See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/365089280

A digital twin framework for aircraft hydraulic systems failure detection using
machine learning techniques

Article in ARCHIVE Proceedings of the Institution of Mechanical Engineers Part C Journal of Mechanical Engineering Science 1989-1996 (vols 203-210) · November 2022
DOI: 10.1177/09544062221132697

CITATIONS READS

6 542

2 authors, including:

Hakki Ozgur Unver

TOBB University of Economics and Technology
64 PUBLICATIONS 954 CITATIONS

SEE PROFILE

All content following this page was uploaded by Hakki Ozgur Unver on 08 November 2022.

The user has requested enhancement of the downloaded file.

 A DIGITAL TWIN FRAMEWORK FOR
AIRCRAFT HYDRAULIC SYSTEMS FAILURE
DETECTION USING MACHINE LEARNING
TECHNIQUES
Furkan Kosova
Turkish Aerospace Industries, Inc., Ankara, Turkey
[email protected], [email protected]

Hakki Ozgur Unver (Corresponding author)

TOBB University of Economics and Technology, Department of Mechanical Engineering,
Ankara, Turkey.
e-mail: [email protected]
ASME Member

ABSTRACT
Since the last decade, aircraft systems, such as flight control and landing gear, have been requiring

increasing power, and consequently, the complexity of hydraulic aircraft systems has escalated. Inevitably,

this complexity has resulted in the need for the troubleshooting of hydraulic aircraft systems that are

dispersed around an aircraft and supply power to critical flight systems. In this context, this study develops

a novel digital twin-based health monitoring system to enable diagnostics of system failures early in the

design cycle using machine learning (ML) methods, instead of in the expensive downstream testing and

verification activities. The scope of the systems is limited to hydraulic systems at the aircraft level using 20

failure scenarios. The support vector machine and several ensemble learning algorithms of ML methods

were used to identify these failures. Comparison of the performance results revealed that the random forest

algorithm is superior to the other ML methods.

KEYWORDS: Digital Twin, Aircraft Hydraulics, Failure Detection, SVM, Ensemble Learning

1
 1. INTRODUCTION

Aircraft maintenance costs, which constitute a major operating cost in the

modern aircraft industry, cause giant aviation companies, such as Boeing and Airbus, to

spend 9% of their revenue on aircraft maintenance in recent years [1]. In this industry,

the main requirement of diagnostics systems is to raise an alarm when a failure occurs

and indicate its source. However, even using existing advanced technologies, online

diagnosis at the vehicle level is still difficult because current air vehicles consist of

numerous complex subsystems and components [2, 3]. Owing to the presence of these

complex systems and the increased demand for shorter the service time, the digital twin

concept was introduced in the "Technology Area (TA)" road map published in 2010 by

NASA. Since then, there have been serious efforts to use it in the next-generation

aircrafts [4, 5].

In modern civil aircrafts, hydraulic systems are composed of highly complex

systems that must be constantly monitored. However, the structure of a hydraulic

system makes it difficult to execute online diagnostics when a failure occurs, as in the

A343 Helsinki Finland 2009 incident. The failure, which emerged as a "too hot

temperature error" in the landing gear during the aircraft takeoff, was detected as a

hydraulic system leak 6 min after the flight [6, 7]. Although, the aircraft landed safely,

one of the two hydraulic lines became empty when the aircraft stopped.

In view of the Helsinki incident, which did not involve loss of life, continuous

online health monitoring at the aircraft level is now vital for civil aircrafts. Because

2
present aircraft systems are highly interconnected and hydraulic systems are major hubs

providing power to systems such as flight control, landing gear, and doors, it is difficult

to detect the source of a failure. The most advanced approach to overcome this

problem is to follow a model-based methodology using a developed digital imitation

model. In this concept, also called as digital twin, current products, such as a modern

aircraft, are modeled in a virtual environment and data are collected from all the

phases, starting from production to delivery to the end-user. Subsequently a concrete

decision is taken based on the data, and it is revealed whether the products perform

well at these levels. In the early design phase, virtual models provide enormous time-

saving when comparing different design options. In the testing and verification phase,

digital twin models allow design engineers to make design changes by acting rapidly

based on the accuracy of their products and the feedback from the customers after the

product delivery.

Using this concept, which has been a remarkable development in recent years,

the objectives are to virtualize products physically and functionally in the product life

cycle, capture data from all phases, and visualize the findings virtually to reduce high

development costs and time losses. With the advancement of simulation technologies,

the digital twin technology is currently being used in several fields, from aviation to the

automotive industries. [8-11].

Since the early 2000s, health management studies have gained importance in

terms of the diagnostics and prognostics of aircraft systems [12-14]. Over the past

decade, artificial intelligence (AI)-based approaches to detect and isolate individual

3
failures in aircraft subsystems have gained tremendous prominence both in academics

and the industry [15-17]. The objective of such fault detection systems, known as

diagnostics, is to identify the device or equipment that fails in a system by constantly

comparing the data of the sensors in the operating system using machine learning (ML)

methods. The application of online diagnostic systems at the aircraft level in many

aircraft models, such as A320 and Boeing 737 and F35, which are currently accepted as

modern aircrafts in the field of aviation, has strengthened the confidence in this

approach.

With the rising prominence of digital twin technology, it has recently become

feasible to use this simulation method for failure detection. [18, 19]. The use of

supervised learning technique, one of the most important types of ML, to detect failures

has been reported in the literature [20, 21]. In this study, support vector machines

(SVMs) and ensemble learning methods, one of the most important types of supervised

learning, were preferred using labeled data. These tags are entered manually by a user

within a digital twin environment with different codes based on the types of health

operation and failures.

The model proposed in this study can determine the source and location of a

randomly generated fault injection using the above mentioned ML methods by injecting

different faults, and the data are collected by labeled values.

As a part of this study, a hydraulic system was modeled on AMESIMTM at the

aircraft level and data were obtained by providing fault injections separately using

Simulink. Considering the scope and depth of this research, there is no study in the

4
literature on digital twin generation to detect aircraft-level failures in a hydraulic system

and in which the findings are contrasted with those of supervised classification

techniques. This paper presents a new and fully integrated hydraulic system digital twin

for a civil aircraft to detect failures by an online diagnostic approach.

2. SYSTEM ARCHITECTURE OF DIGITAL TWIN

For the model-based digital twin model we generate for this study, an aircraft-

level simulation environment is created for the hydraulic system using two engine-

driven pumps and two backup pumps. The hydraulic system is developed considering a

four-engine passenger aircraft. A schematic of the aircraft is shown in Figure 1.

Fig. 1 Four-engine aircraft hydraulic system

The objectives of this study are to model a civil aircraft hydraulic system in a

digital twin setting to analyze the system under healthy and faulty conditions and to

develop and assess ML models to detect hydraulic system failures. The developed model

collects the data from pressure, temperature, level, flow rate, and speed sensors to

monitor the health status of each component. The digital twin model designed in the

AMESIMTM environment is shown in Fig. 2.

Fig. 2 AMESIM model for aircraft hydraulic systems

In commercial aircraft systems, the necessary power is transmitted by at least

two separate hydraulic lines to the flight control and landing gear systems [22]. In this

pilot study, the second hydraulic line is ignored for the purpose of implementing a new

fault detection system.

5
 Figure 3 displays the digital twin for primary and secondary hydraulic flight

control equipment as well as equipment for the landing gear braking system. Because

the main flight control equipment is typically composed of hydraulic actuators, the

digital twin is designed as a hydraulic cylinder when location monitoring must be

precise, as shown for rudders [23], ailerons [24], and elevator systems [25]. Secondary

flight control is appropriate to be modeled using a hydraulic motor, as shown in [26, 27].

Ming et al. [28] simulated a digital twin landing gear brake system, and the results

showed that a diagnostic system can be easily created by reducing the complexity of the

operation system.

Fig. 3 Hydraulic equipment used in digital twin

2.1 Framework

In the literature, there are two types of diagnostic methods for online fault

diagnosis in aircraft systems: model-based and data-driven diagnostic systems.

The working principle of model-based diagnostic systems is based on physics-

based systems. A modern model-based diagnostic system was introduced by

Mosterman and Biswas to create a framework to detect failures in liquid sodium cooling

systems [29]. Model-based systems are traditionally preferred for aircraft systems [30,

31]. However, it is not very realistic to establish a model-based diagnostic reasoner at

the aircraft system level. The main reason is that aircraft vehicle systems, such as

hydraulic systems, include numerous components, which simultaneously impact flight-

critical systems, such as the electric power, flight control, and fuel systems.

6
 Data-driven approaches are also emerging diagnostic methods. A data-driven

approach is designed to identify and separate faults from new data after training the

current data using statistical mathematics and ML algorithms [32, 33]. According to

Tidriri et al., big data-driven methods are preferred for complex systems in real time

because such systems do not require high computation while working dynamically [34].

The significant difference between model- and data-based methods is that in the

former, first the model must be obtained analytically, and subsequently large

calculations must be conducted while dynamically integrating it into the working

system. In comparison, in a data-driven approach, the model is built by training on the

data collected from the system sensors.

However, both methods are inadequate for the scope of the proposed study.

Therefore, the following model based on a hybrid method is suggested as in this study.

Figure 4 presents the hybrid model framework for an aircraft hydraulic diagnostic

reasoner.

Fig. 4 Framework for diagnostic reasoner

In the first step, digital twin system is developed. Faulty and healthy scenarios

are implemented separately on Siemens AMESIMTM software, and sensor outputs are

collected within the specified framework. Subsequently, the data are manually labeled

depending on the fault codes. For on the different scenarios, the sensor outputs are

stored.

In the second step, the raw data are cleaned from missing values. Subsequently,

all the features are scaled between 0 and 1, as shown on the right side in Fig. 4, because

7
data sources such as, temperature, pressure, and velocity, are on different scales.

Before training the model, each feature is presented based on its feature weight, and 4–

5 features with the largest feature weight are selected to be trained. To avoid the

problem of dimensionality, more than five features are not selected. Once the data are

ready, they are trained.

In the last step, the trained model is evaluated based on its accuracy rate and

receiver operating characteristic (ROC) curve. The trained Model is examined using

previously untrained data within the context of this study. The best model is selected

and configured to be used on a selected aircraft.

2.2 Description of Hydraulic System

Current hydraulic systems are supplied by at least two independent hydraulic

systems owing to a dual redundancy requirement. In our digital twin concept, one

hydraulic line is pressurized by two separate engine-driven pumps, as displayed in Figure

5.

Fig. 5 Overview of hydraulic system

Pumps are one of the most dynamic parts that operate continuously in an

aircraft hydraulic system, and they are demonstrated as one of the major sources of

failures [35]. Therefore, it is appropriate to focus on a detailed hydraulic pump design as

the most essential element of the digital twin. As a starting point, each pump is assumed

to be driven at a maximum flow rate of 250 L/min under a 3000-psi nominal operating

pressure condition. The power transmitted by a circular shaft from the engine is

expressed in Eq. (1).

8
 𝑆ℎ𝑎𝑓𝑡 𝑃𝑜𝑤𝑒𝑟(𝑘𝑊) = 𝐻𝑦𝑑𝑟𝑎𝑢𝑙𝑖𝑐 𝑃𝑜𝑤𝑒𝑟(𝑘𝑊) /𝜂𝑝𝑢𝑚𝑝 𝑒𝑓𝑓𝑖𝑐𝑖𝑒𝑛𝑐𝑦 . (1)

The pump and volume efficiency are assumed to be 0.90 and are 0.95,

respectively; therefore, the maximum displacement can be calculated using Eq. (2).

𝐹𝑙𝑜𝑤 𝑅𝑎𝑡𝑒, 𝑄 = 𝑛(𝑟𝑝𝑚) 𝑑𝑖𝑠𝑝. (𝑓𝑡3/𝑟𝑒𝑣) 𝜂𝑣𝑜𝑙𝑢𝑚𝑒𝑡𝑟𝑖𝑐 𝑒𝑓𝑓𝑖𝑐𝑖𝑒𝑛𝑐𝑦 , (2)

where n is the shaft speed.

The maximum displacement of each pump is calculated as approximately 1.60

ft3/rev.

To obtain the required shaft rotation speed to drive each pump, the maximum

torque is assumed to be 400 Nm. Hence, the required rotational speed is calculated

using Eq. (3).

𝑆𝑝𝑒𝑒𝑑(𝑟𝑝𝑚) = 𝑃𝑜𝑤𝑒𝑟 (𝑘𝑊) 9.5488 / 𝑇𝑜𝑟𝑞𝑢𝑒 (𝑁𝑚). (3)

The rotational speed to drive each engine-driven pump is determined to be 3000

rpm. This suggests that a piston-type hydraulic pump should be preferred and that the

above calculations of the operating state of the pumps are remarkably similar to those

of the pumps utilized in the present aircrafts. Piston-type hydraulic pumps are preferred

in the industry because they can have a high flow rate with an overall efficiency of

approximately 0.90 under high-pressure conditions [36]. In addition, these pumps can

operate effectively under both normal pumping mode and depressurized conditions.

In fluid systems, the hydraulic fluid contained in the reservoir is pressurized by a pump

and subsequently delivered to the actuators. Primary and secondary aircraft control

actuators, as shown in Fig. 5, are powered by pressurized hydraulic fluid. While

developing the digital twin in this study, the primary flight control, landing gear, and

9
door actuators were modeled as linear actuators. The linear actuator formula is

correctly obtained by considering the maximum force, called as the stall force, acting on

the actuators, which is calculated using Eq. (4).

𝐹𝑚𝑎𝑥 = 𝑃𝑠𝑚𝑎𝑥 (𝐴𝑒𝑥𝑡𝑒𝑛𝑑 𝑑𝑖𝑟𝑒𝑐𝑡𝑖𝑜𝑛 − 𝐴𝑟𝑒𝑡𝑟𝑎𝑐𝑡 𝑑𝑖𝑟𝑒𝑐𝑡𝑖𝑜𝑛 ), (4)

where Ps is the system pressure.

The key performance parameters of a linear actuator expand and retract its

velocity. To obtain the linear actuator velocity, the contact area of each actuator should

be determined considering the acting stall force. Simple velocity parameter formulas are

given in Eqs. (5) and (6)

𝑉𝑒𝑥𝑡𝑒𝑛𝑑 = 𝑄/𝐴𝑒𝑥𝑡𝑒𝑛𝑑 𝑑𝑖𝑟𝑒𝑐𝑡𝑖𝑜𝑛 . (5)

𝑄
𝑉𝑟𝑒𝑡𝑟𝑎𝑐𝑡 𝑑𝑖𝑟𝑒𝑐𝑡𝑖𝑜𝑛 = 𝐴 . (6)
𝑟𝑒𝑡𝑟𝑎𝑐𝑡 𝑑𝑖𝑟𝑒𝑐𝑡𝑖𝑜𝑛

Another type of hydraulic actuator is a motor. Slats and leading-edge flaps are

mostly driven by hydraulic motors. Hydraulic motors are structured considering the

optimal point in the torque–speed curve. To obtain the maximum motor displacement

Eq. (7) is considered.

𝐷𝑚 = 𝜏𝑚𝑎𝑥 /(𝜂𝑚𝑜𝑡𝑜𝑟 𝑒𝑓𝑓𝑖𝑐𝑖𝑒𝑛𝑐𝑦 𝑃𝑚𝑎𝑥 ) . (7)

The mechanical efficiency is taken as 0.95, and the maximum torque is assumed

to be as 500 Nm. The results calculated with the provided inputs suggest using 12.00

mL/rev hydraulic motor at a rotating speed of 1250 rpm.

Hydraulic circuits need to use a relief valve to keep the system in a certain pressure

range and a shut-off valve to ensure the flow direction. Because these equipment are

composed entirely of mechanical parts, they generally do not contain any sensors.

10
 In addition, Hydraunycoil FH 51 is selected as the hydraulic fluid because it is one

of the most preferred aviation hydraulic fluids in view of its wide working temperature

range and relatively high temperature and viscosity performance.

Heat exchangers are other essential equipment composed only of mechanical

components. Oil coolers are typically found in the hydraulic systems of aircraft suction

lines. From thermodynamic calculations, the relationship between the pressure drop

and the temperature rise of an enclosed hydraulic circuit can be calculated using Eq. (8).

Δ𝑇 = Δ𝑃 (𝜌 𝐶𝑝)−1 , (8)

where Cp is the specific heat.

The calculated pressure drop is 36 psi; hence, a heat exchanger is selected

considering the calculation results.

Piston-type high pressure accumulators are preferred in aircraft hydraulic

circuits to store hydraulic power under pressure. Accumulator size is obtained based on

the universal gas law and is calculated using Eq. (9).

𝑉𝑜 = Δ(𝑃2 /𝑃0 )/((𝑃2 /𝑃1 )1/1.6 − 1).(9)

The adiabatic index is assumed as 1.6. The accumulator volume is calculated to be 15 L.

Finally, the diameters of the suction and return lines are assumed to be 50 mm

each, and the delivery line is assumed to be 30 mm.

The healthy working conditions of the digital twin equipment are calculated from the

results obtained based on the above equations. The values listed in Table 1 reflect the

healthy working conditions at a nominal pressure of 3000 psi.

11
2.3 Healthy Conditions

The ranges in which aircraft hydraulic systems are expected to work under

healthy conditions are listed in Table 1.

Table 1 Assumed healthy working conditions

Using the above defined systems and inputs, the sensor data are obtained for a

simulated healthy flight condition reduced to 100 s, which are shown in Figure 6.

Fig. 6 Healthy working results of digital twin

2.4 Fault Injection Testing

In the designed digital twin, the hydraulic line is provided by two engine-driven

pumps. The operating pressure of the system is ensured by a relief valve so that these

pumps do not exceed the defined pressure rate (3000 psi). Srivyas et al. [37] identified

the following six common failure types of hydraulic pumps: defective bearing, high

vibration cavitation, misalignment, seal breakage, and leakage. To diagnose the

malfunction of a pump, it is important to observe its vibration, examine the pressure in

both the frequency and time ranges, and calculate the flow rate values immediately. If

one of the two pumps in each line is out of service, the other pump can provide the

necessary pressurized oil to the entire system. It is also consistent to track both the

pumps simultaneously. Faults are injected into the pumps by driving the shaft and

assigning random numbers to the variable values on the pump. Failing to open, closing

unexpectedly, leakage, and irregular operation are the most common errors observed in

pumps.

12
 Other critical equipment in hydraulic systems are solenoid-controlled valves. Ji x.

et al. [38] proposed that the most common types of faults in these valves are oil leaks.

As previously mentioned, the function of accumulators in a hydraulic circuit is to store

the required pressurized liquid under pressure. Hence, oil leakage is a major issue for

accumulators as well as valves. In the accumulators of the hydraulic circuits, fault

injection is provided directly over the accumulator pressure difference sensor and by

changing the adiabatic index. There are many methods to inject faults for hydraulic

actuators; however, there as only three common types of actuators that can be

simulated, i.e., fail-open, fail-closed, and fail-locked, by providing the maximum

extraction and retraction in the AMESIM software.

In this study, to observe the defects in the secondary flight control surfaces, the

faults in the hydraulic motor are simulated by cutting the fluid flowing to the hydraulic

motors. In addition, adding noise to the rotational speed sensor is another viable option

to simulate hydraulic motor faults.

While simulating a blockage, which is the most common type of fault in hydraulic

filters, a fault is injected by reducing the filter flow rate. The pressure drop that

inevitably occurs is followed by a pressure differential sensor.

The two most common forms of faults found in heat exchangers used for heat

balance in the hydraulic circuit are leakage and plugging. Leakage is simulated by

modifying the leakage flow rate, whereas plugging is simulated by changing the outlet

flow rate.

13
 The relief and shut-off valves consist of entirely mechanical elements that

maintain the hydraulic system of an aircraft at the required pressure and flow rates and

enable the system to operate safely. The flow rates of these valves can be adjusted to

simulate the failure of the opening and closing after a fault is injected.

The list of faults injected in the digital twin is provided in Table 2.

Table 2 Injected faults

All the faults and important parameters are obtained using the digital twin

developed using AMESIM software and are recorded by the MATLAB/Simulink interface.

The following assumptions are made to obtain the appropriate results from the

digital twin:

1. All the sensors are calibrated

2. There is no any oil leakage in the hoses

3. Cavitation is ignored

The results of the faults injected in the digital twin at specific locations and listed

in Table 2 are shown in Figure 7.

Fig. 7 Healthy and faulty working conditions of equipment

As can be seen from Figure 7, the failures detailed in Table 2 were injected

separately in the digital model. All these faults are shown in red in the figure, whereas

those shown in blue are under the healthy working conditions of the equipment. For

instance, the failure of the landing gear and flight control system actuators were

modeled similarly under different operating conditions. Moreover, the unsteady flow

inside each pump was modeled, the pump failure was injected to operate randomly. For

14
all the aircraft hydraulic systems and hydraulic systems, hydraulic equipment of

consumers, such as actuators, valves, motors, filters, and faults, were separately

injected in different times, and the sensor outputs were recorded in a data warehouse.

2. MACHINE LEARNING TECHNIQUES FOR AIRCRAFT HYDRAULIC SYSTEM FAILURE

DETECTIONS

In this section, SVMs and two ensemble learning methods, AdaBoost and

random forest algorithms, which are methods used for failure detection in aircraft

hydraulic systems, are discussed. Interest in ensemble learning approaches has been

increasing in recent years because their performance is comparable to those of SVMs

for troubleshooting and fault detection tasks [39].

3.1 Theoretical background for Support Vector Machines and Ensemble Learning
Methods
3.1.1. Support Vector Machines

According to Cervantes et al., SVMs is one of the most optimal classification

techniques and also one of the most well-known classification approaches [40]. The

SVM technique has been used for classification data sets in recent years and been

proven to have achieved satisfactory results in various different studies [41-43]. First,

the SVM algorithm was invented by Vapnik and Learner in 1963 for pattern recognition.

[44]. Subsequently, the widespread use of this classification technique by ML engineers

increased with the increase in the speed of computer technologies since the 2000s.

As detailed by Alpaydin, the theory of SVM can be summarized as follows [45,

46].

15
 Please consider the dataset shown in Eq. (10).

𝐷 (𝑋1 , 𝑌1 ), (𝑋2 , 𝑌2 ), . . . , (𝑋𝑛 , 𝑌𝑛 ), 𝑋𝑖 ∈𝑅 𝑑 , (10)

where X is the training tuples and Y is the classification label.

Let these two classes adopt labels +1 for 𝑌 + and -1 for 𝑌 − .

To determine w and 𝑤0 for the purpose of fully separated data, Eqs. (11) and (12)

are written as follows:

𝑤 𝑇 𝑥 𝑡 + 𝑤0 − 1 ≥ 0 𝑓𝑜𝑟 𝑌 + = +1 and (11)

𝑤 𝑇 𝑥 𝑡 + 𝑤0 + 1 ≤ 0 𝑓𝑜𝑟 𝑌 + = −1. (12)

It is assumed that w is the coefficient vector and 𝑤0 is the bias vector

Eqs. (11 & 12) can be converted into a single equation, as shown in Eq. (13),

𝑌𝑖 (𝑤 𝑇 𝑥 𝑡 + 𝑤0 ) ≥ 1. (13)

The process is summarized in Figure 8.

Fig. 8 Support vector machines

The SVM algorithm is based on measuring the margin at the highest possible

depth, thereby offering a hyperplane that divides the data into two distinct labels. To

achieve this, it is necessary to determine the maximum value of the margin separator,

as shown in Fig 9. Hence, the margin distance is calculated as follows:2/‖𝑤‖2

The vectors called as the SVMs are presented as broad green circles in Fig. 9,

where they are intersected by margin lines.

Obtaining the solution is an optimization problem. Using Lagrange multipliers 𝛼 𝑡

to reduce the complexity of Eqs. (14) and (15), the maximum margin can be determined

as expressed in Eq. (16).

16
 𝑇
𝐿(𝛼) = ∑𝑛𝑖=1 ∑𝑛𝑗=1 𝛼 𝑖 𝛼 𝑗 𝑌 𝑖 𝑌𝑗 (𝑥 𝑖 ) 𝑥 𝑗 + ∑𝑛𝑖 𝛼 𝑖 , (16)

s.t., ∑𝑛 𝑖 𝑖
𝑖 𝛼 𝑌 = 0, and 𝛼 𝑖 ≥ 0, ∀𝑖.

Quadratic optimization methods are sufficient to overcome the above problem.

In most cases, the majority of the samples in the model cannot be linearly separated,

and several exceptions can also be found in the margin, as seen in Fig. 9.

Fig. 9 Support vector machines and hyperplane example

Therefore, the solution is obtained by adding a slack variable , as shown in Eqs. (14)

and (15).

1
𝑚𝑖𝑛 (2 ‖𝑤‖2 + 𝐶 ∑𝑛𝑖=1 𝑖 ), (14)

𝑌𝑖 (𝑤𝑋𝑖 + 𝑏) ≥ 𝜌 − 𝑖
{ for i=1, 2, …, n,
>0

where 𝜌 is an optimization variable so that the new margin distance is 2𝜌/‖𝑤‖2 .

Using the Lagrange multipliers as previously described, the dual problem

solution is as expressed in Eqs. (16) and (17).

1 𝑇
𝐿(𝛼) = ∑𝑛𝑖 𝛼 𝑖 − 2 ∑𝑛𝑖 ∑𝑛𝑗 𝛼 𝑖 𝛼 𝑗 𝑌 𝑖 𝑌𝑗 (𝑋 𝑖 ) 𝑋𝑗 , (16)

s.t. ∑𝑛 𝑖 𝑖
𝑖 𝛼 𝑌 = 0, and 𝐶 ≥ 𝛼 𝑖 ≥ 0, ∀𝑖, (17)

where C is a penalty factor, which should be determined by cross-validation.

This case needs to be carefully examined. The vectors being inside the margin

and being difficult to differentiate are the primary reasons of the increased overall

error.

In addition to the above two cases, most of the examples encountered in the

real world are not separated by a linear line. The data obtained in the real world are

17
rarely suitable for linear separation. To avoid this problem, the kernel trick method

allows acquiring a linear separation vector by mapping the input vectors to high-

dimensional spaces, as shown in Fig. 10.

Fig. 10 Mapping for support vector machines

The kernel trick is a technique that enables linear separation by mapping the

available data to higher dimensions, instead of fitting a hyperplane directly to a non-

linearly distributed dataset. Briefly, the kernel trick ensures that the data that cannot be

separated linearly for a given dimension are separated linearly for a higher dimension.

The following formulas are obtained with a new higher dimensional acceptance,

which can be expressed by its simple functions for kernel trick operation, and the

discriminant is accepted as expressed in Eqs. (18)–(20).

𝑍 = ∅(𝑋), 𝑖𝑓 𝑍𝑑 = ∅𝑟 (𝑋), 𝑑 = 1, 2, . . . , 𝑘, (18)

𝑔(𝑍) = 𝑊 𝑇 𝑍, and (19)

𝑔(𝑍) = ∑𝑘𝑑=1 𝑊𝑑 ∅𝑑 (𝑋). (20)

In the following calculations, the Lagrangian equations are repeated as before.

The only distinction is that the defined constraints are tailored to new spaces, such as

expressed in Eq. (21).

𝑟 𝑡 𝑊 𝑇 = ∅(𝑋 𝑇 ) ≥ 1 − 𝑡 . (21)

Eqs. (22) and (23) show the results after applying the specified constraints to the

Lagrange multipliers.

1
𝐿(𝛼) = ∑𝑛𝑖 𝛼 𝑖 − 2 ∑𝑛𝑖 ∑𝑛𝑗 𝛼 𝑖 𝛼 𝑗 𝑌 𝑖 𝑌𝑗 𝐾 (𝑋 𝑖 , 𝑋𝑗 ), (22)

18
 s.t., 𝑔(𝑋) = ∑𝑛𝑖 𝛼 𝑖 𝑌 𝑖 𝐾 (𝑋 𝑖 , 𝑋) = 0. (23)

3.1.2. Ensemble Learning Methods

Ensemble learning is a set of algorithms that aims to achieve improved results by

combining several ML methods in the same algorithm. The ensemble learning structure

can be adapted to both supervised and unsupervised learning methods. Therefore,

ensemble learning methods have become highly well known in recent years [47, 48].

3.1.2.1. AdaBoost Algorithm

Among the boosting algorithms, the AdaBoost algorithm is generally preferred

for classification studies, instead of regression studies. It has begun to be favored

recently because it yields remarkable results in systems with challenging fault diagnosis.

[49]. Because diagnostic studies require a classification-based approach, this study is

conducted using the AdaBoost algorithm as the ensemble learning method, which is one

of the classification methods.

The AdaBoost algorithm was developed in 1997 by Schapire and Freund [50]. The

main objective of the AdaBoost algorithm is to increase the margin maximally, similar to

the SVM algorithm.

The AdaBoost algorithm consists of the following stages, which are shown in Fig.

11.

Fig. 11 AdaBoost algorithm steps

The stages can be formulated as follows:

19
 1. The train set is prepared as follows:

𝐷 (𝑋1 , 𝑌1 ), (𝑋2 , 𝑌2 ), . . . , (𝑋𝑛 , 𝑌𝑛 ), 𝑋𝑖 ∈𝑅 𝑑 , (24)

where X is the training tuples and Y is the classification labels.

Suppose these two classes adopt labels +1 for 𝑌 + and -1 for 𝑌 − .

For the weight vector, w, the following assumption is accepted:

1
𝑤0 = 𝑁, (25)

where N is the number of input vectors.

2. Weight Vector is calculated for each input vector by comparing the error rate

with 𝑌𝑗 using the weight vectors starting with the above assumption, as

expressed in Eq. (26).

∈𝑒𝑟𝑟𝑜𝑟 = ∑𝑛𝑖 𝑃𝑗𝑖 . 1(𝑦𝑗𝑖 ≠ 𝑌 𝑖 ). (26)

3. After the error rate is calculated, ∈𝑒𝑟𝑟𝑜𝑟 < 1/2 is compared for all the j inputs.

4. If the fourth stage is valid, the 𝛽𝑗 value is calculated for the whole input set,

as expressed in Eq. (27).

∈
𝛽𝑗 = 1−∈𝑗 (27)
𝑗

5. In this stage, the voting algorithm operates. If 𝑦𝑗𝑖 = 𝑌 𝑖 for each (X, Y) value,

𝑖
𝑃𝑗+1 = 𝛽𝑗 𝑃𝑗𝑖 is selected. Otherwise, 𝑃𝑗+1
𝑖
= 𝑃𝑗𝑖 is selected.

6. This algorithm is repeated until all the weak data are converted into strong

ones.

3.1.2.2. Random Forest Algorithm

20
 The random forest algorithm is an improved version of the decision tree

algorithm developed by Leo Breiman in 2001 [51]. The most important feature that

distinguishes this approach from decision trees is its lack of sensitivity to noisy data. The

random forest algorithm basically makes an inference expressed as a forest by

performing a regression over randomly generated decision trees. Because the random

forest decision algorithm is a very well-known algorithm among ensemble learning

methods, it is observed recently that many researchers prefer it while developing a fault

diagnosis model [52-53].

This algorithm uses the Gini index, which is expressed in Eq. (28), to measure the

homogeneity of the labeled data, as in the decision tree algorithm.

2
𝐺𝑖𝑛𝑖(𝑇) = 1 − ∑𝑛𝐽=1(𝑃𝐽 ) . (28)

After the Gini index is calculated, each class is defined on the decision trees.

Subsequently, based on the number of decision trees provided by the user in the

algorithm, n elements are randomly selected for each cluster, and an inference is made

from the training and testing along with the Gini index.

3.2 Feature Generation, Selection, and Extraction

The data generated in the digital twin were transferred to the MATLABTM

environment. In this study, a total of 201 distinctive features were extracted.

Approximately 80% of the dataset was separated for training, whereas 20% of the

remaining dataset was split for testing. Because the entire generated dataset is in the

21
time domain, the features are derived from the raw data in the forms of mean, median,

kurtosis, skewness, root mean square, peak to peak, and entropy.

However, given the vast scope of the analysis and injection of several different

fault scenarios, an automatic fault selection approach is required. Within the framework

of this study, using principal component analysis (PCA), which provides particularly

satisfactory results in fault detection projects involving supervised learning, is the most

reasonable option[54].

This study requires a separate PCA for each fault type based on the location of

each equipment. For instance, the important variables for the landing gear actuator

jamming fault differ from the flight control gear actuator one. Accordingly, the PCA

study was designed to run automatically for each fault detection classification. Fig. 12

shows the PCA results for this fault type, which indicates the important features of an

actuator.

Fig. 12 Landing gear actuator jamming failure PCA results

4. FAULT DETECTION RESULTS AND DISCUSSIONS

A comparison of the classification performance results is provided in this section.

After obtaining healthy and faulty simulation data and performing feature selection

processes, separate training on SVM and ensemble models are obtained for all the fault

codes listed in Table 2.

SVM algorithms were implemented directly on the classification application of

MATLABTM software. To train the dataset by SVM, the most well-known SVM types—

22
linear, quadratic, cubic, and Gaussian kernels —were used, and the best SVM type was

determined.

To prevent high-variance and high-bias problems, the dataset was trained by

splitting it into six separate parts using the K-fold method. Under the K-fold method, five

separate datasets were randomly chosen and trained, and the remaining set was used

as a validation tool. In addition, as Gangsar proposed, one of the most critical parts for

the SVM method is to determine the optimal regularization parameter, C, and the

kernel parameter,  [55]. These steps are repeated until the most optimal parameters

are obtained.

In addition, in the dataset training using the random forest algorithm, it was

suitable to select the learning rate as 0.1, maximum split number as 20, and number of

learners as 30.

The data collected from 201 different sensors in total were investigated

individually for 23 different failure types, and subsequently, different failure codes were

created for hundreds of components. Based on the results, both the SVM and ensemble

learning algorithms can classify the failures in the relief valve, shutoff valve, and filter

equipment with an accuracy rate of almost 100%.

Fig. 13 shows the ability of the random decision forest and SVM ML models to

detect faults in the aircraft hydraulic system for the test set. One can easily notice that

the random decision algorithm yields better results than the SVM method, particularly

in difficult to detect faults. Although both methods detect blockage and stuck failure

types with similar failure rates in the selector valve, it is observed that the random

23
forest algorithm provides better results in detecting oil leakage. The superiority of the

random forest algorithm over SVM is also seen for crucial equipment, such as hydraulic

pumps, hydraulic motors, and actuators.

Fig. 13 SVM vs. random forest algorithm

To determine which of these two proposed methods performs best in the

classification process comparison, it is important to specify the evaluation criteria. The

comparison of the accuracy, confusion matrix, and ROC curves as the classification

evaluation criteria was considered acceptable for this research.

Accuracy is a factor that defines the proximity of a measured value to the actual

value. For this research, the accuracy values were calculated via the AdaBoost algorithm

using both the SVM types and a Gaussian kernel. The accuracy results for all the SVMs

and the AdaBoost algorithm are tabulated in Table 3.

Table 3 Prediction accuracy

To further understand the findings obtained, it is important to analyze the ROC

curves more carefully. An ROC curve is typically expressed as a sensitivity/specificity for

each point, and it is a curve by which the classification properties of the methods can be

better expressed in the binary classification. In this curve, the slope of the upper left

part is considered to be better. As seen in Fig. 14, the failure of the LG1-Actuator, fail to

extraction, is represented using the random decision forest, and for the SVM, it is

represented by the ROC curve.

Fig. 14 ROC curve

24
 As shown by the accuracy rate and the ROC curve, the AdaBoost algorithm

clearly outperforms the SVM method for diagnosing failures in aircraft hydraulic

systems.

5. CONCLUSION

In this research, we aimed to diagnose failures by introducing the significance of

hydraulic aircraft systems and building an aircraft-level digital hydraulic twin, which

excels owing to the benefits it provides in the current aviation world. As shown in the

Results section, the ensemble learning algorithm, AdaBoost, provides excellent results

at the aircraft level compared to SVMs, with the appropriate feature selection approach.

The great success of this study compared to other research is that, it definitely

allows detecting all the failures fully by developing a digital twin in a virtual environment

before installing a test rig on the ground. Following this research, a digital twin hydraulic

system should be developed at the aircraft level, which will be highly precise and

specific to each aircraft and enable diagnostic detection of all the system failures.

The proposed method will assist in the detection of failures in vehicle hydraulic

systems, which are the main power distributors in current aircrafts, during health

monitoring studies, which have become critical for the aircraft industry. Moreover, it

will help in managing certification tasks owing to the high accuracy of the hydraulic

reasoning system.

The digital twin methodology developed in this study can be adapted to the

development cycles of many vehicle programs, helping designers to develop and test

25
diagnostics systems early in the design phase. This will prevent expensive time loss and

failures in the downstream phases, such as product verification and certification.

6. FUTURE WORKS

The results of this study, which are discussed above, should be proven with real

flight data, exhibiting that ML models yield correct classification. Other future research

activities would involve using deep learning methods, such as convolutional neural

networks and long-short term memory, for the detection of failures in vehicle systems.

This is because these new methods are proven to be very effective in systems where big

data are generated.

In the next phase of this project, a physical test rig will be built to verify the

operation of the proposed digital twin approach. The high-level architecture of the

envisioned digital twin architecture with integrated test rig is given in Fig 15.

Fig. 15 The envisioned digital twin architecture. (The Physical model is an

illustration from the GrabCAD repository https://grabcad.com/library/hydraulic-test-

bench-100-kw-1/details?folder_id=1290612)

Considering the flight envelope parameters and failure modes and effect analysis

(FMEA) results, simulation data will be generated by the AMESIMTM model for healthy

system parameters and fail system scenarios. Afterward, the simulated data will be used

to train multiple ML models. When the test rig is integrated, Hardware-in-the-loop (HIL)

26
simulation technique will be required as a fault isolator and fail-safe unit between the

physical and digital twins. In order to fine-tune and calibrate the sensor

types/numbers/locations on the test rig, first, datasets for many alternative scenarios can

be easily generated by the AMESIM model in the digital twin, and then the ML models

trained with the simulated datasets will be tested in the physical twin. Finally, when the

digital twin is verified in an actual flight environment, the digital twin system will be ready

for aviation certification procedures.

ACKNOWLEDGMENT

The authors would like to express their deep gratitude to Osman Oğuz, Sinan

Küçük, and Dr. Özkan Altay for their contributions and useful critiques of this study.

Finally, the authors would like to thank all the TF-X project members of the

Turkish Aerospace Industries, Inc. for their invaluable support.

27
REFERENCES

[1] IATA, "Airline Maintenance Cost Executive Commentary,”

https://www.iata.org/contentassets/bf8ca67c8bcd4358b3d004b0d6d0916f/mctg-fy2018-
report-public.pdf, (2019, Accessed at 15 Feb 2021)

[2] Mengshoel OJ, Darwiche A, Cascio K, Chavira M, Poll S, Uckun NS. Diagnosing Faults
in Electrical Power Systems of Spacecraft and Aircraft. InAAAI 2008 Jul 13 (pp. 1699-
1705).

[3] Byington CS, Roemer MJ, Galie T. Prognostic enhancements to diagnostic systems for
improved condition-based maintenance [military aircraft]. InProceedings, ieee
aerospace conference 2002 Mar 9 (Vol. 6, pp. 6-6). IEEE.

[4] Piascik B, Vickers J, Lowry D, Scotti S, Stewart J, Calomino A. Materials, Structures,

Mechanical Systems, and Manufacturing Roadmap-Technology Area 12. National
Aeronautics and Space Administration (NASA): Washington, DC, The United States of
America. 2012.

[5] Glaessgen E, Stargel D. The digital twin paradigm for future NASA and US Air Force
vehicles. In53rd AIAA/ASME/ASCE/AHS/ASC structures, structural dynamics and
materials conference 20th AIAA/ASME/AHS adaptive structures conference 14th AIAA
2012 Apr 23 (p. 1818).

[6] "A343, Helsinki Finland, 2009,”

https://www.skybrary.aero/index.php/A343,_Helsinki_Finland,_2009 (Accessed at 15
May 2021)

[7] Hradecky S, "Incident: Finnair A343 on Jun 22nd 2009, Main Wheel Tyre Damage
Leads to Hydraulics Leak and Steering Failure,” http://avherald.com/h?article=41b9cd1f
(2009, Accessed at 15 Feb 2021)

[8] Xu W, Cui J, Li L, Yao B, Tian S, Zhou Z. Digital twin-based industrial cloud robotics:
Framework, control approach and implementation. Journal of Manufacturing Systems.
2021 Jan 1;58:196-209.

[9] Yang J, Lee S, Kang YS, Noh SD, Choi SS, Jung BR, Lee SH, Kang JT, Lee DY, Kim HS.
Integrated platform and digital twin application for global automotive part suppliers.
InIFIP International Conference on Advances in Production Management Systems 2020
Aug 30 (pp. 230-237). Springer, Cham.

[10] Moslått GA, Padovani D, Hansen MR. A digital twin for lift planning with offshore
heave compensated cranes. Journal of Offshore Mechanics and Arctic Engineering. 2021
Jun 1;143(3).

28
[11] Guerreiro G, Figueiras P, Costa R, Marques M, Graça D, Garcia G, Jardim-Gonçalves
R. A digital twin for intra-logistics process planning for the automotive sector supported
by big data analytics. InASME International Mechanical Engineering Congress and
Exposition 2019 Nov 11 (Vol. 59384, p. V02BT02A021). American Society of Mechanical
Engineers.

[12] Massam S, McQuillan S. Verification of PHM capabilities: A joint customer/industrial

perspective. In Proceedings, IEEE Aerospace Conference 2002 Mar 9 (Vol. 6, pp. 6-6).
IEEE.

[13] Litt JS, Simon DL, Garg S, Guo TH, Mercer C, Millar R, Behbahani A, Bajwa A, Jensen
DT. A survey of intelligent control and health management technologies for aircraft
propulsion systems. Journal of Aerospace Computing, Information, and Communication.
2004 Dec;1(12):543-63.

[14] Li A, Zhang W, Tan J. Survey on aircraft health management technology. Electronics

Optics & Control. 2007 Jun;14(3):79-83.

[15] Perhinschi MG, Moncayo H, Davis J. Integrated framework for artificial immunity-
based aircraft failure detection, identification, and evaluation. Journal of Aircraft. 2010
Nov;47(6):1847-59.

[16] Zolghadri A, Cieslak J, Efimov D, Henry D, Goupil P, Dayre R, Gheorghe A, Leberre H.

Signal and model-based fault detection for aircraft systems. IFAC-PapersOnLine. 2015
Jan 1;48(21):1096-101.

[17] Swischuk R, Allaire D. A machine learning approach to aircraft sensor error

detection and correction. Journal of Computing and Information Science in Engineering.
2019 Dec 1;19(4).

[18] Zhou K, Yang S, Guo Z, Long X, Hou J, Jin T. Design of automatic spray monitoring
and tele-operation system based on digital twin technology. Proceedings of the
Institution of Mechanical Engineers, Part C: Journal of Mechanical Engineering Science.
2021 Dec;235(24):7709-25.

[19] Seshadri BR, Krishnamurthy T. Structural health management of damaged aircraft

structures using digital twin concept. In25th aiaa/ahs adaptive structures conference
2017 (p. 1675).

[20] Pose CD, Etchemaite JI, Giribet JI. Multirotor fault detection based on supervised
learning. InProceedings of the 2020 Argentine Conference on Automatic Control
(AADECA), CABA, Argentina 2020.

29
[21] Resendiz-Ochoa E, Saucedo-Dorantes JJ, Benitez-Rangel JP, Osornio-Rios RA,
Morales-Hernandez LA. Novel methodology for condition monitoring of gear wear using
supervised learning and infrared thermography. Applied Sciences. 2020 Jan;10(2):506.

[22] Wang S, Tomovic M, Liu H. Commercial Aircraft Hydraulic Systems: Shanghai Jiao
Tong University Press Aerospace Series. Academic Press; 2015 Oct 9.

[23] Li Y, Quan Q, Li H, Tang D, Li Z, Fan W, Deng Z. Air rudder mechanism dynamics

considering two elements: Joint clearance and link flexibility. Journal of Mechanical
Science and Technology. 2017 Jul;31(7):3189-97.

[24] Zong W, Wan F, Wei Y. Real-time monitoring for the actuator mechanism of the
aileron. In2017 Prognostics and System Health Management Conference (PHM-Harbin)
2017 Jul 9 (pp. 1-5). IEEE.

[25] Mosterman PJ, Pereira Remelhe MA, Engell S, Otter M. Simulation for analysis of
aircraft elevator feedback and redundancy control. InModelling, Analysis, and Design of
Hybrid Systems 2002 (pp. 369-390). Springer, Berlin, Heidelberg.

[26] Rea J. Boeing 777 high lift control system. InProceedings of the IEEE 1993 National
Aerospace and Electronics Conference-NAECON 1993 1993 May 24 (pp. 476-483). IEEE.

[27] Biedermann O, Geerling G. Power control units with secondary controlled hydraulic
motors-a new concept for application in aircraft high lift systems.

[28] Ming Z, Hong N, Xiao‐hui W, Xiaomei Q, Enzhi Z. Modeling and simulation of aircraft
anti‐skid braking and steering using co‐simulation method. COMPEL-The international
journal for computation and mathematics in electrical and electronic engineering. 2009
Nov 13.

[29] Mosterman PJ, Biswas G. Diagnosis of continuous valued systems in transient

operating regions. IEEE Transactions on Systems, Man, and Cybernetics-Part A: Systems
and Humans. 1999 Nov;29(6):554-65.

[30] Gorinevsky D, Dittmar K, Mylaraswamy D, Nwadiogbu E. Model-based diagnostics

for an aircraft auxiliary power unit. InProceedings of the international conference on
control applications 2002 Sep 18 (Vol. 1, pp. 215-220). IEEE.

[31] Polverino P, Esposito A, Pianese C, Ludwig B, Iwanschitz B, Mai A. On-line

experimental validation of a model-based diagnostic algorithm dedicated to a solid
oxide fuel cell system. Journal of Power Sources. 2016 Feb 29;306:646-57.

[32] Wu Y, Tao R, Zhu D, Yao Z, Xiao R. A machine-learning approach to predicting the

energy conversion performance of centrifugal pump impeller influenced by blade

30
profile. Proceedings of the Institution of Mechanical Engineers, Part C: Journal of
Mechanical Engineering Science. 2021 Dec;235(24):7363-84.

[33] Naderi E, Khorasani K. Data-driven fault detection, isolation and estimation of

aircraft gas turbine engine actuator and sensors. Mechanical Systems and Signal
Processing. 2018 Feb 1;100:415-38.

[34] Tidriri K, Chatti N, Verron S, Tiplica T. Bridging data-driven and model-based

approaches for process fault diagnosis and health monitoring: A review of researches
and future challenges. Annual Reviews in Control. 2016 Jan 1;42:63-81.

[35] Xu G, Ma C, Gao Z, Hu X, Luo Y. Modeling and simulation of aero-hydraulic pump

wear failure. In2017 Prognostics and System Health Management Conference (PHM-
Harbin) 2017 Jul 9 (pp. 1-7). IEEE.

[36] Eaton, 2013, "Eaton’s Pump and Motor Products",

https://www.eaton.com/ecm/groups/public/@pub/@eaton/@hyd/documents/content
/pll_2509.pdf

[37] Srivyas PD, Singh S, Singh B. Study of Various Maintenance Approaches Types of
Failure and Failure Detection Techniques Used in Hydraulic Pumps: A REVIEW.
INDUSTRIAL ENGINEERING JOURNAL. 2017 May.

[38] Ji X, Ren Y, Tang H, Shi C, Xiang J. An intelligent fault diagnosis approach based on
Dempster-Shafer theory for hydraulic valves. Measurement. 2020 Dec 1;165:108129.

[39] Rezaei Aderiani A, Wärmefjord K, Söderberg R, Lindkvist L. Individualizing locator

adjustments of assembly fixtures using a digital twin. Journal of Computing and
Information Science in Engineering. 2019 Dec 1;19(4).

[40] Cervantes J, Garcia-Lamont F, Rodríguez-Mazahua L, Lopez A. A comprehensive

survey on support vector machine classification: Applications, challenges and trends.
Neurocomputing. 2020 Sep 30;408:189-215.

[41] Yu R, Li X, Tao M, Ke Z. Fault Diagnosis of Feedwater Pump in Nuclear Power Plants

Using Parameter-Optimized Support Vector Machine. In International Conference on
Nuclear Engineering 2016 Jun 26 (Vol. 50015, p. V001T03A013). American Society of
Mechanical Engineers.

[42] Fan X, Zhang L, Liang W, Wang Z. Leak Detection Method Based on Support Vector
Machine. InInternational Pipeline Conference 2008 Jan 1 (Vol. 48579, pp. 517-522).

[43] Bordoloi, D.J. and Tiwari, R., 2013, “Optimization of support vector machine based
multi-fault classification with evolutionary algorithms from time domain vibration data

31
of gears,” Proceedings of the Institution of Mechanical Engineers, Part C: Journal of
Mechanical Engineering Science, 227(11), pp.2428-2439.

[44] Vapnik V. Pattern recognition using generalized portrait method. Automation and
remote control. 1963;24:774-80.

[45] Alpaydin E. Introduction to machine learning. MIT press; 2020 Mar 24.

[46] Alpaydin, E., https://www.cmpe.boun.edu.tr/~ethem/i2ml2e/index.html

[47] Wu Z, Lin W, Ji Y. An integrated ensemble learning model for imbalanced fault

diagnostics and prognostics. IEEE Access. 2018 Feb 19;6:8394-402.

[48] Li Z, Goebel K, Wu D. Degradation modeling and remaining useful life prediction of

aircraft engines using ensemble learning. Journal of Engineering for Gas Turbines and
Power. 2019 Apr 1;141(4).

[49] Shen F, Langari R, Yan R. Exploring Sample/Feature Hybrid Transfer for Gear Fault
Diagnosis Under Varying Working Conditions. Journal of Computing and Information
Science in Engineering. 2020 Aug 1;20(4).

[50] Freund Y, Schapire RE. A decision-theoretic generalization of on-line learning and an

application to boosting. Journal of computer and system sciences. 1997 Aug
1;55(1):119-39.

[51] Breiman L. Random forests. Machine learning. 2001 Oct;45(1):5-32.

[52] Tian J, Ai Y, Zhao M, Fei C, Zhang F. Fault Diagnosis Method for Inter-Shaft Bearings
Based on Information Exergy and Random Forest. InTurbo Expo: Power for Land, Sea,
and Air 2018 Jun 11 (Vol. 51128, p. V006T05A017). American Society of Mechanical
Engineers.

[53] Patil S, Patil A, Handikherkar V, Desai S, Phalle VM, Kazi FS. Remaining useful life
(rul) prediction of rolling element bearing using random forest and gradient boosting
technique. InASME international mechanical engineering congress and exposition 2018
Nov 9 (Vol. 52187, p. V013T05A019). American Society of Mechanical Engineers.

[54] Ayati M, Shirazi FA, Ansari-Rad S, Zabihihesari A. Classification-based fuel injection

fault detection of a trainset diesel engine using vibration signature analysis. Journal of
Dynamic Systems, Measurement, and Control. 2020 May 1;142(5):051003.

[55] Gangsar P, Tiwari R. Multifault diagnosis of induction motor at intermediate

operating conditions using wavelet packet transform and support vector machine.
Journal of Dynamic Systems, Measurement, and Control. 2018 Aug 1;140(8).

32
33
 Figure Captions List

Fig. 1 Four-engine aircraft hydraulic system

Fig. 2 AMESIM model for aircraft hydraulic systems

Fig. 3 Hydraulic equipment used in digital twin

Fig. 4 Framework for diagnostic reasoner

Fig. 5 Overview of hydraulic system

Fig. 6 Healthy working results of digital twin

Fig. 7 Equipment’s healthy and faulty working conditions

Fig. 8 Support vector machines

Fig. 9 Support vector machines and hyperplane example

Fig. 10 Mapping for support vector machines

Fig. 11 AdaBoost algorithm steps

Fig. 12 Landing gear actuator jamming failure PCA results

Fig. 13 SVM vs. random forest algorithm

Fig. 14 ROC curve

Fig 15 The envisioned digital twin architecture

34
Table Caption List

Table 1 Assumed healthy working conditions

Table 2 Injected faults

Table 3 Prediction accuracy

35
Fig. 1 Four-engine aircraft hydraulic system

36
Fig. 2 AMESIM model for aircraft hydraulic systems

37
Fig. 3 Hydraulic equipment used in digital twin

38
Fig. 4 Framework for diagnostic reasoner

39
Fig. 5 Overview of hydraulic system

40
Fig. 6 Healthy working results of digital twin

41
Fig. 7 Healthy and faulty working conditions of equipment

42
 K(X,X1) Wo

Bias
α1
K(X,X1)
α2

Input
Vectors αn-1

K(X,X1) αn

K(X,X1)

Fig. 8 Support vector machines

43
 2
2/ w

X2

X1

Fig. 9 Support vector machines and hyperplane example

44
Fig. 10 Mapping for support vector machines

45
 Initialize the weights values to train

Classify assigned weights

Update incorrectly classified weights

again

Exit algorithm and show results if all weak data evolved to strong one;
otherwise, go back to previous process

Fig. 11 AdaBoost algorithm steps

46
Fig. 12 Landing gear actuator jamming failure PCA results

47
Fig. 13 SVM vs. random forest algorithm

48
Fig. 14 ROC curve

PHYSICAL TEST PRODUCT Hardware-in-

DIGITAL TWIN
Loop Simulator
Online Diagnostic Software

Collect Real1 Time

Sensor Data
Failure Scenarios
DIGITAL MODEL
CANBUS (AMESIM)
Healthy Scenarios
Drive Pumps and Valves

Trained ML
Models
Digital Model Sensor Outputs

Fig. 15 The envisioned digital twin architecture (The Physical model is an illustration
from the GrabCAD repository https://grabcad.com/library/hydraulic-test-bench-100-kw-
1/details?folder_id=1290612)

49
Table 1 Assumed healthy working conditions
Aircraft Hydraulic Sensor Lists
Healthy
Sensor Inputs Unit
Working Range
Pump_1_Engine_Shaft 2500–3350 rpm

Pump_2_Engine_Shaft 2500–3350 rpm

Pump_1_Torque 330–450 Nm

Pump_2_Torque 330–450 Nm

Pump_1_Flow_Rate 65–75 gpm

Pump_2_Flow_Rate 65–75 gpm

Motor_Flap_1_Rot 0–1500 rpm

Motor_Flap_2_Rot 0–1500 rpm

Filter_1_Pressure_Differential 0–10 psi

Filter_2_ Pressure_Differential 0–10 psi

Reservoir_1_Volume_Probe 0.700–0.720 L

Reservoir_2_ Volume_Probe 0.700–0.720 L

Relief_Valve_1_Pressure 3000–3300 psi

Relief_Valve_2_Pressure 3000–3300 psi

Relief_Valve_1_Flow_Rate 250–275 L/min

Relief_Valve_2_Flow_Rate 250–275 L/min

Aileron_Pressure_Inlet_Port 3000–3300 psi

Aileron_Displacement 0–0.40 m

Rudder_Pressure_Inlet_Port 3000–3300 psi

Rudder_Displacement 0–0.40 m

Elevator_Pressure_Inlet_Port 3000–3300 psi

Elevator _Displacement 0–0.40 m

LG_Doors_Displacement 0.8 m

LG_Doors_Pressure_Inlet_Port 3000–3300 psi

LG_Actuator_Pressure 3000–3300 psi

LG_Flow_Rate 80–88 L/min

Steering_Angular_Displacement 0–40 °

LG_Break_Flow_Rate 2–2.5 L/min

Flap_Angular_Displacement 0–40 °

Slat_Angular_Displacement 0–40 °

HS_Angular_Displacement 0–40 °

50
 Table 2 Injected faults
Fault Location Fault
Fault Type Fault ID
Time
Hydraulic Pump, Fail Hydraulic Pump 1
P1 0
to Run
Hydraulic Pump, Hydraulic Pump 1
P2 10
Closes Unexpectedly
Hydraulic Pump, Oil Hydraulic Pump 1
P3 15
Leakage
Hydraulic Pump, Hydraulic Pump 1
P4 0
Unsteady Flow
Selector Valve, Oil Selector Valve 1
SV1 15.5
Leakage
Selector Valve, Stuck SV2 Selector Valve 1 1
Selector Valve, Selector Valve 1
SV3 40
Blockage
Actuator, Fail to Primary FLCS
PA1 0
Extraction Actuator
Actuator, Fail to Primary FLCS
PA2 15
Retraction Actuator
Primary FLCS
Actuator, Jamming PA3 15.5
Actuator
Actuator, Fail to Landing Gear
LG1 5
Extraction Actuator
Actuator, Fail to Landing Gear
LG2 0
Retraction Actuator
Landing Gear
Actuator, Jamming LG3 11.75
Actuator
Secondary FLCS
Motor, Fail to Run SA1 0
Motor
Secondary FLCS
Motor, Jamming SA2 39.5
Motor
Filter, Blockage F1 Return Line Filter 0
Relief Valve, Fail to Pressure Line Relief
R1 0
open Valve
Relief Valve, Fail to Pressure Line Relief
R2 0
close Valve
Priotry Valve, Fail to Pressure Line Shut-
S1 0
open off Valve
Priotry Valve, Fail to Pressure Line Shut-
S2 0
open off Valve

51
 Table 3 Prediction accuracy
Prediction Accuracy
Fault Random Linear Quadratic Cubic Gaussian
AdaBoost
Code Forest SVM SVM SVM Kernel
P1 100 100 100 100 99.3
49.3
P2 100 99.8 99.5 92.0 99.5
80
P3 100 100 100 100 100
82.5
P4 75 100 100 100 100 100

SV1 82.8 100 100 100 100 100

SV2 75.5 99.3 83.5 95.8 98.0 96.8

SV3 95.0 99.8 99.8 99.8 79.0 97.8

PA1 75 100 100 100 100 100

PA2 75 100 100 100 100 100

PA3 82.3 99.5 99.5 99.5 99.0 96.3

R1 75 100 99.3 99.8 100 100

R2 75 100 100 100 100 99.5

S1 75 100 100 99 99.8 98.3

S2 75 100 100 99 99.8 98.3

SA1 75 100 93 100 100 100

SA2 94.8 99.8 99.8 99.5 99.8 99.8

F1 75 100 98.5 99 99.8 94.5

LG1 77.5 99.8 96.8 96.3 96.3 96.3

LG2 90.3 100 99 99.8 100 99.3

LG3 82.3 99.0 99.8 99.8 99.0 99.0

52

View publication stats