Driver Attention in Automatic Transmission Cars

ORIGINAL ARTICLE

Data-driven system health monitoring technique using autoencoder for the safety management of commercial aircraft

Hyunseong Lee1 • Hyung Jin Lim1 • Aditi Chattopadhyay2

Received: 15 April 2020 / Accepted: 11 July 2020 / Published online: 23 July 2020 � Springer-Verlag London Ltd., part of Springer Nature 2020

Abstract This paper presents the development of a real-time and data-driven aircraft health monitoring technique to monitor in-flight

operations by detecting off-nominal flight operation such as aircraft upset, which is strongly related to aviation safety and is

a primary contributor to fatal accidents worldwide. A deep autoencoder is adopted to effectively estimate the complex

flight responses from historical flight datasets and capture potential upset precursors. Key sensing variables highly relevant

to upset primary factors are selected based on aircraft accident reports. The variables are then preprocessed using deci-

mation and Savitzky–Golay filter for sampling frequency synchronization and denoising. Statistical detection baselines are

employed to extract statistically unusual operation patterns that highly correlate with upset precursors. The performance of

the developed methodology is evaluated by demonstrating real-time detection of upset precursors in flight datasets. An

actual accident scenario is introduced to validate detection robustness. The results show that the proposed aircraft upset

detector technique can enhance pilots’ situational awareness by providing early safety alert for more effective flight safety

management.

Keywords Aircraft health monitoring � Aircraft upset detection � Autoencoder � Aviation safety

1 Introduction

Despite the significant advances in commercial aircraft

guidance and control technologies in recent years, the risk

of encountering aircraft upset conditions remains a critical

safety issue [1]. Aircraft upset conditions can be defined as

‘‘any uncommanded or inadvertent event with an anoma-

lous aircraft attitude, attitude rate, acceleration, airspeed, or

trajectory,’’ all of which can lead to aircraft accidents [2].

One of the most important causes of aircraft accidents is

loss of control in-flight (LOC-I), a main contributor to fatal

aviation accidents worldwide, with the highest number of

fatalities [3]. An LOC-I event is highly complex event and

is often caused by aircraft upset conditions with an

individual or multiple causal factors. A statistical analysis

of commercial aircrafts conducted by the Boeing Company

showed a total of 51 fatal accidents occurring between

2008 and 2019, out of which a total of 1183 fatalities

(46.7% of total fatalities) resulted from 13 LOC-I accidents

(25.5% of fatal accidents) [4]. A comprehensive study by

the National Aeronautics and Space Administration

(NASA) Langley Research Center of 126 LOC-I accidents

(Part 121) occurring between 1979 and 2009 showed sev-

eral contributing factors and the worst scenarios created by

the combination of these factors [3, 5]. Three primary

factors were identified: (1) system faults and errors (69

frequencies (33.3% of total factors)) with 1544 fatalities

(29% of total fatalities)), (2) icing (27 frequencies (21.4%

of total factors)) with 590 fatalities (9.7% of total fatali-

ties)), and (3) inappropriate crew response (18 frequencies

(14.3% of total factors)) with 1095 fatalities (14.3% of total

fatalities)). Specifically, system faults and errors events

were associated with autopilot, engine, and control surface

malfunctions or faults; icing events were related to the

airfoil contamination that led to severe stall condition; and

inappropriate crew response events were associated with

& Hyung Jin Lim [email protected]

1 School for Engineering of Matter, Transport, and Energy,

Arizona State University, Tempe, AZ 85287, USA

2 Mechanical and Aerospace Engineering, School for

Engineering of Matter, Transport, and Energy, Arizona State

University, Tempe, AZ 85287, USA

123

Neural Computing and Applications (2021) 33:3235–3250 https://doi.org/10.1007/s00521-020-05186-x(0123456789().,-volV)(0123456789().,-volV)

the lack of pilots’ situational awareness, which caused

failure to recover upset conditions.

In recent studies, a report by the International Air

Transport Association (IATA) has shown that the number

of worldwide flight passengers and air traffic is expected to

double by 2035 [6]. This study shows that an increase in

safety-related issues may not be able to be completely

managed by current aviation safety infrastructures, partic-

ularly since the current aviation system relies heavily on

manual operations, inspections, and communications

between pilots and air traffic controllers (ATCs). Thus,

there is an increased need for automated and proactive

safety management techniques for aviation safety. Tech-

niques that monitor aircraft system to accurately diagnose

upset conditions at initial stage and to prevent further

development of catastrophic accidents with potential

domino effects on the air traffic system are an emerging

need. As an example, the Federal Aviation Administration

(FAA) is currently developing the Next Generation Air

Transportation System (NextGen) to efficiently manage

aircraft and air traffic safety through an interconnected

information process between aircraft and ATCs [7]. As

such, there is a need for a reliable and automated in-flight

aircraft system health monitoring (SHM) technique to

prevent aircraft accidents and enhance overall aviation

safety. Furthermore, such monitoring techniques are also

expected to enhance the pilots’ situational awareness to

mitigate aircraft upsets at early stage.

Many techniques have been proposed to monitor aircraft

system and detect upset conditions, and generally they fall

into two broad categories: (1) model-based and (2) data-

driven. Most model-based techniques adopt a physics-

based numerical model to identify upset events [8–15].

These techniques use aircraft flight dynamics models to

estimate safe operating envelopes and to capture upset

events when the aircraft operates near the edge or out of the

envelopes. In addition, such techniques adaptively estimate

the envelopes to provide safety margins under varying

flight conditions. For instance, an adaptive and predictive

control algorithm was developed to predict the safety

margins in real time as the aircraft operation approaches

the edge of the safe operating envelopes [9]. An adaptive

estimation technique using differential vortex lattice algo-

rithm in conjunction with extended Kalman filter was

proposed to estimate impaired aircraft performance

envelopes [13]. However, the main limitations of these

model-based approaches are (1) the high cost of modeling a

complete aircraft, (2) a lack of generalization (strictly

system dependent), and (3) limited ability to quantify

uncertainties in realistic flight operations [16].

Recently, data-driven approaches have been extensively

used to perform aircraft SHM because of their efficient

information fusion of multivariate time series data and

robust diagnostic capabilities [17–31]. In general, flight

data recorder (FDR) datasets obtained from aircraft

onboard sensors or simulated flight data can be used to

train machine learning (ML) algorithms to model aircraft

system [21]. Based on these trained ML models, the current

status of the aircraft system is assessed, and future flight

performances are predicted. For instance, a predictive

monitoring technique that used a neural network (NN) and

Markov chain (MC) to predict landing state such as vertical

speed, acceleration, and distance was proposed to mitigate

landing safety issue, which can efficiently prevent further

potential domino effects near airport [22]. Similarly, an

aviation safety monitoring technique using a Gaussian

mixture model was proposed to detect abnormal flight

operations during takeoff and approach flight phases in the

FDR datasets. The detected anomalies were validated by

aviation experts, and the detection capability was examined

by comparing the detection results with conventional

exceedance detection (industry standard), multiple kernel

anomaly detection, and cluster-based anomaly detection

methods [23]. Additionally, a detection technique using

logarithmic multivariate Gaussian (LMG) model was

developed to capture the operational anomalies in the FDR

datasets. Specifically, multiple LMG models were sepa-

rately trained for each flight phase to eliminate the effect of

correlation changes in the sensing variables at different

flight phases, and the anomaly detection was performed for

each flight phase [24]. While many data-driven approaches

to perform aircraft SHM are available, the diagnostic

capabilities of current approaches are limited since they do

not address the following issues: (1) real-time monitoring

of entire in-flight operation to detect abnormal flight con-

ditions prior to actual upsets; (2) system health assessment

by monitoring key sensing variables that are directly rela-

ted to upset events; and (3) performance evaluation under

actual upset scenario.

Motivated by these challenges, a real-time and data-

driven SHM technique that uses a deep autoencoder (DAE)

to detect potential upset precursors of commercial aircraft

is presented in this paper. This SHM technique tackles the

limitation of previous data-driven approaches, as men-

tioned earlier, and is expected to improve the situational

awareness of pilots while providing a sufficient margin of

time for pilots to prevent hazardous situations. The FDR

datasets recorded from commercial aircraft are utilized to

develop the proposed technique and to capture possible

upset precursors in the datasets. Based on the investigation

of aircraft accident reports, key sensing variables that

highly correlate with upset contributing factors are selected

from the FDR datasets. Decimation and Savitzky–Golay

(SG) filtering are utilized to preprocess the sensing vari-

ables to implement sampling rate synchronization and

denoising. Subsequently, a DAE model is constructed and

3236 Neural Computing and Applications (2021) 33:3235–3250

123

trained using the preprocessed variables to model the

complex relationships and statistical patterns among the

key sensing variables. The architecture and hyperparame-

ters of the DAE are optimized based on the trained model

performances. The potential upset precursors are detected

and localized using the detection baselines derived from

the reconstruction error distributions. The proposed tech-

nique is then validated through a demonstrative real-time

precursor detection in the FDR datasets. The associated

possible upset initiation and coupled response are analyzed.

Finally, a FDR dataset obtained from an actual accident

report is used in order to further demonstrate the diagnostic

robustness of the proposed approach. The developed air-

craft SHM technique can serve as a part of an integrated

aviation health management system (e.g., the NextGen

system) that automatically interconnects information

between aircraft and air traffic controllers (ATCs) for

efficient and proactive health assessment.

The remaining part of the paper is organized as follows.

The research background for the DAE model and the FDR

datasets used in this study are described in Sect. 2. In

Sect. 3, the aircraft SHM technique using the DAE model

and detection baseline is detailed. The performance of the

developed technique is evaluated in Sect. 4. Finally, in

Sect. 5, the concluding remarks are presented.

2 Research background

2.1 Deep autoencoder (DAE)

Autoencoder (AE) is an unsupervised artificial neural net-

work model and is used to estimate reconstructed input

variables (outputs) that are close to their original inputs by

accounting for the complex relationships and statistical

patterns in multivariate variables [32, 33]. As shown in

Fig. 1, the AE model mainly consists of three types of

layers: (1) input (Layer 1), (2) bottleneck hidden (Layer 2),

and (3) output (Layer 3). In particular, the AE model with a

single hidden layer includes a process of encoding (Layer 1

to 2) and decoding (Layer 2 to 3); the equation of the

encoding and decoding can be formulated as Eq. 1 and

Eq. 2, respectively [34]:

z ¼ a Wxzx þ bxzð Þ ð1Þ x̂ ¼ a Wzxz þ bzxð Þ: ð2Þ

The encoding process in Eq. 1 maps the input variables

x in Layer 1 to a hidden representation z in Layer 2 through

an affine transformation (weight W and bias b) with a

nonlinear transformation (activation) function a. The original inputs are reconstructed (x̂) by mapping the hidden

representation in Layer 2 back to its original input space

(Layer 3) in the decoding process (Eq. 2). Subsequently,

the reconstruction errors � between the original and

reconstructed inputs are estimated as

� ¼ x � x̂k k; ð3Þ

where �k k is often set to be ‘2-norm. In the training process, the AE model is optimized by minimizing the reconstruc-

tion error. In general, the AE model with more than one

hidden layer is called DAE, and the additional encoding

and decoding processes are implemented in each added

hidden layer. By adding multiple hidden layers, DAE can

effectively represent the complex patterns and nonlinear

relations over input variables [35].

In addition, DAE model can be utilized for deviation-

based anomaly detection where anomalous instances are

scored by reconstruction errors. For instance, in a semi-

supervised manner, normal instances (assumed majority of

the datasets) can be used to train the DAE model so that the

trained model provides an accurate estimation with small

reconstruction errors when the inputs are normal events. In

contrast, large reconstruction errors are observed when

inputs are anomalous events (minority of the datasets)

[36, 37]. This concept can be extended to aircraft upset

detection scenarios where the instances of normal events

are abundant, while upset events are rare. Hence, in this

study, the DAE model is adopted to model standard flight

operations; abnormal operations or potential upsets are

detected using the reconstruction error and the statistical

detection baselines. The detailed description of the DAE

model construction and detection baselines for the upset

precursor detection is addressed in Sect. 3.4.

Fig. 1 Schematic of general autoencoder (AE) model

Neural Computing and Applications (2021) 33:3235–3250 3237

123

2.2 Dataset construction and assumptions

In general, FDR datasets (also called the ‘‘black-box’’

datasets) are recorded from onboard sensors in commercial

aircraft during flight operations and used for flight analytics

and forensics after aircraft incidents or accidents. In this

study, large FDR datasets of commercial aircraft, which

have been sanitized and obtained from NASA’s Discovery

in Aeronautics Systems Health (DASH) link network [38],

are utilized (referred to as DASHlink datasets in this work).

The DASHlink datasets involve the recorded sensing

variables in time series from 186 onboard sensors as dis-

crete or continuous forms. For instance, the discrete vari-

ables are generally integers that indicate a particular flight

state or mode, such as flight phase and landing gear posi-

tion. The continuous variables display the current aircraft

performances and operations, such as aircraft position,

velocity, and engine operation-related variables. The flight

phases in the datasets are categorized into seven groups: (1)

parking, (2) taxi, (3) takeoff, (4) ascent, (5) cruise, (6)

descent, and (7) landing. Each digit represents the current

flight phase of the aircraft. The sampling rate of the

recorded variables varies between 0.25 and 16 Hz, and the

flight duration range is from 56 min to 3 h and 22 min,

which hinders the direct usage of the datasets for aircraft

monitoring [21]. Furthermore, the DASHlink datasets are

unlabeled, so any explicit information about aircraft type,

system health metrics, and flight plan (departure, way-

points, and arrival) is not provided.

The aforementioned problems prevent direct compar-

ison with reported flight incidents or accidents in public

records and limit the performance of data-driven, in-flight

monitoring techniques since flight performances are highly

dependent on flight plans. Therefore, a decision boundary

was assumed that flight operations under potential upset

events will exhibit statistically significant off-nominal

tendencies with the abnormal flight dynamics. In addition,

the DASHlink datasets were categorized by the flight plans,

and the potential aircraft type was identified to mitigate

these issues. A total of 92,329 flights were investigated; the

recorded latitude (LATP) and longitude (LONP) at the

beginning and end of in-flight (weight on wheels switch

off) were used to identify the departure and arrival airports,

respectively. In particular, the locations of the airports were

pinpointed by searching the closest airports from the

recorded LATP and LONP within 0.005 decimal degrees.

With this process, the DASHlink datasets were categorized

by the flight plans. Among them, 1179 flights categorized

by three flight plans, where the largest number of flight

events is available among the datasets, were investigated.

Table 1 details the investigated DASHlink datasets

including the airports and number of flight events.

Subsequently, the identification of the potential aircraft

type was conducted by matching flight history record with

two additional databases: (1) Bureau of Transportation

Statistics [39] and (2) NASA Sherlock [40]; these data-

bases include the past flight schedules with the aircraft

types and airport information. For instance, the departure

and arrival time at a certain airport in the DASHlink

datasets were matched with the recorded flight schedules in

these databases. The potential aircraft type was presumed

as a small-sized and single-aisle aircraft such as Bom-

bardier CRJ700, De Havilland Canada Dash 8 (DHC-8)

Q400, and Embraer E170 series aircrafts. Based on the

presumed aircraft type, accidents from Part 121 occurred

between 2000 and 2019 (20 years) were investigated from

the National Transportation Safety Board (NTSB) database

[41]. Among the reported accidents, a scenario of the

severe aerodynamic stall was found and investigated [42].

Hence, two FDR datasets were used in this study: The

DASHlink datasets were utilized to develop the proposed

aircraft SHM technique and demonstrate its detection

capabilities, and the accident dataset obtained from the

National Transportation Safety Board (NTSB) database

(referred to as NTSB dataset in this work) was used to

further validate the diagnostic capabilities of the proposed

SHM method under actual upset scenario. In-flight opera-

tions, including takeoff, ascent, cruise, and descent flight

phases, were considered in this work to examine the

robustness of the developed technique under flight phase

variations.

3 Aircraft system health monitoring using deep autoencoder

3.1 Overview

The proposed detection approach is mainly divided into a

training (off-line) phase and a monitoring (online) phase, as

shown in Fig. 2. In the training phase, the data prepro-

cessing and DAE model training are implemented. Key

sensing variables that are directly related to primary air-

craft upset contributing factors are selected from the orig-

inal DASHlink datasets for accurate and efficient aircraft

SHM, as detailed in Sect. 3.2. The selected key variables

are preprocessed using the decimation in conjunction with

the SG filter to synchronize sampling rates and reduce

inherent noise in the sensing variables, as described in

Sect. 3.3. Subsequently, the preprocessed variables are

utilized to train the DAE model to estimate the standard

flight performances during in-flight operations. The statis-

tical detection baselines derived from the reconstruction

errors are modeled to detect abnormal flight operations that

are related to upset precursors; this is further expanded

3238 Neural Computing and Applications (2021) 33:3235–3250

123

upon in Sect. 3.4. In the monitoring phase, the FDR data-

sets are introduced into the trained DAE model with the

detection baselines to capture upset precursors. In partic-

ular, the precursors are initially detected by the global

baselines and then localized by the local baselines,

respectively. The training process of the DAE model and

the detection results are presented in Sect. 4.

3.2 Key sensing variable selection

In general, data-driven SHM techniques, especially per-

formed on large datasets associated with irrelevant vari-

ables, suffer from expensive computational cost and

overfitting problems. These issues can adversely affect the

real-time diagnostic abilities of the monitoring technique.

Therefore, a relevant variable selection is vital to improve

the detection accuracy with computational efficiency for

real-time monitoring. Furthermore, the relevant sensing

variable selection based on an understanding of the actual

upset scenarios and its main contributing factors (primary

causes of aircraft upsets) is necessary to enhance the

detection accuracy while minimizing false alarm rates by

incorporating a priori physics-based knowledge into the

data-driven SHM technique. For instance, Turkish Airlines

Flight 1951 was stalled during the approach phase and

crashed at a runway in Amsterdam Schiphol Airport,

Table 1 Investigated DASHlink datasets categorized by flight plans

Flight

plan

Departure airport Arrival airport Number of

flights

1 Minneapolis–Saint Paul International Airport

(KMSP)

Cleveland Hopkins International Airport (KCLE) 424

2 KMSP Cincinnati/Northern Kentucky International Airport

(KCVG)

387

3 Hartsfield–Jackson Atlanta International Airport

(KATL)

Memphis International Airport (KMEM) 368

Fig. 2 Schematic of the proposed real-time aircraft health monitoring method

Neural Computing and Applications (2021) 33:3235–3250 3239

123

Netherlands, led to 9 fatalities; this event was caused by a

radio altimeter fault, which changed the auto-throttle con-

figuration to decrease engine thrust level to idle [43]. In

this case, monitoring changes in the altitude and engine

operations is key to capturing the upset initiations, as a

means to prevent the severe upset event. Based on the

investigations of upset accidents with corresponding con-

tributing factors [3, 5], 41 key sensing variables that are

directly related to the primary factors were selected from

the DASHlink datasets (the number of original sensing

variables n = 186) to train the DAE model and to capture

potential upset precursors in the datasets. The selected key

sensing variables include flight dynamics and engine

operation-related variables, as summarized in Table 2.

3.3 Data preprocessing

The main issues in the DASHlink datasets are the different

sampling rates of each sensing variable and random noises,

which cannot be directly employed for the proposed tech-

nique. Therefore, the decimation and SG filter were

adopted to conduct the sampling rate synchronization and

for denoising the selected key sensing variables. The dec-

imation technique is a down-sampling technique with two

steps: (1) eliminating high-frequency components with a

digital low-pass filter for anti-aliasing and (2) down-sam-

pling with an integer factor [44]. In each flight event p

(p ¼ 1; 2; . . .; q; q is the number of flights), the selected

key sensing variable can be described as x j i , where i is a

sensing variable index (i ¼ 1; 2; . . .; m; m is the number of selected variables) and j is a time step ðj ¼ 1; 2; . . .; liÞ. li is the number of the time steps for the ith sensing variable and

varies due to the different sampling rates in each variable.

The resampled ith sensing variable with the synchronized

time step k (k ¼ 1; 2; . . .; L; L is the number of synchro- nized time steps) using the decimation method can be

described as

~xki ¼ XW�1

w¼0 x jD�v½ � i � h w½ �; ð4Þ

where ~xki , D, h, and W denote the resampled variable,

integer factor, impulse response, and its length, respec-

tively. In this study, a finite impulse response (FIR) filter

was used to prevent aliasing with computational efficiency

because of no iterations in the filtering process. The order

of the FIR filter was set as 30 to perform the filtering

process on various sampling rates of the sensing variables.

With this approach, the sampling rates of the selected

variables were synchronized to 1 Hz; the different data

lengths of the variables li were synchronized to the length

L. For instance, the sampling rate of the engine fan speed 1

(FS 1) is synchronized from 4 to 1 Hz, as shown in Fig. 3.

Furthermore, the SG filtering was performed to denoise the

sensing variables originally recorded under 1 Hz. The

performance of this filter has shown promising denoising

capabilities with a decreased signal distortion compared to

Table 2 List of selected key sensing variables from

DASHlink datasets

Variable name Description Unit Category

ALT Altitude Ft Flight dynamics

LATP Latitude Deg Flight dynamics

LONP Longitude Deg Flight dynamics

PTCH Pitch angle Deg Flight dynamics

ROLL Roll angle Deg Flight dynamics

MH Magnetic heading Deg Flight dynamics

DA Drift angle Deg Flight dynamics

VS Vertical speed ft/s Flight dynamics

GS Ground speed Knots Flight dynamics

IAS Indicated airspeed Knots Flight dynamics

LONG Longitudinal acceleration G Flight dynamics

VRTG Vertical acceleration G Flight dynamics

LATG Lateral acceleration G Flight dynamics

PLA 1-4 Engine power lever angle input 1-4 Deg Engine operation

FS 1-4 Engine fan speed 1-4 %rpm Engine operation

CS 1-4 Engine core speed 1-4 %rpm Engine operation

FF 1-4 Fuel flow rate 1-4 lbs/h Engine operation

EGT 1-4 Engine exhaust gas temperature 1-4 Deg Engine operation

OIT 1-4 Oil temperature 1-4 Deg Engine operation

OIP 1-4 Oil pressure 1-4 Psi Engine operation

3240 Neural Computing and Applications (2021) 33:3235–3250

123

the moving average filters in the literature [45]. Therefore,

the filtered ith sensing variable for the time step k using the

SG filter can be represented as follows:

~xk�i ¼ PM

j¼�M cjx jþk½ � i

N ; ð5Þ

where ~xk�i , cj, N, and M are the filtered sensing variable, coefficient for jth filtering, number of datapoints, and half-

width of the smoothing window, respectively. The window

size is defined as 2M þ 1.

3.4 Aircraft upset precursor detection using deep autoencoder

This section discusses the proposed detection technique

using the DAE model with the detection baselines for the

upset precursor detection and localization. First, the pre-

processed 41 key sensing variables were scaled between 0

and 1 to avoid the optimization issue during the training

process. Subsequently, the scaled variables were intro-

duced into the input layer to train the DAE model. The

number of 41 neurons was defined in the input and output

layers. The reconstructed variables (output) were then

scaled back to their original scales in the monitoring phase.

Therefore, each reconstructed variable is estimated as a

function of all input variables as follows:

x̂i ¼ f x1; x2; . . .xnð Þ: ð6Þ

This process allows the estimation of the flight perfor-

mance by modeling the complex relationships and inter-

actions between the variables. For the activation function, a

rectified linear unit (ReLU) was used to avoid gradient

vanishing problem and fulfill computational efficiency

[46]. Mean square error (MSE) was selected as a loss

function to train the DAE model. In addition, batch nor-

malization was implemented for the efficient training

process and enhanced model generalization by reducing the

internal covariate shift phenomenon [47]. In this case, the

off-nominal flight operations in the DASHlink datasets are

assumed to be an overfitting contributor since these

behaviors are statistically a minor portion of the datasets

and may significantly deviate from a major portion (nom-

inal flights). Therefore, the off-nominal trends are most

likely ignored during the training process with the batch

normalization, and the DAE model parameters are mainly

trained by the majority of the datasets (nominal flights).

With this assumption, the abnormal operations potentially

related to upset conditions are expected to be detected by

large reconstruction errors. The reconstruction error from

the trained DAE model was used to define the statistical

detection baselines. The reconstruction error of the ith key

sensing variable is standardized as

��i ¼ ��i � li ri

; ð7Þ

where � and �� indicate the ‘2-norm reconstruction error and

standardized reconstruction error, and l and r are the mean and standard deviation of the reconstruction error, respec-

tively. The standardized errors were used to define the

detection baselines based on assumption that each error is

normally distributed. In particular, the global detection

baseline estimated by the multivariate Gaussian (MG)

distribution was used to detect the upset precursors based

on global estimation. After the detection, the sources of the

upset precursors were localized by testing each variable

using the local detection baselines estimated by the uni-

variate Gaussian (UG) distributions. With the number of 41

reconstruction errors, the global probability density func-

tion p ��; l; Rð Þ of the MG distribution [48] can be defined as follows:

p ��; l; Rð Þ ¼ 1 ffiffiffiffiffiffiffiffiffiffiffi 2p Rj j

p exp � 1

2 �� � lð ÞTRT �� � lð Þ

� � ; ð8Þ

where R is the covariance matrix. In this study, the detection baselines were defined with 0.1% of off-nominal

ratio. Therefore, the statistically off-nominal responses are

captured when the standardized reconstruction errors vio-

late the defined global baseline. The global off-nominal

Fig. 3 Comparison between the original and preprocessed FS 1

during a entire flight phases and b descent and approaching flight phases (close look of a)

Neural Computing and Applications (2021) 33:3235–3250 3241

123

metric H at jth time step can be defined by binary variables as

H j ¼ 1; if p �� j; l; Rð Þ\w

0; otherwise

� ; ð9Þ

where 1 and 0 represent the global off-nominal and nom-

inal indicators, respectively, and w is a threshold value defined by the off-nominal ratio. The local off-nominal

metrics h for the ith sensing variable at jth time step is defined by

h ji ¼ 1; if ��

j i

�� �� [ / 0; otherwise

� ð10Þ

where 1 and 0 denote the local off-nominal and nominal

indicators, respectively, and / is a threshold value. In this study, / ¼ 6r was used to only identify significant off- nominal trends induced by the dramatic changes in flight

dynamics, while minimizing the unnecessary alarms that

can cause the pilot fatigue [49, 50]. The upset precursors

are detected when the standardized reconstruction errors

violate at both global and local baselines for at least 5 s to

ignore sensor and data acquisition errors that are not

indicative of any upset behaviors. The optimization process

of the DAE model architecture (i.e., the number of hidden

layers and neurons), the training process, and the demon-

stration of the assumption for the detection baselines will

be discussed in Sect. 4.1.

4 Results and discussion

4.1 Model training process and evaluation

This section presents the training process and the evalua-

tion of the trained DAE model. A total of 1179 flights

categorized by flight plans in Table 1 were divided into

two segments to train and validate the model: (1) training

set (70% of total flights) and (2) validation set (30% of total

flights). For instance, the datasets of the flight plan 1 (424

flights) were split into 297 flights and 127 flights for

training and validating the DAE model, respectively. The

validation process was mainly performed to prevent over-

fitting issues that decrease the estimation accuracy [51].

Mini-batch stochastic optimization using an adaptive

moment estimation (Adam) optimizer [52] was performed

with the 2048 batch size and learning rate of 1e-8 to

minimize the MSE between the inputs (ground truth) and

reconstructed inputs. In addition, parametric studies were

performed to obtain the optimized DAE model with dif-

fering three parameters: the numbers of (1) neurons, (2)

hidden layers, and (3) epochs (iterations). The reconstruc-

tion errors for each case of the trained DAE model were

quantitatively measured by root-mean-square error

(RMSE) to evaluate their respective performances. The

DAE model training, validation, and testing were con-

ducted on a TensorFlow 2.0.0 with Python 3.7.4 environ-

ment and desktop with an AMD Ryzen 7 3700X (eight

cores with a 3.6 GHz core clock), NVIDIA GeForce GTX

1050 graphics processing unit (GPU) (640 CUDA cores

with a 1.3 GHz core clock), and 32 GB DDR4 memory

running on a Windows 10 (64-bit) operating system.

Initially, five cases of the DAE model training were

performed by varying the number of neurons in hidden

layers, as described in Table 3. For a consistent evaluation,

the number of neurons was decreased or increased linearly

over hidden layers for all cases. Case 2 showed the best

performance (7.22e-5 RMSE in training and 2.83e-4

RMSE in validation) compared to the other cases, as shown

in Fig. 4a. With the optimized number of neurons, the DAE

models were trained by differing the number of hidden

layers. Figure 4b compares the RMSE values for each case;

the DAE with five hidden layers presented the best per-

formance (6.05e-5 RMSE in training and 2.04e-4 RMSE

in validation). Subsequently, the number of epochs was

optimized to obtain reconstruction accuracy, while mini-

mizing the overfitting issue. In particular, the DAE models

were separately trained using three different flight plans

with the optimized model architecture (i.e., optimized

number of neurons and hidden layers). The optimized

epochs were determined by capturing the point of inflection

in the validation RMSE. Figure 4c illustrates a comparison

of the RMSE between training and validation in the case of

flight plan 1. The training RMSE showed a monotonic

decrease over epochs, while the validation RMSE slightly

increased after 2602 epochs (1.67e-4 RMSE) due to the

possible initiation of the overfitting. In the same way, the

optimized epochs for flight plans 2 and 3 were determined

as 2432 and 2118 epochs, respectively. Therefore, the

optimized DAE architecture and number of epochs were

used to monitor aircraft system health. As an example,

Fig. 4d shows a comparison between the original and

reconstructed longitudinal acceleration (LONG) during the

entire in-flight operation, showing good correlation.

In addition, the reconstruction errors were analyzed to

demonstrate the assumption that the evaluated errors from

the optimized DAE model follow Gaussian distribution.

Therefore, the standardized reconstruction errors of 41 key

sensing variables were examined using the datasets of the

flight plan 1 (424 flights). Furthermore, chi-square tests

were implemented to quantitatively measure the degree of

normality [53]. For instance, the error distributions of the

pitch angle (PTCH) and indicated airspeed (IAS) show

close similarities with UG distribution, as presented in

Fig. 5a and b, respectively. The chi-square tests with a 5%

significance level were performed on the standardized

3242 Neural Computing and Applications (2021) 33:3235–3250

123

errors from the 41 key sensing variables and verified that

all error distributions follow UG distributions within the

significance level. These results demonstrated the

feasibility of using MG and UG distributions for the upset

precursor detection and localization, respectively. In

addition, the processing time of the proposed detection

Fig. 4 Optimization process and model evaluation of the

DAE model: a number of neurons, b number of hidden layers, c number of epochs, and d comparison between the original and reconstructed

LONG

Fig. 5 Standardized reconstruction error histograms:

a PTCH and b IAS

Table 3 Cases of number of neurons in hidden layers

Case 1st encoder 2nd encoder Bottleneck 1st decoder 2nd encoder

1 37 34 31 34 37

2 35 30 25 30 35

3 33 26 19 26 33

4 31 22 13 22 31

5 29 18 7 18 29

Neural Computing and Applications (2021) 33:3235–3250 3243

123

technique was measured to examine the computational

efficiency for real-time monitoring. The 41 key sensing

variables with a time step of 50 s were preprocessed and

introduced into the trained DAE model. Ten measurements

were performed on the same computational environment,

and the average processing time was measured as 0.766 s;

the processing time will be less for a single time step input

(1 s). Thus, the computational efficiency of the proposed

technique for real-time monitoring was demonstrated. It is

noted that the proposed technique is possibly applicable to

aviation health management systems, such as NextGen,

where the aircraft and air traffic monitoring are imple-

mented on a distributed, but interconnected computational

environment between the aircraft, the air traffic controller,

and dedicated processing nodes.

4.2 Real-time aircraft upset precursor detection

In this section, the proposed monitoring technique was

validated through the demonstrations of real-time upset

precursor detection. Furthermore, the detected precursors

associated with the coupled responses were analyzed. In

order to detect and analyze the upset precursors in the

DASHlink datasets, the preprocessed 41 input key sensing

variables from 1179 flight cases (Table 1) were separately

introduced into each case of the trained DAE model in real

time. The precursors were captured when the standardized

reconstruction errors violate the global and local detection

baselines for at least 5 s. In addition, a case of aircraft

accident dataset (NTSB dataset) was utilized to further

demonstrate the detection capability of the proposed

method under the actual upset scenario. The cases of

analyzed flights are detailed in Table 4.

4.2.1 Spiral descent

In case 1, the spiral descent trajectories with the associated

responses were detected during the descent flight phase.

Spiral mode is an unstable flight condition and often ini-

tiated by a disturbance in the sideslip angle, which typi-

cally follows a disturbance in the roll angle and leads to a

wing to drop. Once the spiral mode is excited, the aircraft

maneuvers in a slow diverging path in both the roll and

yaw angles. The aircraft loses its altitude since the vertical

forces are no longer in equilibrium. Hence, the associated

flight path is most likely a spiral descent. It is noted that

most spiral descents can be recognized and addressed by

pilots within a sufficient margin of time since the rate at

which the mode diverges is very slow [54]. However, the

pilots’ ability to recognize spiral descent can be signifi-

cantly reduced under poor weather conditions, which may

be further developed and lead to severe aircraft upsets. This

abnormal behavior was observed in the DASHlink datasets

as shown in Fig. 6. In particular, the initiation of the spiral

descent was detected at 4835 s when the reconstruction

errors violated the defined global detection baseline as

shown in Fig. 6a. Subsequently, the precursors in each

variable were captured and localized by the local detection

baselines. Figure 6b–e shows the comparisons between the

original and the reconstructed variables with the error plots

showing the detected precursors. In particular, the initiation

and continuation of disturbances (precursors) in roll angle

(ROLL) and magnetic heading (MH) were observed as

shown in Fig. 6b and c. The abrupt changes in ROLL were

initially detected at 4919 s, and the two additional abrupt

changes were captured at 5202 s and 5440 s, respectively,

as the spiral descent continued. Similarly, the instantaneous

changes in MH were observed at 5835 s and 6687 s,

respectively. These rapid changes were also observed in

PTCH and captured at 5607 s, 6253 s, and 6952 s,

respectively as shown in Fig. 6d. In addition, the distur-

bances in IAS were captured at 6886 s and 6975 s,

respectively, while there was no detection occurring in

vertical speed (VS), as shown in Fig. 6e and f. It is

hypothesized that the potential upset behavior was initiated

in ROLL, and the coupled effects were caused due to the

propagation of the potential upsets in ROLL; these pre-

cursors were accurately captured by the proposed tech-

nique. The obtained results demonstrated the ability of the

proposed method to detect and localize the upset

precursors.

4.2.2 Normal flight operation

In case 2, nominal flight operation was investigated to

demonstrate the reliability of the precursor detection. In

particular, the normal operation, where the reconstruction

errors did not violate any detection baselines, was selected

from the datasets in flight plan 2 (Table 1). Figure 7a

illustrates the flight trajectories during the entire in-flight

operation; any sudden changes or off-nominal responses

were not observed. In addition, the key sensing variables

Table 4 Cases of analyzed flights

Case Scenario Flight phase Data source

1 Spiral descent Descent DASHlink

2 Normal flight operation All DASHlink

3 Severe aerodynamic stall Approach NTSB accident report

3244 Neural Computing and Applications (2021) 33:3235–3250

123

between the original and the reconstruction during the in-

flight operation are compared in Fig. 7b–f. In general, the

variables showed stable responses during the cruise flight

phase, while dynamic and nonlinear behaviors were

observed in the other flight phases. The trained DAE model

accurately estimated the flight operation in both stable and

dynamic states, showing smaller reconstruction errors

compared to case 1. It can be seen that the patterns of

normal operation (case 2) were recognized as the common

patterns in the trained DAE model.

4.2.3 Severe aerodynamic stall due to icing

A case of a reported upset accident primarily caused by a

severe aerodynamic stall (NTSB accident ID:

DCA09MA027) [42] was selected from the NTSB database

and utilized to further validate the proposed technique. The

accident aircraft was a Bombardier DHC-8 Q400 operated

by Colgan Air, Inc., and it crashed into a residence near

Buffalo-Niagara International Airport, Buffalo, NY, during

the approach flight phase, which led to 50 fatalities. One of

the main contributing factors identified by the NTSB

investigation team was the flight crew’s inappropriate

response to the activation of the stick shaker that resulted in

an irrecoverable aerodynamic stall. Another main con-

tributing factor to the accident was the flight crew’s failure

to monitor airspeed under icing conditions, indicative of

the importance of situation awareness in such a hazardous

situation. In this study, 18 key sensing variables were

selected from the NTSB dataset, also available in

Fig. 6 Detected upset precursors in case 1 (spiral

descent in DASHlink datasets):

a flight trajectories, b ROLL, c MH, d PTCH, e IAS, and f VS

Neural Computing and Applications (2021) 33:3235–3250 3245

123

DASHlink datasets, including flight dynamics and engine

operation variables, as detailed in Table 5. As the same

flight plan of the accident data was not available in the

DASHlink datasets, the flight plan 3 dataset in Table 1 was

selected because it headed in a similar direction as the

accident data. With this approach, 18 variables within flight

plan 3 were extracted from the DASHlink datasets and

these variables were used to train the DAE model. It is

noted that the longitude (LONP) and latitude (LATP)

variables were excluded in the training datasets since the

same flight plan was not available, and the duration of the

scenario was too short to consider the flight plan-dependent

aircraft performances. Furthermore, the DAE model

architecture was modified as the number of input sensing

variables was changed. The optimization process of the

DAE model was implemented with varying the numbers of

(1) neurons, (2) hidden layers, and (3) epochs. Conse-

quently, five hidden layers were determined, and the

number of neurons was optimized as 15, 12, 9, 12, and 15

neurons for 1st, 2nd, 3rd, 4th, and 5th layer, respectively.

The optimal number of epochs was determined as 1448

epochs. Subsequently, the preprocessed 18 sensing vari-

ables from the NTSB dataset were injected into the trained

DAE model, and the upset precursors and actual upsets

were marked when the reconstruction error violated the

detection baselines for at least 5 s.

The upset precursor in the altitude (ALT) was detected

at 22:15:22 eastern standard time (EST), which was 68 s

Fig. 7 Comparisons between the original and the

reconstructed variables in case 2

(normal flight in DASHlink

datasets): a flight trajectories, b ROLL, c PTCH, d FS 2, e IAS, and f VS

3246 Neural Computing and Applications (2021) 33:3235–3250

123

earlier than the activation of the stick shaker (alarm of

aerodynamic stall), and the upset was detected at 22:16:41

EST after the stall had occurred, as illustrated in Fig. 8a.

Similarly, the upset precursor in IAS was detected at

22:14:41 EST (109 s earlier detection than the stick sha-

ker), and the upset was captured at 22:16:11 EST (19 s

early detection) and continued until the time when the stick

pusher was activated as shown in Fig. 8b. In addition, the

precursors were detected in PTCH and vertical acceleration

(VRTG) (key features of aerodynamic stall). The upset

precursors in PTCH were captured at 22:16:17 EST (13 s

early detection) and continued until the end of the flight, as

shown in Fig. 8c. The precursor in VRTG was detected at

22:16:12 EST and continued until 22:16:36 EST, as pre-

sented in Fig. 8d. In addition, the precursors were detected

in ROLL (Fig. 8e) and the engine fan speed 2 (FS 2)

(Fig. 8f) at 22:15:30 EST (60 s early detection) and

22:14:28 EST (122 s early detection), respectively, and the

upsets were detected at 22:16:21 EST (9 s early detection)

and 22:16:32 EST, respectively. The results demonstrated

the capability to accurately detect the aircraft upset pre-

cursors in real time prior to actual upset event, which can

improve the pilots’ situational awareness to prevent aircraft

upsets or LOC-I in early stage.

5 Conclusion

In this study, a real-time and data-driven aircraft system

health monitoring (SHM) technique was developed to

capture potential aircraft upset precursors during in-flight

operations to improve aviation safety. Flight data recorder

(FDR) datasets recorded from onboard sensors in com-

mercial aircraft were used to develop the proposed

technique and demonstrate its diagnostic capabilities.

Reported aircraft accident scenarios with the main con-

tributing factors were investigated, and relevant key

sensing variables that were physically related to the upset

contributing factors were employed to enhance detection

accuracy with computational efficiency for real-time

detection. Data preprocessing using the decimation and

Savitzky–Golay (SG) filter was performed for the sam-

pling rate synchronization and denoising the selected

variables. The deep autoencoder (DAE) model was con-

structed and trained using the preprocessed variables to

estimate the standard flight performances and operations.

The architecture and hyperparameter of the DAE model

were optimized through parametric studies, and their

respective performances for each case were evaluated.

The processing time of the proposed technique with a

time step of 50 s was measured as 0.766 s, which

demonstrated the computational efficiency necessary for

real-time health monitoring in the integrated aviation

safety management system. With the trained DAE model

and the statistical detection baselines, a demonstrative

real-time upset precursor detection was performed in the

DASHlink FDR datasets. Initiation of a spiral descent was

detected by the disturbance in the roll angle; the associ-

ated coupled responses in other sensing variables were

also analyzed. In addition, the FDR dataset from the

actual upset accident was used for further demonstration.

The precursors indicating the initiation of upsets were

successfully captured by the proposed technique prior to

the severe upset situation. In particular, the precursors

were detected prior to the activation of the stick shaker

(maximum 122 s early). The results demonstrate the early

detection capability and detection accuracy of the devel-

oped method. By providing the precursor alerts to pilots

Table 5 List of selected sensing variables from NTSB accident

dataset used for validation of the

proposed detection technique

Variable name Description Unit Category

ALT Altitude Ft Flight dynamics

PTCH Pitch angle Deg Flight dynamics

ROLL Roll angle Deg Flight dynamics

MH Magnetic heading Deg Flight dynamics

DA Drift angle Deg Flight dynamics

GS Ground speed Knots Flight dynamics

IAS Indicated airspeed Knots Flight dynamics

LONG Longitudinal acceleration G Flight dynamics

VRTG Vertical acceleration G Flight dynamics

LATG Lateral acceleration G Flight dynamics

PLA 1-2 Engine power lever angle input 1-2 Deg Engine operation

FS 1-2 Engine fan speed 1-2 %rpm Engine operation

CS 1-2 Engine core speed 1-2 %rpm Engine operation

FF 1-2 Fuel flow rate 1-2 lbs/h Engine operation

Neural Computing and Applications (2021) 33:3235–3250 3247

123

prior to upset situations, the developed framework can

improve the pilots’ situational awareness and provide a

sufficient margin of time to recover from these situations.

Future research will include extension to other flight plans

and the integration of the developed technique with flight

simulations to investigate generalized diagnostic capabil-

ities under various aircraft upset scenarios.

Acknowledgements The research reported in this paper was sup- ported by funds from NASA University Leadership Initiative program

(Contract No. NNX17AJ86A, Project Officer: Dr. Anupa Bajwa,

Project Manager: Dr. Koushik Datta). The authors would also like to

thank Dr. P. K. Menon for his domain expertise and helpful discus-

sions. All supports are gratefully acknowledged.

Compliance with ethical standards

Conflict of interest The authors declared that they have no conflicts of interest to this work.

References

1. Gill SJ, Lowenberg MH, Neild SA, Krauskopf B, Puyou G,

Coetzee E (2013) Upset dynamics of an airliner model: a non-

linear bifurcation analysis. J Aircr 50:1832–1842. https://doi.org/

10.2514/1.C032221

2. Lambregts A, Nesemeier G, Newman R, Wilborn J (2008) Air-

plane upsets: old problem, new issues. In: AIAA modeling and

Fig. 8 Detected upset precursors in case 3 (severe

aerodynamic stall in NTSB

accident report): a ALT, b IAS, c PTCH, d VRTG, e ROLL, and f FS 2

3248 Neural Computing and Applications (2021) 33:3235–3250

123

simulation technologies conference and exhibit. American Insti-

tute of Aeronautics and Astronautics, Reston, Virigina

3. Belcastro C, Foster J (2010) Aircraft loss-of-control accident

analysis. In: AIAA guidance, navigation, and control conference.

American Institute of Aeronautics and Astronautics, Reston,

Virigina, pp 1–41

4. Boeing Commercial Airplanes (2019) Statistical summary of

commercial jet airplane accidents: worldwide operations

1959–2018

5. Belcastro CM, Foster JV, Shah GH, Gregory IM, Cox DE, Crider

DA, Groff L, Newman RL, Klyde DH (2017) Aircraft loss of

control problem analysis and research toward a holistic solution.

J Guid Control Dyn 40:733–775. https://doi.org/10.2514/1.

G002815

6. International Air Transport Airport (2018) Future of airline

industry 2035

7. Federal Aviation Administration (2016) NextGen implementation

plan

8. Pfifer H, Venkataraman R, Seiler P (2017) Quantifying loss-of-

control envelopes via robust tracking analysis. J Guid Control

Dyn 40:1042–1050. https://doi.org/10.2514/1.G001748

9. Stepanyan V, Krishnakumar K, Dorais G, Reardon S, Barlow J,

Lampton A, Hardy G (2017) Loss-of-control mitigation via pre-

dictive cuing. J Guid Control Dyn 40:831–846. https://doi.org/10.

2514/1.G001731

10. Schuet S, Lombaerts T, Acosta D, Kaneshige J, Wheeler K, Shish

K (2017) Autonomous flight envelope estimation for loss-of-

control prevention. J Guid Control Dyn 40:847–862. https://doi.

org/10.2514/1.G001729

11. Dlamini Z, Jones T (2016) Fly-by-wire robustness to flight

dynamics change under horizontal stabiliser damage. Aeronaut J

120:1005–1023. https://doi.org/10.1017/aer.2016.42

12. Lombaerts T, Schuet S, Acosta D, Kaneshige J (2015) On-line

safe flight envelope determination for impaired aircraft. Adv

Aerosp Guid Navig Control. https://doi.org/10.1007/978-3-319-

17518-8_16

13. Menon PK, Sengupta P, Vaddi S, Yang BJ, Kwan J (2013)

Impaired aircraft performance envelope estimation. J Aircr

50:410–424. https://doi.org/10.2514/1.C031847

14. Li T, Tang X, Ge J, Fei S (2020) Event-based fault-tolerant

control for networked control systems applied to aircraft engine

system. Inf Sci (NY) 512:1063–1077. https://doi.org/10.1016/j.

ins.2019.10.039

15. Tao X, Li N, Li S (2016) Multiple model predictive control for

large envelope flight of hypersonic vehicle systems. Inf Sci (NY)

328:115–126. https://doi.org/10.1016/j.ins.2015.08.033

16. Yoon S, MacGregor JF (2000) Statistical and causal model-based

approaches to fault detection and isolation. AIChE J

46:1813–1824. https://doi.org/10.1002/aic.690460910

17. Basir O, Yuan X (2007) Engine fault diagnosis based on multi-

sensor information fusion using Dempster–Shafer evidence the-

ory. Inf Fusion 8:379–386. https://doi.org/10.1016/j.inffus.2005.

07.003

18. Schwartz S, Montero Jimenez JJ, Salaün M, Vingerhoeds R

(2020) A fault mode identification methodology based on self-

organizing map. Neural Comput Appl. https://doi.org/10.1007/

s00521-019-04692-x

19. Melnyk I, Banerjee A, Matthews B, Oza N (2016) Semi-Markov

switching vector autoregressive model-based anomaly detection

in aviation systems. In: Proceedings of ACM SIGKDD interna-

tional conference on knowledge discovery and data mining

13–17-August, pp 1065–1074. https://doi.org/10.1145/2939672.

2939789

20. Rouhani S, Sheikholeslami A, Hosseini H, Kazemi H (2013)

Application of fuzzy subtractive clustering for optimal transient

performance of automatic generation control. IETE J Res 59:753.

https://doi.org/10.4103/0377-2063.126967

21. Lee H, Li G, Rai A, Chattopadhyay A (2020) Real-time anomaly

detection framework using a support vector regression for the

safety monitoring of commercial aircraft. Adv Eng Inform

44:101071. https://doi.org/10.1016/j.aei.2020.101071

22. Zhou S, Zhou Y, Xu Z, Chang W, Cheng Y (2019) The landing

safety prediction model by integrating pattern recognition and

Markov chain with flight data. Neural Comput Appl 31:147–159.

https://doi.org/10.1007/s00521-018-3669-9

23. Li L, Hansman RJ, Palacios R, Welsch R (2016) Anomaly

detection via a Gaussian mixture model for flight operation and

safety monitoring. Transp Res Part C 64:45–57. https://doi.org/

10.1016/j.trc.2016.01.007

24. Li G, Lee H, Rai A, Chattopadhyay A (2020) Analysis of oper-

ational and mechanical anomalies in scheduled commercial

flights using a logarithmic multivariate Gaussian model. Transp

Res Part C Emerg Technol 110:20–39. https://doi.org/10.1016/j.

trc.2019.11.011

25. Hosseini H, Tousi B, Razmjooy N, Khalilpour M (2013) Design

robust controller for automatic generation control in restructured

power system by imperialist competitive algorithm. IETE J Res

59:745. https://doi.org/10.4103/0377-2063.126971

26. Lee H, Li G, Rai A, Chattopadhyay A (2019) Health monitoring

framework for aircraft engine system using deep neural network.

In: PHM, vol 2019, pp 1–8. https://doi.org/10.36001/phmconf.

2019.v11i1.869

27. Razmjooy N, Ramezani M, Namadchian A (2016) A new LQR

optimal control for a single-link flexible joint robot manipulator

based on grey wolf optimizer. Majlesi J Electr Eng 10:53

28. Oehling J, Barry DJ (2019) Using machine learning methods in

airline flight data monitoring to generate new operational safety

knowledge from existing data. Saf Sci 114:89–104. https://doi.

org/10.1016/j.ssci.2018.12.018

29. Kiakojoori S, Khorasani K (2016) Dynamic neural networks for

gas turbine engine degradation prediction, health monitoring and

prognosis. Neural Comput Appl 27:2157–2192. https://doi.org/

10.1007/s00521-015-1990-0

30. Khalilpour M, Valipour K, Shayeghi H, Razmjooy N (2013)

Designing a robust and adaptive PID controller for gas turbine

connected to the generator. Res J Appl Sci Eng Technol

5:1544–1551. https://doi.org/10.19026/rjaset.5.4902

31. Mir M, Dayyani M, Sutikno T, Mohammadi Zanjireh M,

Razmjooy N (2020) Employing a Gaussian particle swarm opti-

mization method for tuning multi input multi output-fuzzy system

as an integrated controller of a micro-grid with stability analysis.

Comput Intell 36:225–258. https://doi.org/10.1111/coin.12257

32. Hinton GE (2006) Reducing the dimensionality of data with

neural networks. Science (80-) 313:504–507. https://doi.org/10.

1126/science.1127647

33. Vincent P, Larochelle H, Lajoie I, Bengio Y, Manzagol PA

(2010) Stacked denoising autoencoders: learning useful repre-

sentations in a deep network with a local denoising criterion.

J Mach Learn Res 11:3371–3408

34. An J, Cho S (2015) Variational autoencoder based anomaly

detection using reconstruction probability. Spec Lect IE. https://

doi.org/10.1007/BF00758335

35. Zhou C, Paffenroth RC (2017) Anomaly detection with robust

deep autoencoders. In: Proceedings of ACM SIGKDD interna-

tional conference on knowledge discovery and data mining part

F1, vol 296, pp 665–674. https://doi.org/10.1145/3097983.

3098052

36. Chalapathy R, Chawla S (2019) Deep learning for anomaly

detection: a survey, pp 1–50

37. Peng Z, Xiao X, Hu G, Kumar Sangaiah A, Atiquzzaman M, Xia

S (2020) ABFL: an autoencoder based practical approach for

Neural Computing and Applications (2021) 33:3235–3250 3249

123

software fault localization. Inf Sci (NY) 510:108–121. https://doi.

org/10.1016/j.ins.2019.08.077

38. National Aeronautics and Space Administration (2020) DASH-

link sample flight data. https://c3.nasa.gov/dashlink/projects/85/

39. United States Department of Transportation (2020) Bureau of

transportation statistics. https://www.transtats.bts.gov/

40. Arneson HM, Hegde P, La Scola ME, Evans AD, Keller RM,

Schade JE (2019) Sherlock data warehouse

41. National Transportation Safety Board (2020) Aviation accident

database. https://www.ntsb.gov/_layouts/ntsb.aviation/index.aspx

42. National Transportation Safety Board (2010) NTSB ID:

DCA09MA027

43. The Dutch Safety Board (2010) Turkish Airlines, crashed during

approach, Boeing 737–800, near Amsterdam Schiphol Airport

44. Milic L (2009) Multirate filtering for digital signal processing.

IGI Global, Hershey

45. Liu Y, Dang B, Li Y, Lin H, Ma H (2016) Applications of

Savitzky–Golay filter for seismic random noise reduction. Acta

Geophys 64:101–124. https://doi.org/10.1515/acgeo-2015-0062

46. Nair V, Hinton GE (2010) Rectified linear units improve

restricted Boltzmann machines. In: Proceedings of the 27th

international conference on machine learning (ICML-10),

pp 807–814

47. Ioffe S, Szegedy C (2015) Batch normalization: accelerating deep

network training by reducing internal covariate shift. In: 32nd

international conference on machine learning ICML 2015, vol 1,

pp 448–456

48. Amraee S, Vafaei A, Jamshidi K, Adibi P (2018) Anomaly

detection and localization in crowded scenes using connected

component analysis. Multimed Tools Appl 77:14767–14782.

https://doi.org/10.1007/s11042-017-5061-7

49. Kromanis R, Kripakaran P (2013) Support vector regression for

anomaly detection from measurement histories. Adv Eng Inform.

https://doi.org/10.1016/j.aei.2013.03.002

50. Gang N, Lau D, Pecht M (2010) Improving computer manufac-

turing management through Lean Six Sigma and PHM. In: 2010

prognostics and system health management conference. IEEE,

pp 1–7

51. Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdi-

nov R (2014) Dropout: a simple way to prevent neural networks

from overfitting. J Mach Learn Res 15:1929–1958

52. Kingma DP, Ba J (2014) Adam: a method for stochastic opti-

mization, pp 1–15

53. Washington SP, Karlaftis MG, Mannering F (2003) Statistical

and econometric methods for transportation data analysis.

Chapman and Hall/CRC, Boca Raton

54. Cook MV (2012) Flight dynamics principles: a linear systems

approach to aircraft stability and control. Butterworth-Heine-

mann, Oxford

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

3250 Neural Computing and Applications (2021) 33:3235–3250

123