, and resolving environmental issues. The incorporation of aerospace robotics,
artificial intelligence, and cyber-physical systems (CPS) has expedited the acceptance of automated decision-making
procedures and the progressive shift towards reliable autonomous operations in both civil and military contexts. Within the
realm of commercial aviation, CPS technologies are now being used to facilitate the advancement of aircraft that may be
handled by a single pilot. In this context, the traditional co-pilot role may be substituted by either a digital assistant or a
remote pilot situated on the ground. The One-to-Many (OTM) concept enables a solitary remote pilot to concurrently operate
several human and unmanned spacecraft.
Artificial intelligence (AI) has emerged as a crucial component in the fields of big data and data mining analysis. It
facilitates the accomplishment of many tasks such as categorization, planning, prediction, optimization, diagnosis,
computing, and consumer analysis. Neural networks, deep learning, and machine learning are computational techniques that
use algorithms to analyze extensive datasets, enabling the generation of desired results and the identification of trends,
patterns, and forecasts. Virtual assistants driven by AI, such as Amazon Alexa, Google Assistant, Apple Siri, and Cortana,
have become indispensable in everyday routines, with an estimated 27% of individuals using these technologies [1]. Francis,
Bernard, Nowak, Daniel, and Bernard [2] have conducted thorough investigations and evaluations regarding the efficacy
and operational capabilities of diverse virtual assistants that are now accessible in the marketplace.
Machine learning (ML) is a specialized domain within the study of AI that is primarily concerned with the improvement
of techniques and algorithms that enable computers to obtain knowledge and enhance their effectiveness through experience,

90
ISSN: 3005-9852 Journal of Robotics Spectrum 1(2023)

without requiring explicit programming instructions. ML methodologies may be classified into several categories, including
reinforcement learning, supervised learning, semi-supervised learning, and unsupervised learning. The objective of
unsupervised learning is to identify patterns within a dataset by leveraging the similarities that exist between individual data
points. On the other hand, semi-supervised learning involves the use of both labeled and unlabeled data in order to make
predictions or uncover patterns. Supervised learning is a ML approach that use training data with known labels to establish
a mapping between input and output variables. Through an iterative process, the model is continuously refined until it attains
the required degree of accuracy.
Deep learning (DL), which falls under the umbrella of ML and AI, is recognized as a fundamental technology of the
Fourth Industrial Revolution (4IR). It finds extensive application across diverse sectors like healthcare, image identification,
text analytics, and cybersecurity. Deep learning models, like machine learning models, undergo comparable processing
phases; however, they possess the capability to automate the process of feature extraction. Although the existing body of
research on aviation prediction and diagnosis is small, scholars have identified four primary architectures for this purpose:
Deep Autoencoders (DAE), Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), and Deep Belief
Networks (DBN). The process of input data mapping into a compressed representation is a characteristic of the DAE, while
the LSTM is a kind of recurrent neural network specifically intended to capture and understand long-term relationships
within data. CNNs and DBNs are widely used neural network architectures within the field of aircraft maintenance. These
networks are commonly employed for activities such as the classification of defects and the evaluation of the remaining
useable life of aircraft systems.
The utilization of feature selection (FS) and feature extraction (FE) algorithms assumes a pivotal function in the
integration of deep learning (DL) and machine learning (ML) methodologies in the aerospace industry. Feature selection
(FS) is a computational procedure that involves the automatic identification and extraction of relevant properties from a
provided dataset. Conversely, feature extraction (FE) approaches result in new features through transformation or
combination on the original feature set. They are aimed at improving flight data monitoring algorithms that play an important
role in increasing the efficiency and safety of aircraft operations. Predictive maintenance (PdM) is an advanced methodology
that utilizes sophisticated analytic techniques such as machine learning to assess the operational condition of machinery
through the analysis of collected data on a regular basis. Predictive maintenance (PdM), which improves equipment’s
efficiency as well as its lifespan by actively predicting future possible situations and promptly rectifying any faults that might
arise, successfully manages risks and minimizes unnecessary maintenance costs. The examination of fleet-wide aircraft
utilization information using machine learning technology holds promise in predicting the internal load distribution of
aircraft structures. The utilization of this data analytics methodology facilitates accurate predictions of the loading conditions
experienced by the structure, offering notable benefits to the civil aviation industry.
The FDM is a critical practice for airlines, as it plays a crucial role in improving safety and maximizing operating
efficiency. In the past, flight data analysis predominantly utilized statistical methodologies. The advent of computing
advancements, namely in the realm of machine learning, has presented new opportunities for the analysis of flight data.
Machine learning algorithms are very suitable for the classification of flights into safe and risky categories, as well as for
conducting comprehensive analyses of flight data to elucidate the underlying factors contributing to unsafe flights. The use
of AI, ML and DL technologies within the aviation domain has brought about significant transformations across several
facets of the sector, including aircraft operations, maintenance, safety, and efficiency. The aforementioned technologies has
the capability to augment predictive maintenance, increase flight data analysis, and optimize aircraft performance, hence
yielding advantages for the whole of the aviation ecosystem.
This study examines the influence of technical advancements on the aviation industry and their role in bolstering its
significance within the global economy. This study focuses on the utilization of cutting-edge technologies, including AI,
ML, and DL, to enhance efficiency and operational capabilities within the aviation industry. The study also examines the
use of these technologies in domains such as air traffic management, aircraft maintenance, and flight data analysis. This
paper examines the possible advantages and obstacles associated with the use of these technologies within the aviation sector.
The rest of the paper has been organized as follows: Section II presents an overview of advanced technologies such as
artificial intelligence, deep learning, and machine learning. Section III presents a detailed discussion of the application of
enhanced aircraft maintenance, and flight data analysis. In this section, concepts such as feature selection, predictive
maintenance, and flight data monitoring, are critically analyzed. Lastly, Section IV presents a conclusion to the research and
presents a proposal of future research.

II. OVERVIEW OF ADVANCED TECHNOLOGIES

Artificial Intelligence
Artificial intelligence (AI) is extensively employed in various domains for the purpose of addressing computation,
classification, diagnosis, planning, optimization problems, prediction, as well as analyzing and collecting customer
information. This enables the extraction of valuable insights into customer preferences and requirements, which can
subsequently inform decision-making processes. The objective of this article is to provide a detailed overview of the latest
advancements in the area of analysis and provide valuable insights for academics actively engaged in studying algorithms
and applications of AI. Big data and data mining are pervasive in several domains, making it crucial to effectively maintain
the vast volume of created data to ensure no valuable information is overlooked. The use of artificial intelligence is often

91
ISSN: 3005-9852 Journal of Robotics Spectrum 1(2023)

employed for the purpose of processing this particular category of data. Artificial Intelligence (AI) and its several subfields,
such as ML, DL, and Neural Networks, are fundamentally rooted in algorithmic principles. Algorithmic approaches are often
used on vast quantities of data, sometimes referred to as Big Data, in order to generate desired outcomes and identify trends,
patterns, and forecasts.
Artificial Intelligence (AI) facilitates the execution of complex analytical operations on Big Data at a speed beyond
human conception. In their study, Zhang, Patuwo, and Hu [3] constructed an artificial neural network using a systematic
scientific approach that focuses on optimizing a criteria often known as the learning rule. The training data, which includes
input and output information, plays a crucial role in neural networks as it facilitates the acquisition of essential data necessary
for achieving optimal performance. Moreover, the nonlinearity inherent in neural network processing components
contributes to the versatility of the system. One commonly accepted definition is on the comparison of the intelligence shown
by computer devices with that of the general population. An alternative interpretation pertains to the concern of the
performance of machines, which has historically been associated with the realm of intelligence.
In the contemporary day, virtual assistants have assumed a pivotal role in facilitating many daily tasks and responsibilities
of individuals. Based on the 2018 study report conducted by Këpuska and Bohouta [4], it was found that a significant
proportion, namely 27%, of individuals use AI-powered virtual assistants like Apple Siri, Cortana, Amazon Alexa, and
Google Assistant,. These virtual assistants are mostly employed for the execution of basic tasks, with a particular emphasis
on those created with natural language processing capabilities. Pot et al. [5] have conducted research and examined the
operational framework and effectiveness of several virtual assistants now present in the marketplace. The intelligent virtual
assistant developed by Page and Gehlbach [6] was meant to be seamlessly incorporation with Google virtual services and
operate using the Google virtual assistant interface. The effectiveness of the proposed virtual assistant is calculated by
inputting a comparative study of message and traffic transmission, together with the duration of discussion, over a period of
around three days.
The architecture of the virtual assistant is depicted in Fig 1. It illustrates the sequential flow of the system, beginning
with the user input. Subsequently, the system determines the appropriate conversation strategy module to employ, which is
a response generated by the dialog management module. Concurrently, the classification module interacts with the NLP
module to generate a response. The conversation history database is used for the analysis of the knowledge base development
module, which subsequently provides responses to the domain knowledge base, as elaborated in Fig 1.

Fig 1. Proposed Architecture for Conversational Agents

Machine Learning
Machine Learning (ML) is a specialized domain within the study of AI that centers on the improvement of models and
algorithms that enable computers to acquire knowledge and develop performance through experience, without explicit
programming for certain tasks. Machine learning methodologies may be categorized into four main types: reinforcement,
unsupervised, supervised, and semi-supervised learning, as shown in Fig 2. Unsupervised learning involves the use of
unlabeled data. The primary objective of the machine learning model is to identify previously unidentified patterns within
the dataset, often via the identification of similarities between individual data points.
Algorithms are therefore designed in a manner that enables them to autonomously identify patterns and structures within
the data. Semi-supervised learning involves using a combination of unlabeled and labeled data pieces as input. Supervised
learning involves the use of machine learning models that rely on training data that has been labeled. In this context, the
model is assigned appropriate labels for the desired outcome and endeavors to acquire knowledge of the relationship between
inputs and outputs, often via iterative adjustments. The procedure is iteratively performed until the model attains a

92
ISSN: 3005-9852 Journal of Robotics Spectrum 1(2023)

predetermined degree of precision on the training dataset and is capable of accurately forecasting the outcomes for novel
occurrences.
In the context of machine learning, reinforcement learning is a methodology that leverages a trial and error approach,
whereby the agent explores various behaviors and balances the trade-off between exploitation and exploration. The primary
objective is to identify and choose actions that result in the highest possible rewards. In their study, Susto, Schirru, Pampuri,
McLoone, and Beghi [7] examined several scholarly contributions pertaining to the field of predictive maintenance (PdM)
in conjunction with deep reinforcement learning, which involves the integration of reinforcement learning (RL) techniques
with deep learning methodologies. However, in accordance with our established search criteria for machine learning-based
predictive maintenance in automotive systems, it was observed that none of the examined publications included
reinforcement learning techniques.

Fig 2. Common Classification of Machine Learning using Fig 3. The most Essential Machine Learning tasks for
the ML tasks PdM.

The use of machine learning applications is seeing a notable surge and has already shown substantial advantages across
several sectors of the business, including PdM (see Fig 3) [8]. These advantages include various fields, including but not
limited to traffic forecasting, logistics optimization, and cancer treatment. Consequently, intelligent algorithms are
increasingly being embraced as a means to effectively address chronic difficulties faced by society. Given the growing
demand for air travel and the intensifying competition within the industry, it has become imperative for makers of
commercial aircraft to actively embrace the digital revolution.

Deep Learning
Deep learning (DL), which falls under the umbrella of ML and AI, is widely recognized as a fundamental technology in the
context of the Industry 4.0. DL technology, derived from ANN, has gained significant attention in the field of computers
due to its capacity to learn from data. It is widely employed in diverse domains like cybersecurity, healthcare, text analytics,
image identification, and others. In contemporary times, the phrases AI, ML, and DL are often used interchangeably to
denote systems or software that exhibit intelligent behavior. Fig 4 depicts the relative positioning of DL in relation to most
learning algorithms. Based on the information shown in Fig 5, it can be inferred that DL is included under both ML and the
broader field of AI. In broad terms, AI is the integration of human behavior and intellect into computers or systems. On the
other hand, ML is the approach used to acquire knowledge from data or experience, therefore enabling the automated
construction of analytical models.
Deep learning models generally adhere to the same sequential processing phases as machine learning modeling. Figure
4 illustrates a deep learning pipeline designed to address practical challenges. This workflow encompasses three distinct
stages: data comprehension and preprocessing, deep learning model development, and training, as well as validation and
interpretation. In contrast to ML modeling, the DL model automates the process of feature extraction instead of relying on
operator intervention. Despite the limited amount of research completed in the field of aircraft prognosis/diagnosis, a study
conducted by Khan et al. [9] has shown the use of four primary architectures: DAE, CNN, LSTM, and DBN. Table 1
presents a comprehensive overview of the existing research conducted in the field.
The Deep Autoencoders (DAE) model, as described by Zhou and Paffenroth [18], is a neural network architecture that
employs a mapping function to transform input information into a compressed representation, which is then decoded to
provide an approximation of the original input. The aforementioned procedure compels the encoder to decrease the
dimensionality of the data, and under certain circumstances, it acquires the ability to disregard noise. The process of
codification entails the condensation of the inherent features of the input data. The LSTM is a specific kind of RNN that is
structured with chain units including output, forget, and input gates. A sigmoid function is applied to each gate. The input
gate regulates the impact of the present input. The forget gate mechanism included in each unit is responsible for regulating
the amount of information that should be preserved. The output gate is responsible for determining whether the flow of
information will be sent to the subsequent LSTM unit. This architectural design facilitates the acquisition of extended
temporal relationships within datasets. LSTM is being used in the aerospace Maintenance, Repair, prognosis of excessive

93
 ISSN: 3005-9852 Journal of Robotics Spectrum 1(2023)

vibration within aero-engines, Overhaul (MRO) sector for the purposes of fault identification, and calculation of remaining
usable life.
Table 1. Deep Learning Architectures for Diagnosing and Prognosticating Aircraft
Literature Architecture Application
Gao, Ma, Song, and Liu Deep Quantum-based Deep Belief Fuel system fault diagnostics in an aircraft
[10] Networks and Neural Network
Fan, Ding, Zheng, Xiao, Multi-objective Deep Belief Network Identification of faults and assessment of an
and Ai [11] Ensemble aero-engine's remaining usable life
Abdel-Zaher and Eldeib Deep Belief Network Aero-engineering classification of health
[12] conditions
Lawrence, Zhang, Lim, and Convolutional Neural Network and Diagnostics for rolling bearing faults
Phillips [13] Particle Swarm Optimization
ElSaid, Jamiy, Higgins, Long Short-Term Memory and Ant Aero-engine prognosis for excessive vibration
Wild, and Desell [14] Colony Optimization
Ji, Xu, Yang, and Yu [15] Convolutional Neural Network Aero-engine fault diagnosis and remaining
usable life calculation
Zhang, Xiong, He, and Long Short-Term Memory Identification of faults and assessment of an
Pecht [16] aero-engine's remaining usable life
Lai, Chen, Wang, Lu, Tsao, Deep denoising autoencoder Life forecasting in combined modular
and Lee [17] avionics

Fig 4. An Illustration of DL Position, Comparing with AI Fig 5. Performance Measures Between ML and other ML
and ML Algorithms
The CNNs are a specific type of neural network structure characterized by the inclusion of convolutional layers featuring
nonlinear activation functions. These convolutional layers are subsequently followed by a fully connected layer, which is
responsible for computing the network's outputs. The incoming data undergoes sequential processing through many
convolution layers, with each layer employing unique filters. Ultimately, the outputs generated by these layers are combined
in order to obtain a conclusive outcome. This enables CNN to achieve improved performance when applied to data sets
characterized by a high degree of spatial correlation with neighboring data points. DBN are used in the aerospace industry
for several applications, including aero-engine Remaining Useful Life (RUL) estimate, aero-engine fault classification, and
aviation fuel system problem detection. The DBN, also known as a deep belief network, is a kind of generative graphical
model that incorporates stacked RBMs (Restricted Boltzmann Machines).
Each RBM consists of a hidden layer and a visible layer, with links established between these layers but not inside each
individual layer. The DBN is trained based on the application of the greedy layer-wise unsupervised learning approach in
order to extract properties from the input information. In their study, Lee, Grosse, Ranganath, and Ng [19] use a DBN
classifier to accurately determine the health condition of an aero-engine. The C-MAPSS data is used. The used DBN
classifier has three hidden layers. The conjugate gradient algorithm is employed to refine the DBN classifier by fine-tuning,
subsequent to its pre-training and training phases. The comparison of DBN fault classification in aero-engines is conducted
with other methods such as Mahalanobis Distance, SVM, Self-Organizing Maps, and Backpropagation Neural Network
(BNN). The results indicate that DBN demonstrates the highest accuracy in fault classification for five out of the six
operational circumstances.

III. APPLICATION OF ENHANCED AIRCRAFT MAINTENANCE AND FLIGHT DATA ANALYSIS

The aerospace sector and aviation management have seen a diverse range of applications pertaining to ML, AI, and DL. This
section presents a discussion of some of these applications:

Feature Selection
Flight data comprises a multitude of flight characteristics, such as altitude and estimated air speed, among others. The
analysis of flight data and the detection of anomalies heavily rely on these characteristics. However, it should be noted that

94
 ISSN: 3005-9852 Journal of Robotics Spectrum 1(2023)

not all of these metrics have equal importance, and some parameters are deemed more significant, resulting in their frequent
recording. In the present state of the aviation sector, professionals make determinations about the essential factors that pertain
to Flight Data Monitoring (FDM). Hence, the selection of appropriate parameters has always been seen as an issue of prudent
decision-making. Inefficient FDM analysis may result from inadequate parameter selection in the first stages. A more
effective approach involves using a learning algorithm that has the capacity to determine the optimal set of parameters for a
given circumstance. This approach will facilitate the development of a FDM analysis that has the ability to effectively handle
and adjust to various forms of data. The preselection of parameters has the potential to introduce bias into the learning
process, resulting in worse performance in anomaly detection compared to allowing the machine, or learning algorithm, to
make the parameter decision.
The phenomenon in which a learning system autonomously selects properties from a given dataset is often referred to as
Feature Selection (FS). According to Zebari, Abdulazeez, Zeebaree, Zebari, and Saeed [20], feature reduction or selection
is the process of identifying the most significant characteristics from a given set of features. The objective is to minimize the
quantity of parameters while preserving a significant amount of their class discriminating information. Feature Extraction
(FE) algorithms, as defined by Pohjalainen, Räsänen, and Kadıoğlu [21], include a variety of techniques that generate
additional features by applying modifications or combinations to the original feature selection. This method has the potential
to decrease the expenses associated with recognition by minimizing the amount of data that must be gathered. Additionally,
in some scenarios, it may provide improved classification accuracy as a result of finite sample size effects. Functional size
(FS) aids in the process of generalizing performance by using more computationally efficient approaches and identifying
crucial characteristics.
The study referenced by Farrow, Zhang, Szabó, Torchia, and Kay [22] employs the Spectral Density Function (SDF)
method, whereas utilizes Kalman filtering, vector feature selection, and scalar feature selection techniques. In [23], Tuncer
uses SDF, a reliable technique for extracting time-series features, to improve the performance of pattern classification. The
process has four primary stages:
a) Dividing the time series data sets into segments and creating symbol sequences.
b) Using the corresponding symbol sequences to build Probabilistic Finite State Automata (PFSA).
c) Extracting features from PFSA as state probability vectors or probability matrices.
d) Pattern categorization using the characteristics that were extracted.
Algorithms developed within the SDF framework shown enhanced performance in terms of early anomaly identification
and resistance to measurement noise when compared to other methodologies such as Bayesian approaches, NN, and Principal
Component Analysis (PCA). The feature extraction approach based on Spectral Density Function (SDF) exhibits sensitivity
to signal distortions while also demonstrating robustness to spurious signals and measurement noise. Additionally, it exhibits
adaptability in the context of low-resolution sensing as a result of the spatial partitioning's coarse graining. In the study
conducted by Bolón-Canedo, Sánchez-Maroño, and Alonso-Betanzos [24], the use of scalar feature selection procedure is
employed to effectively choose features that have been found by the correlation coefficient and ambiguity function approach.
The ambiguity function is used for the purpose of analyzing the degree of overlap among features, hence facilitating the
reduction of feature count. On the other hand, the correlation coefficient serves the purpose of identifying characteristics that
exhibit a high degree of correlation.
Highly connected features are substituted with the most important characteristic. Two flights from distinct classes, such
as anomalous and normal, are considered. A parameter that has a normal distribution is selected as the preferred choice. The
variance and mean are calculated for the two flights. Statistical tests, such as the t-test, are used for the analysis of this
particular characteristic [29]. Following the completion of the test, if the two means fall within the same interval of
importance, they are deemed to be equivalent. Consequently, the characteristic under consideration does not contribute any
discriminating data between the two flights and is then excluded. The subset of features that has been reduced is then
subjected to a vector feature selection method [30]. The primary aim of this data processing task is to identify and choose
the most effective characteristics that result in the highest level of differentiation among the various flights within the dataset.
Various approaches, including Branch and Bound, Floating Sequential search, and Backward & Forward selection, are used
to exhaustively explore all possible combinations of characteristics [31]. An exhaustive examination of these methodologies
beyond the boundaries of this research article [32].
The Kalman filter is used to assess the significance of individual features in distinguishing different flights by classifying
samples with low probability density function values as atypical. The extent to which FE methods have been used in
implementing the FDM is currently restricted. Consequently, there is untapped potential in exploring the application of
Independent Component Analysis (ICA), PCA, and NN for FDM experimentation. The study conducted in [25] has shown
that the use of SDF (Statistical Detection Framework) has resulted in significant improvements in the identification of
abnormalities seen in actual flight data.
Predictive maintenance (PdM) is an advanced approach that leverages condition-based monitoring techniques to enhance
the efficiency and longevity of equipment. It does this by continuously evaluating the health of the tool in real-time. Through
the use of sensor data collection and the application of sophisticated analytical methodologies, such as machine learning
(ML), the practice of predictive maintenance may effectively identify, detect, and promptly treat problems as they arise.
Furthermore, this approach enables the prediction of probable future states of equipment, therefore mitigating risk. The
determination of maintenance methods and maturity is contingent upon several aspects, including the cost of assets and their

95
 ISSN: 3005-9852 Journal of Robotics Spectrum 1(2023)

potential replacement, the criticality of such assets, the patterns of use, and the consequences of failure on safety,
environmental impact, operational efficiency, financial stability, and public perception.

Predictive Maintenance

Fig 6. Application of ML in the Predictions of Aircraft Fatigue

Predictive maintenance is recognized as one of the primary maintenance tactics used by enterprises, alongside reactive
maintenance, which addresses problems as they arise, and preventive maintenance, which depends on a predetermined
maintenance plan to detect defects. Predictive maintenance, due to its proactive nature, improves the effectiveness of
preventive maintenance by offering ongoing insights into the real-time status of equipment, as opposed to depending only
on anticipated conditions derived from previous data. Predictive maintenance entails the execution of corrective maintenance
only when it is deemed required, hence circumventing the accumulation of superfluous maintenance expenses and the
occurrence of machine downtime. The practice of predictive maintenance involves using past time series data and failure
records to forecast the prospective condition of equipment and proactively identify possible issues. This allows firms to
enhance maintenance scheduling and enhance dependability.
In order to ensure the provision of customized aircraft maintenance solutions that prioritize safety while also optimizing
scheduling, a crucial need is a comprehensive understanding of actual aircraft use in practical settings. By utilizing
comprehensive data on aircraft usage parameters across the entire fleet, such as flight-by-flight recordings and employing
validated machine learning algorithms and meteorological conditions, it becomes feasible to make highly precise predictions
regarding the internal loading factors of a structure. These predictions are based on the recorded and measured aircraft
parameters, including accelerations, speed, flight configurations, and altitude. One of the primary advantages of using data
analytics is the potential to accurately replicate the authentic loading sequence encountered by the structure. In order to
comprehensively substantiate any findings derived from ML, it is essential to acknowledge the need of including
supplementary information within the framework of the study. This may include several forms of data, such as extra in-
service data, theoretical studies, comprehensive test results, and so on.
The potential capabilities of ML applications in predicting the internal load distribution of aircraft structures have
significant implications. This ability enables the comparison and linkage of practical aircraft usage with that of an initial or
average fleet assumptions. Unlocking this potential can bring about transformative benefits in the field of civil aviation. In
order to provide predictions that are relevant and trustworthy, a comprehensive dataset serves as the primary need. In a broad
sense, the majority of puzzle components, including this one, are now accessible. The primary hurdle for machine learning
is in effectively integrating and validating these components in a cohesive manner, hence augmenting analytical prowess
and acquiring novel predictive capacities. The use of machine learning in predicting fatigue stress on aircraft is shown in
Fig 6.

Flight Data Monitoring

FDM is a practice undertaken by airlines largely for the purpose of enhancing and monitoring the operational and safety
aspects of their aircraft. The flight data recorder on an aircraft captures and stores data, which is subsequently extracted and
examined using a range of tools and methodologies. The primary aim of this analysis is to enhance civil aviation operations
by implementing improvements in maintenance schedules, pilot training, and operational procedures, while ensuring safety
remains uncompromised. The primary aims of the FDM are as follows:
a) The system has the capability to identify technological deficiencies, hazardous practices, or deviations from
planned operational protocols in their initial stages, hence reducing the risk of possible mishaps or accidents.

96
 ISSN: 3005-9852 Journal of Robotics Spectrum 1(2023)

b) The provision of an objective tool for monitoring and evaluating the effectiveness of remedial measures is
facilitated by the implementation of enhanced training programs or modifications to Standard Operating
Procedures.
The advantages of FDM are often emphasized in accident investigations. An illustrative instance is the accident
involving Gulf Air's A320 aircraft GF072 on August 23, 2000, near Bahrain International Airport [26]. According to the
conclusive report of the Accident Investigation Board (AIB), it was determined that the flight data analysis system was
experiencing unsatisfactory performance at the time of the accident [27]. The airline's lack of access to flight data analysis
resulted in the deprivation of a crucial tool for safety analysis. Upon the retrieval of flight data from the aircraft, a
comprehensive analysis of the whole period of the flight, referred to as a timeline, is conducted. Historically, the FDM
methodology has included the use of statistical tools for the purpose of data analysis.
The airline uses specialized software to analyze flight data, which identifies instances when certain aspects of the flight
data have beyond predetermined thresholds. This strategy is based on established "exceedances," which are predetermined
concerns. Events are occasionally described as occurrences in which there is a deviation from the predetermined schedule.
The conventional algorithmic approach for identifying exceedance or incidents is analyzing the data for any deviations from
the specified limits specified in the flight manual, standard operating procedures (SOPs), and principles of sound airmanship
[28]. This methodology relies on the proficiency of human experts in constructing a framework that use pre-determined
thresholds to detect known safety issues through the assessment of a restricted set of variables. A solitary event is constituted
by many instances that are captured in a consecutive manner.
The process of event detection might be laborious. Nevertheless, the progress in computer methodologies has presented
novel opportunities for the analysis and understanding of flight data. The utilization of Machine Learning (ML)
methodologies has been crucial in driving progress within the banking and online gambling industries. ML is a subfield of
AI that focuses on the creation and advancement of computer systems with the ability to access data, recognize patterns,
acquire knowledge, and enhance performance through experiential learning, without requiring direct human intervention.
The utilization of computational methodologies holds the promise of substantially augmenting the efficiency and efficacy
of the aviation industry.
The aviation industry operates within a limited financial margin. Despite the ongoing expansion of air travel, the aviation
industry remains susceptible to external factors, particularly fluctuations in oil prices. Hence, the thorough analysis of flight
data has the capacity to optimize flight operations, leading to a reduction in fuel consumption, diminished maintenance and
insurance costs, and an enhanced level of safety. However, the traditional statistical techniques employed for analyzing flight
data, which rely on pre-established criteria, are inadequate in delivering full information and are also burdensome. This
constraint may be mitigated by using machine learning methods. The aforementioned approaches provide the capability to
conduct comparisons of flight data parameters over a significant number of flights, hence enabling the identification of novel
or unfamiliar patterns. These observed trends may indicate atypical or incongruous conduct in relation to the majority of
flights. The examination of outliers is a subject of interest that requires more inquiry.
An illustration of an atypical flight trajectory occurs during the descent phase when the aircraft deviates from the
established protocol for a stable approach, specifically by failing to deploy the landing gear by the time the altitude reaches
1000 feet. Another illustration of an atypical flying occurrence may be an extreme inclination of the aircraft's nose during
the process of taking off. In conventional approaches to flight data analysis, an aircraft's pitch exceeding a predetermined
threshold is indicated by a red flag. However, machine learning algorithms have the capability to identify aberrant pitch
deviations, as seen in Fig 7.

Fig 7. The Angle of Aircraft Pitch During Take-Off

Contemporary machine learning (ML) algorithms are highly suitable for not only classifying flights as safe or unsafe,
but also for conducting in-depth analysis of flight particulars and providing comprehensive explanations for the factors
97
 ISSN: 3005-9852 Journal of Robotics Spectrum 1(2023)

contributing to flight unsafety. The advancement of machine learning algorithms has led to increased efficiency in analyzing
flight data, making it more appropriate for generating predictions based on such data. The following sections delineate the
machine learning strategies that have been used for FDM. The last section of the paper encompasses the conclusion, when
the main findings and implications are summarized, and the subsequent discussion of potential avenues for further research.

IV. CONCLUSION AND FUTURE SCOPE

In summary, the aviation industry has seen notable technological progress, namely in the domains of AI and ML. The
advancements have significantly influenced many facets of the field, including air traffic management, aircraft upkeep, and
analysis of flight data. The use of AI and ML algorithms has significantly enhanced the efficiency and operational capacities
within the aviation industry. The potential to transform commercial aviation exists with the development of single-pilot
controlled aircraft, aided by digital assistants and remote pilots stationed on the ground. The operational concept known as
One-to-Many (OTM) enables a solitary remote pilot to exercise control over several manned and unmanned vehicles,
therefore enhancing operational efficiency and yielding cost reductions.
Artificial intelligence (AI) and ML methodologies have been used in the analysis of flight data, particularly within the
domain of Flight Data Monitoring (FDM). Historically, the field of FDM has mostly used statistical methodologies to
evaluate flight data and identify departures from pre-established thresholds. Nevertheless, the emergence of machine learning
algorithms has opened new possibilities for the study of flight data. Machine learning algorithms have the capability to not
only categorize flights as either safe or hazardous, but also provide comprehensive analysis into the underlying factors
contributing to risky flights. Moreover, the implementation of predictive maintenance (PdM) has emerged as an essential
strategic approach within the aviation sector. Using sophisticated analytical tools and machine learning algorithms, the
practice of PdM enables the ongoing evaluation of equipment health in real-time. This approach facilitates proactive
maintenance strategies, hence enhancing equipment performance and mitigating the likelihood of failures. This methodology
improves the effectiveness of preventative maintenance by offering ongoing observations on the real-time state of the
equipment, hence mitigating superfluous repair expenses, and reducing machine downtime.
The aviation industry has significant potential for future breakthroughs in AI and ML. The advancement of AI-driven
virtual assistants and the incorporation of AI technology into air traffic control systems have the potential to augment
operational efficiency and safety measures. Furthermore, the investigation of novel machine learning methodologies, such
as deep learning, has the potential to enhance predictive capacities in the domains of aircraft maintenance and flight data
analysis. Nevertheless, there exist several issues that require attention and resolution in the next period. The incorporation
and verification of extensive datasets for machine learning applications continue to pose a substantial challenge.
Furthermore, it is imperative to prioritize the establishment of robust security measures and dependable protocols for AI and
ML systems within the aviation industry. This is essential to uphold the safety and overall soundness of aviation operations.
The aviation industry has adopted technical advancements in AI and ML, resulting in notable enhancements in efficiency,
safety, and operational capacities. The prospective developments in virtual assistants, air traffic control systems, predictive
maintenance, and flight data analysis indicate a significant future scope for these technologies in the sector. Nevertheless, it
is essential to acknowledge that the effective application of these technologies in the aviation industry will heavily rely on
the resolution of obstacles pertaining to data integration, security, and dependability.

Data Availability
No data was used to support this study.

Conflicts of Interests
The author(s) declare(s) that they have no conflicts of interest.

Funding
No funding agency is associated with this research.

Ethics Approval and Consent to Participate

The research has consent for Ethical Approval and Consent to participate.

Competing Interests
There are no competing interests.

References
[1]. T. Bolton, T. Dargahi, S. Belguith, M. Al-Rakhami, and A. H. Sodhro, “On the Security and Privacy Challenges of Virtual Assistants,” Sensors,
vol. 21, no. 7, p. 2312, Mar. 2021, doi: 10.3390/s21072312.
[2]. E. Francis, S. A. Bernard, M. L. Nowak, S. R. Daniel, and J. A. Bernard, “Operating room virtual reality immersion improves Self-Efficacy
amongst preclinical physician assistant students,” Journal of Surgical Education, vol. 77, no. 4, pp. 947–952, Jul. 2020, doi:
10.1016/j.jsurg.2020.02.013.
[3]. G. Zhang, B. E. Patuwo, and M. Y. Hu, “Forecasting with artificial neural networks:,” International Journal of Forecasting, vol. 14, no. 1, pp.
35–62, Mar. 1998, doi: 10.1016/s0169-2070(97)00044-7.

98
ISSN: 3005-9852 Journal of Robotics Spectrum 1(2023)

[4]. V. Këpuska and G. Bohouta, “Next-generation of virtual personal assistants (Microsoft Cortana, Apple Siri, Amazon Alexa and Google Home),”
2018 IEEE 8th Annual Computing and Communication Workshop and Conference (CCWC), Jan. 2018, doi: 10.1109/ccwc.2018.8301638.
[5]. M. Pot et al., “Effectiveness of a Web-Based tailored intervention with virtual assistants promoting the acceptability of HPV vaccination among
mothers of invited girls: randomized controlled trial,” Journal of Medical Internet Research, vol. 19, no. 9, p. e312, Sep. 2017, doi:
10.2196/jmir.7449.
[6]. L. C. Page and H. Gehlbach, “How an artificially intelligent virtual assistant helps students navigate the road to college,” AERA Open, vol. 3,
no. 4, p. 233285841774922, Oct. 2017, doi: 10.1177/2332858417749220.
[7]. G. A. Susto, A. Schirru, S. Pampuri, S. McLoone, and A. Beghi, “Machine Learning for Predictive Maintenance: A multiple classifier approach,”
IEEE Transactions on Industrial Informatics, vol. 11, no. 3, pp. 812–820, Jun. 2015, doi: 10.1109/tii.2014.2349359.
[8]. S. Zander, T. Nguyễn, and G. Armitage, “Automated traffic classification and application identification using machine learning,” The IEEE
Conference on Local Computer Networks 30th Anniversary (LCN’05)L, Jan. 2005, doi: 10.1109/lcn.2005.35.
[9]. K. Khan et al., “Recent trends and challenges in predictive maintenance of aircraft’s engine and hydraulic system,” Journal of the Brazilian
Society of Mechanical Sciences and Engineering, vol. 43, no. 8, Aug. 2021, doi: 10.1007/s40430-021-03121-2.
[10]. Z. Gao, C. Ma, D. Song, and Y. Liu, “Deep quantum inspired neural network with application to aircraft fuel system fault diagnosis,”
Neurocomputing, vol. 238, pp. 13–23, May 2017, doi: 10.1016/j.neucom.2017.01.032.
[11]. C. Fan, C. Ding, J. Zheng, L. Xiao, and Z. Ai, “Empirical Mode Decomposition based Multi-objective Deep Belief Network for short-term
power load forecasting,” Neurocomputing, vol. 388, pp. 110–123, May 2020, doi: 10.1016/j.neucom.2020.01.031.
[12]. A. M. Abdel-Zaher and A. M. Eldeib, “Breast cancer classification using deep belief networks,” Expert Systems With Applications, vol. 46, pp.
139–144, Mar. 2016, doi: 10.1016/j.eswa.2015.10.015.
[13]. T. Lawrence, L. Zhang, C. P. Lim, and E.-J. Phillips, “Particle swarm optimization for automatically evolving convolutional neural networks
for image classification,” IEEE Access, vol. 9, pp. 14369–14386, Jan. 2021, doi: 10.1109/access.2021.3052489.
[14]. A. ElSaid, F. E. Jamiy, J. Higgins, B. Wild, and T. Desell, “Optimizing long short-term memory recurrent neural networks using ant colony
optimization to predict turbine engine vibration,” Applied Soft Computing, vol. 73, pp. 969–991, Dec. 2018, doi: 10.1016/j.asoc.2018.09.013.
[15]. S. Ji, W. Xu, M. Yang, and K. Yu, “3D convolutional neural networks for human action recognition,” IEEE Transactions on Pattern Analysis
and Machine Intelligence, vol. 35, no. 1, pp. 221–231, Jan. 2013, doi: 10.1109/tpami.2012.59.
[16]. Y. Zhang, R. Xiong, H. He, and M. Pecht, “Long Short-Term Memory recurrent neural network for remaining useful life prediction of Lithium-
Ion batteries,” IEEE Transactions on Vehicular Technology, vol. 67, no. 7, pp. 5695–5705, Jul. 2018, doi: 10.1109/tvt.2018.2805189.
[17]. Y. Lai, F. Chen, S.-S. Wang, X. Lu, Y. Tsao, and C. Lee, “A deep denoising autoencoder approach to improving the intelligibility of vocoded
speech in cochlear implant simulation,” IEEE Transactions on Biomedical Engineering, vol. 64, no. 7, pp. 1568–1578, Jul. 2017, doi:
10.1109/tbme.2016.2613960.
[18]. C. Zhou and R. Paffenroth, “Anomaly Detection with Robust Deep Autoencoders,” KDD ’17: Proceedings of the 23rd ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, Aug. 2017, doi: 10.1145/3097983.3098052.
[19]. H. Lee, R. Grosse, R. Ranganath, and A. Y. Ng, “Convolutional deep belief networks for scalable unsupervised learning of hierarchical
representations,” ICML ’09: Proceedings of the 26th Annual International Conference on Machine Learning, Jun. 2009, doi:
10.1145/1553374.1553453.
[20]. R. R. Zebari, A. M. Abdulazeez, D. Q. Zeebaree, D. A. Zebari, and J. N. Saeed, “A comprehensive review of dimensionality reduction techniques
for feature selection and feature extraction,” Journal of Applied Science and Technology Trends, vol. 1, no. 2, pp. 56–70, May 2020, doi:
10.38094/jastt1224.
[21]. J. Pohjalainen, O. Räsänen, and S. Kadıoğlu, “Feature selection methods and their combinations in high-dimensional classification of speaker
likability, intelligibility and personality traits,” Computer Speech & Language, vol. 29, no. 1, pp. 145–171, Jan. 2015, doi:
10.1016/j.csl.2013.11.004.
[22]. N. A. Farrow, O. Zhang, A. Szabó, D. A. Torchia, and L. E. Kay, “Spectral density function mapping using 15N relaxation data exclusively,”
Journal of Biomolecular NMR, vol. 6, no. 2, pp. 153–162, Sep. 1995, doi: 10.1007/bf00211779.
[23]. E. Tuncer, “Extracting the spectral density function of a binary composite withouta prioriassumptions,” Physical Review B, vol. 71, no. 1, Jan.
2005, doi: 10.1103/physrevb.71.012101.
[24]. V. Bolón-Canedo, N. Sánchez-Maroño, and A. Alonso-Betanzos, Feature selection for High-Dimensional Data. Springer, 2015.
[25]. A. Khalil et al., “Efficient frameworks for statistical seizure detection and prediction,” The Journal of Supercomputing, vol. 79, no. 16, pp.
17824–17858, May 2023, doi: 10.1007/s11227-022-05030-0.
[26]. W. Tan and J. Yin, “An overview of paper documentation moving to Onboard Information System (OIS) for commercial aircraft,” in Lecture
Notes in Computer Science, 2020, pp. 116–126. doi: 10.1007/978-3-030-60128-7_9.
[27]. V. Foreman, F. Favarò, J. H. Saleh, and C. W. Johnson, “Software in military aviation and drone mishaps: Analysis and recommendations for
the investigation process,” Reliability Engineering & System Safety, vol. 137, pp. 101–111, May 2015, doi: 10.1016/j.ress.2015.01.006.
[28]. E. T. Martins, I. T. Martins, and M. M. Soares, “The encourage operators to promote manual flight operations- a pandemic in modern aviation,”
in Springer eBooks, 2014, pp. 317–325. doi: 10.1007/978-3-319-07635-5_31.
[29]. A. Agarkar, Md. Z. Hussain, J. T. Raja, A. Haldorai, S. Selvakanmani, and M. Thangamani, “Analysis of taskable mobile IoT sensing systems
for coverage and throughput,” International Journal of System Assurance Engineering and Management, Feb. 2023, doi: 10.1007/s13198-023-
01872-w.
[30]. D. K. Mishra, A. Thomas, J. Kuruvilla, P. Kalyanasundaram, K. R. Prasad, and A. Haldorai, “Design of mobile robot navigation controller using
neuro-fuzzy logic system,” Computers and Electrical Engineering, vol. 101, p. 108044, Jul. 2022, doi: 10.1016/j.compeleceng.2022.108044.
[31]. R. Sankaranarayanan, K. S. Umadevi, N. Bhavani, B. M. Jos, A. Haldorai, and D. V. Babu, “Cluster-based attacks prevention algorithm for
autonomous vehicles using machine learning algorithms,” Computers and Electrical Engineering, vol. 101, p. 108088, Jul. 2022, doi:
10.1016/j.compeleceng.2022.108088.
[32]. G. S, D. T, and A. Haldorai, “A Supervised Machine Learning Model for Tool Condition Monitoring in Smart Manufacturing,” Defence Science
Journal, vol. 72, no. 5, pp. 712–720, Nov. 2022, doi: 10.14429/dsj.72.17533.

99