Intelligent Diagnostics for Aircraft Hydraulic Equipment

Oliver Ritter1, Gerko Wende2, Rocco Gentile3, Francesco Marino4, Antonio Carlo Bertolino5, Andrea Raviola6, Giovanni
Jacazio7

1,2
Lufthansa Technik, Hamburg, 22335, Germany
[email protected]
gerko.wende@ lht.dlh.de
3,4,5,6,7
Politecnico di Torino – Department of Mechanical and Aerospace Engineering, Torino, 10129, Italy
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]

ABSTRACT 3. High-fidelity mathematical modeling of considered test

units, for deeper physics analysis of possible failures;
In aviation industry, unscheduled maintenance costs may
4. Implementation of Machine Learning reasoner, able to
vary in a large range depending on several factors, such as
process experimental and simulated data.
specific aircraft system, operational environment, aircraft
usage and maintenance policy. These costs will become more
noteworthy in the next decade, due to the positive growing of 1. CRITICAL ISSUES IN MAINTENANCE OF AIRCRAFT
EQUIPMENT
worldwide fleet and the introduction of more technologically
advanced aircraft. New implemented technologies will bring Cooper, Smiley, Porter, & Precourt (2017) predict an
new challenges in the Maintenance, Repair and Overhaul optimistic growth of commercial airline fleet. According to
(MRO) companies, both because of the rising number of new their report, in the next decade, the number of units will
technologies and high volume of well-established devices, increase 3.4% net annually, rising the number of commercial
such as Electro-Hydraulic Servo Actuators for primary flight aircraft more than 35000. This growing demand in both cargo
control. Failures in aircraft hydraulic systems deeply and commercial fleets will be fulfilled on one hand
influence the overall failure rate and so the relative introducing about 20000 new-generation aircraft and, on the
maintenance costs. For this reason, overhaul procedures for other hand, replacing older units to about 10000 during this
these components still represents a profitable market share period (Figure 1). These numbers indicate that about 58% of
for all MRO stakeholders. Innovative solutions able to the global aircraft fleet will include new-generation units and
facilitate maintenance operations can lead to large cost will radically influence many different technological issues.
savings. The presence of technologically advanced aircraft implies
This paper proposes new methodologies and features of the also considerable investments in MRO market. Cooper et al.
Intelligent Diagnostic system which is being developed in (2017) report that “commercial airline MRO growth will be
partnership with Lufthansa Technik (LHT). The
implementation of this innovative procedure is built on a set
of failure detection algorithms, based on Machine Learning
techniques. This development requires first to bring together
the results from different parallel research activities:
1. Identification of critical components from historical data
2. Designing and testing automatic and adaptable
procedure for first faults detection;

Oliver Ritter et al. This is an open-access article distributed under the

terms of the Creative Commons Attribution 3.0 United States License,
which permits unrestricted use, distribution, and reproduction in any
medium, provided the original author and source are credited. Figure 1: Global Aircraft Demand

1
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2018

healthy at 3.8% Compound Annual Growth Rate (CAGR) is a task that often requires large effort and time
over the 10-year period, growing from the current demand of (Byington, Watson, & Edwards, 2004). Furthermore,
$ 75.6 billion to just over $ 109 billion by 2027”. This rising many of these LRUs are installed in parallel for
investment expectation takes into account four main redundancy (for example the Flight Primary Control
segments of MRO market: Airframe, Engines, Components Systems), so usually for safety reason both the devices
and Line Maintenance. By the way, in order to improve the are removed though only one is failed.
maintenance capability of these four sections, it is necessary The goal of this paper is to describe the actual development
to shift the paradigm at the base of the maintenance approach: of fault diagnosis system for a given class of aircraft
from a reactive to a condition-based strategy. Both OEM equipment, to obtain a high-quality, faster and cost-effective
players and MRO providers are dealing with new challenges maintenance procedure. The presented case study is focused
related to new-generation fleet: on the flight control actuators. These units can still be
1. More sophisticated avionics components will require considered a key point for cost saving of maintenance
advanced health monitoring systems, which will be able procedures for three main reasons:
to identify faults of the units; 1. The components of the hydraulic system give a
2. The number of collected health-monitoring parameters significant contribution to the overall failure rate and
will considerably rise, and all the MRO stakeholders hence to the total associated maintenance cost;
have to find out what is the best way to take advantage 2. Large passenger aircraft in service are equipped with two
from this amount of valuable data; or three hydraulic systems for safety reasons;
3. Line maintenance providers need to invest new capitals 3. Hydraulic systems are predestinated to be equipped with
on both modern testing equipment and trainings about more sensors or are already equipped with a huge variety
new health monitoring and fault isolation systems. of sensors that are very good usable for prognostic
As reported by Groenenboom (2017), health-monitoring algorithms and methods;
systems for aircraft were already regulated during the 1970s, 4. Hydraulic technology is projected to be the most
when MSG-2 introduced the “Condition Monitoring” common solution in the next future for flight controls
maintenance concept. During the last years, to improve their and landing gear of the new versions of these large
competitiveness in this field, also large MRO providers like passenger aircraft (Mornacchi et al., 2015). Moreover,
Lufthansa Technik (LHT), Airfrance Industries and KLM the benefits gained from more efficient maintenance
Engineering & Maintenance have introduced their own health procedures will thus not be limited to legacy aircraft but
management and prognostic services. Today both OEMs and could be applied to new platforms.
MRO providers aim to introduce health-monitoring
innovative solutions able to facilitate maintenance operations 2. PROJECT OBJECTIVES, METHODOLOGY AND USE CASE
that can lead to large cost savings. Mornacchi, Vachtsevanos
and Jacazio (2015) indicated two areas that can benefit from The contents of this article summarize one year of
an efficient health-monitoring system: collaboration between Polytechnic of Turin and LHT, which
is aimed to investigate and design new suitable health-
1. Improvement of aircraft operation reliability and monitoring procedures for primary flight control actuators.
dispatch-ability by avoiding on ground immobilization This task constitutes a key cost saving factor for two main
time and flight delays or re-routing or cancellations; reasons:
2. Reduction of direct maintenance costs by rescheduling
maintenance operations, improving failures 1. Considerable numbers of Shop Load Event (SLE);
troubleshooting and performing maintenance tasks of 2. OEM-built test procedures for fault detection and
anticipated failures; isolation are not cost-effective in terms of maintenance
time.
Two main critical issues that negatively influence both cost
areas are: On this basis, the definition of a new automatic procedure
could represent a remarkable advantage. An automatic
1. No Fault Found (NFF) cases: high number of equipment procedure combined with an algorithm capable of detecting
are removed from the aircraft because of a wrong fault failures can really improve the overall maintenance quality.
indication, even though they result fully serviceable. The Marino (2017) reports that it is worthy to choose components
NFF rate is on average about 1/6 of the total number of that are widely installed on today and next-generation
signaled faults, nevertheless some equipment can reach aircraft, to obtain a significant impact on future savings. The
a NFF rate above 25%; choice of the prototyping unit is based on both technical and
2. Difficulty in identifying the actual faulty component economic criteria:
which are classified as Line Replaceable Units (LRU).
1. Number of subcomponent of the unit and overall
These units are complex assembly consisting of several
complexity of the assembly;
sub-assemblies, and the identification of the failed part

2
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2018

2. Number of records and unique SLE; 2. Symmetrical actuator: the position of the ram is
3. Maintenance Total Costs (Overhauling Costs); determined by the difference of pressure between its two
control chambers;
4. Time Since Installation (TSI);
3. Mode Switching Valve (MSV): central valve controlled
During this cooperation with LHT, the first important
by two normally opened Electro-valves (EVs), which are
research goal is the development of a new condition detection
placed in series, constituting a logic NOR function. This
algorithm.
component is used to deactivate the unit and to set its
Several different approaches for fault detection and function in damping mode. When both the two EVs are
identification have been already adopted by Karpenko, not activated, the spool of the MSV is directly connected
Sepehri and Scuse (2003) who have identified three possible with the supply line: in this way, the pressure force
families of algorithms: the first one is a model-based pushes the spool against its contrasting spring and the oil
approach (Isermann, 2005), while the second and the third flows to the cylinder. If at least one of the EVs is turned
use Machine Learning techniques in two different ways. on, the MSV spool is connected to the return line, and
Bernieri, D’Apuzzo, Sansone and Savastano (1994) have the EHSV is then bypassed (bypass-mode). The cylinder
trained a neural network to reproduce the dynamic behaviour chambers are connected through a calibrated orifice
of the system, while different other approaches have used which is responsible of damping action.
Machine Learning algorithms as pattern classifiers in order to 4. Accumulator: this component supplies flow to avoid
recognise failure modes directly from the data of the dynamic pressure to drop below the cavitation value and, in case
system (Byington et al., 2004; Crowther, Edge, Burrows, of hydraulic supply interruption, it allows the actuator
Atkinson, & Woollons, 1998; Garimella & Yao, 2004; Le, rod to come back in its neutral position.
Watton, & Pham, 1998; Mornacchi et al., 2015).
The control part of the scheme includes a position sensor for
To make the last approach efficient, it is necessary to collect closed loop position and a Mechanical Recentering Device
results from different engineering areas: which mechanically connects the linear ram with the spool of
the EHSV. This security component moves the main ram in
1. Techniques of data manipulation, state detection and
its null position in case of electrical failure of the current
identification of degradation; supply in the first stage of the EHSV. All these main sub-
2. Measurement campaigns of in-service components, both assemblies contain several other components that can be
in healthy and degraded conditions; substituted in case of failures. The information about the
3. High fidelity mathematical modeling of physical systems replaced parts are recorded during each SLE. These records
and mathematical description of fault propagation constitute an historical database of all the replaced
physics; components of each adjusted sub-assembly which represents
4. Data Mining and Data Fusion of historical, simulated a valuable tool for the analysis of the most recurrent
and experimental results; exchanged parts during repairing operations.
5. Implementation of Machine Learning Algorithms for
processing data with the knowledge of physics of failure.

3. ANALYSIS OF MOST RECURRENT FAILURES

Once the reference unit has been identified, a further
investigation is needed to recognize which sub-assemblies of
are the most critical. This kind of analysis is also crucial for
the implementation of an excitation signal that can be used to
detect the most recurrent failures. The scheme in Figure 2
displays all the sub-assemblies of the reference Electro
Hydraulic Servo-Actuator (EHSA). In the hydraulic part of
the scheme, it is possible to distinguish:
1. Electro-Hydraulic-Servo-Valve (EHSV): jet-deflector
kind where the first stage receives the current command
by the external controller as input and it moves the
deflector according to the rotation of the torque motor.
This movement creates a pressure drop at the end of the
second stage spool, which drives the correct flow to the
cylinder chambers;
Figure 2: EHSA reference configuration

3
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2018

The results of this investigation have shown that some sub- be classified as a combination of Model Based and Data-
assemblies like the electro-valves or the mechanical feedback Driven approaches, similarly as the Hybrid Diagnosis
are less critical than others, like the EHSV, that is the most Approach for Electro-mechanic Actuators (Narasimhan,
exchanged component. This classification can be considered Roychoudhury, Balaban, & Saxena, 2010).
a first guideline for the definition of the excitation signal, first
step of the Intelligent Diagnostic procedure. 5. MATHEMATICAL MODELING AND SIMULATIONS
The most important advantage of a model-based approach is
4. IMPLEMENTATION OF INTELLIGENT DIAGNOSTIC
its flexibility: in a highly-detailed physical model, it is
As first step of the implementation of an Intelligent possible to set a wide range of possible degradations, with
Diagnostic, the definition of an automatic failure-recognition any kind of severity. The more complex and detailed the
procedure has an important impact in terms of cost saving of model is, the more number of possible case studies increases.
standard maintenance procedure. The base of this Using the same approach of Bertolino, Jacazio, Mauro and
development is a specific command sequence that can Sorli (2017), the final goal of high-fidelity model is to create
identify in a short time the most recurrent failures in the a virtual bench that can replicate the nominal behavior of the
critical subcomponents. In this way, it would be possible to physical unit and to “evaluate its behavior correlating it with
substitute the standard maintenance procedure with a fast and small variations” of some parameters, corresponding to
efficient self-designed Entry Test. Indeed, standard test physical characteristic of the unit. This strategy allows to
procedures contain specific commands to assess the have a greater flexibility on studying the effects of several
performance of the subassemblies in the EHSA, but these flaws injection combinations, due to the absence of hardware.
procedures can last hours, and they must be performed at least Although it speeds up the failure physics investigation, it
two times (one for the Entry and one for the Certification represents an issue as well: indeed, the model has first to be
Test). This approach is not considered time and cost-effective validated and tuned on real measurements.
anymore. The main advantages of an automatic command
sequence can be synthetized: 5.1. Model Structure
1. Automation: Entry Test can run automatically without The high fidelity dynamic non-linear model has been
any human-interaction during the measurements in a developed in Matlab - Simulink® environment and it is the
short period of time, while all the needed Health Features evolution of the model developed by Mornacchi (2016). Its
(HF) can be collected; structure can be divided in three conceptual layers (Figure 3):
2. Standardization: since there are no interactions with the 1. First Layer contains all the necessary data to introduce
operator during the duration of the test, the measured different degradations in terms of severity and starting
time responses are always the same. This means that is time;
possible to compare the responses and increase the
2. Second Layer includes three sub-files with not only
knowledge about similar cases;
physical characteristics of EHSA itself, but also the ones
3. Continuous measurement: it is possible to collect data of its seals and used fluid;
about physical parameters during the entire command
3. Third Layer contains all the files concerning the real
sequence;
physical modeling of the unit. Inside the cardinal EHSA
4. Scalability of the test procedure: since the main blocks in Figure 3 all the physical components in Figure
components of several primary flight control actuators 2 are modelled with a white-box approach.
are the same, the designed automatic procedure for the
reference unit can be adapted to other kind of actuators. The entire model is physically meaningful and each
component motion is determined by the resolution of
During the design of this automatic entry test, both the d’Alambert differential dynamic equations, as in Bertolino et
classification the most critical sub-assemblies and the al., 2017. The behavior of these inner EHSA parts are
historical data of the most recurrent failed certification tests influenced by all the other blocks in Layer 3. Command block
have been considered. With this automatic procedure, it will contains the same excitation-signal that has been adopted for
be possible to collect a large number of data. This database the experimental campaign, and it interacts with Oil
with all the possible historical collected measurements of Properties and Test Bench blocks. In particular, the latter is
degraded and healthy units will considerably enlarge the the only component that has not been modelled in a physical
knowledge of all the single degradations in the EHSA, way, in order to not considerably increase the overall
constituting the base for the development of a Machine complexity. A Discrete- time Autoregressive Exogenous
Learning algorithm for fault detection. By the way, this model (ARX) has been adopted because it results particularly
database would require long-lasting measurement and data suitable for modelling “dynamic process driven by and input
collection campaigns. This issue can be overcome increasing in presence of uncertainties” (Diversi, Guidorzi, & Soverini,
the number of samples with simulated data through a high
fidelity mathematical model. This development strategy can

4
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2018

Jacazio and Vachtsevanos (2017) and Nordin, Galic and

Gutman (1997). The main ram receives as input the control
flows from the EHSV, through the MSV (𝑄1 and 𝑄2 ) and it
returns as output both its position and the external sleeve one
( 𝑥𝑅 and 𝑥𝐶 ). The difference between them is the relative
displacement of the ram and it is used to close the position
control loop. The pressures in both the chambers 𝑝1 and 𝑝2
are calculated as function of the volumetric flow rates, the
fluid properties and the geometric characteristics. In order to
simulate also the seals failures, 𝑄1 and 𝑄2 are affected by two
different types of leakages: internal between the chambers
(𝑄𝑙𝑖 ) and external between the chambers and outside (𝑄𝑙𝑒𝑖 ).
Continuity equations in both the chambers can be written as:
𝜕𝑝1 𝑄1 − 𝑄𝑙𝑒1 − 𝑄𝑙𝑖 − ( 𝑥𝑅̇ − 𝑥𝐶̇ )𝐴
=𝛽 (1)
𝜕𝑡 𝑉0 + 𝑥𝑅 𝐴
𝜕𝑝2 𝑄2 − 𝑄𝑙𝑒2 + 𝑄𝑙𝑖 + ( 𝑥𝑅̇ − 𝑥𝐶̇ )𝐴
Figure 3: Simulink Model Conceptual Structure =𝛽 (2)
𝜕𝑡 𝑉0 − 𝑥𝑅 𝐴
2010). The connection between EHSA and external where 𝛽 is the oil bulk module, 𝑉0 is the initial volume of the
environment is controlled by the blocks Environmental Data
chambers when the ram is centered and 𝐴 represents the
and External Load, which is deactivated at the moment
surface of the head of the piston. The results of Eq. (1) and
because any load test has been carried out so far. When a Eq. (2) are then used to determine the position of ram and
simulation is completed, the Data Storage block collects all sleeve according to the following dynamic equations:
the useful data for post-processing and further analysis. In the
next paragraphs, the models of the EHSA sub-assemblies will 𝑚𝑅 𝑥̈ 𝑅 = (𝑝1 − 𝑝2 )𝐴 − 𝛾(𝑥̇ 𝑅 − 𝑥̇ 𝐶 ) − 𝐹𝑓𝑟𝑖𝑐 − 𝐶𝑒𝑥𝑡 𝑥̇ 𝑅 (3)
be briefly described, and at the end a comparison between 𝑚𝐶 𝑥̈ 𝐶 = (𝑝2 − 𝑝1 )𝐴 + 𝛾(𝑥̇ 𝑅 − 𝑥̇ 𝐶 ) + 𝐹𝑓𝑟𝑖𝑐 − 𝐶𝑠𝑎 𝑥̇ 𝐶 − 𝐾𝑠𝑎 𝑥𝐶 (4)
experimental and simulation results will be shown.
where 𝑚𝑅 and 𝑚𝐶 are the masses of the ram and the sleeve
5.2. EHSA Main Ram respectively, γ is the viscous friction coefficient and 𝐹𝑓𝑟𝑖𝑐
The hydraulic main ram is a 2DOF (Figure 4) model where: stands for the Coulomb friction force. In particular, the model
of friction allows both the static and dynamic conditions and
1. 𝑥𝑅 , 𝑥̇ R and 𝑥̈ 𝑅 are position, velocity and acceleration of the switching between them to be considered. Static friction
the ram; value is evaluated according to the pressure in both the
2. 𝑥𝐶 , 𝑥̇ C and 𝑥̈ 𝐶 are position, velocity and acceleration of chambers and the geometrical and physical data of the seals
the external sleeve structure; (Martini, 1984). The dynamic value instead and the condition
of commutation are modeled as described in Bertolino
The external damper 𝐶𝑒𝑥𝑡 corresponds to the external eye-end
(2016) and Bertolino et al. (2017).
of the ram, and it results particularly useful in case of external
load 𝐹𝑒𝑥𝑡 . As mentioned in the previous paragraph, all the
5.3. Servo-valve EHSV
simulations have been performed without any external acting
force. 𝐶𝑠𝑎 and 𝐾𝑠𝑎 represent the damping coefficient and the The EHSV is the most critical and complex component of
stiffness of the external connection between the sleeve and both the model and real unit. It represents the interface
the fixed frame. Both in these points and in the ram end, it is component between the control logic and the hydraulic part
possible to include the effect of backlash, following the and its behavior influences the entire dynamic response of the
approach of Bertolino (2016); Bertolino et al. (2017); Martin, actuator. The real installed servo-valve in the unit is a jet-
deflector, whereas the simulated one is flapper-nozzles type.
This difference is only in the first stage of the EHSV and it
does not influence the dynamic response or the reliability of
the global physical model of the actuator. The inner structure
of both the types can be divided in:
1. First stage: it transforms the electrical input into a low
power hydraulic signal. It includes the torque motor and
flapper dynamics. The torque motor block receives as
Figure 4: Actuator Model Scheme input the current 𝑖𝑐𝑜𝑚 coming from the controller,

5
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2018

whereas the flapper dynamic gives as output the position

of the flapper 𝑥𝑓 ;
2. Second Stage: it transforms the low power hydraulic
command signal into a high power hydraulic flow. It
contains the blocks of the hydraulic amplifier and spool
dynamics. According to the flapper position, the
hydraulic amplifier block evaluates the difference of
pressure between the command chambers of the sleeve Figure 6: EHSV Spool Free Body Diagram
𝛥𝑝𝑠𝑝𝑜𝑜𝑙 , that controls the final position of the spool 𝑥𝑠 .
through the actuator chambers. The equation of the free body
The model of the torque motor is based on the magnetic diagram in Figure 6 takes into account the force of the
circuit analyzed by Urata (2007), where the sources of feedback spring 𝐹𝑓𝑏 , which connects the flapper with the
magnetic forces are the permanent magnets 𝑉𝑝 and solenoid spool, the Recentering Device equivalent force 𝐹𝑟𝑑 (section
magnetic force 𝑛𝑖𝑐𝑜𝑚 (Figure 5). Each reluctance in the 5.5), Coulomb and viscous friction forces and structural
circuit is calculated taking into account the air-gap thickness, damping. Since there are not seals, the dynamic friction is a
and eventual misalignment of the armature. The generated constant value and does not follow the laws proposed by
torque is expressed considering the magnetic fluxes through Martini (1984). Once 𝑥𝑠 is determined, it is necessary to
the air-gaps: evaluate the flow through each port of the EHSV, considering
4 both supply and return contributes. This issue is modelled
𝐿𝑎 𝐴𝑔 with an equivalent Weathstone circuit (Merritt, 1967), where
𝑇= ∑ 𝐵𝑖 (5)
4µ𝑎 each port is represented as a variable resistance with two
𝑖=1
components in series: a laminar term 𝑅𝐶 and a turbulent one
where 𝐵𝑖 is the flux density between the four air gaps between 𝑅𝐴 (Mornacchi et al., 2015). Once these values are evaluated
magnet and solenoids, 𝐿𝑎 is the distance between the left and for all the control ports, the model calculates the flow through
right pole, 𝐴𝑔 is the cross-sectional area of the air-gap and each control port through Eq. (8).
µ𝑎 is magnetic permeability of the air. At the end of the first
stage of the EHSV; 𝑥𝑓 is calculated combining the Eq. (5) −𝑅𝐶 + √𝑅𝐶2 − 4𝑅𝐴 |∆𝑝𝑠𝑝𝑜𝑜𝑙 |
𝑄𝑆𝑉 = ∗ 𝑠𝑔𝑛(∆𝑝𝑠𝑝𝑜𝑜𝑙 )
(8)
with the dynamic rotation equation of the torque motor 2𝑅𝐴
anchor. The position of the flapper determines the difference
of pressure between the chambers of the sleeve, regulating
the flow through the nozzle channels. These pressures are 5.4. Mode Switching Valve MSV
function of both 𝑥𝑓 and spool velocity: Physically the dynamic of the spool of the MSV is similar to
𝐴𝑠𝑝𝑜𝑜𝑙 the second stage of the EHSV, without any feedback force
𝑝𝐴 = 𝐺𝑃 (𝑥𝑓 − 𝑥̇ ) (6) 𝐹𝑓𝑏 but with an elastic force from its contrasting spring. It is
𝐺𝑄 𝑠
connected not only to the EHSV, but also with the
𝐴𝑠𝑝𝑜𝑜𝑙 accumulator and the orifice between the chambers in case of
𝑝𝐵 = −𝐺𝑃 (𝑥𝑓 − 𝑥̇ ) (7) damping mode. It can reach just two discrete positions
𝐺𝑄 𝑠
(opened or closed) and it can be commanded by just one of
where GP and GQ are respectively pressure and flow gain. The the two installed EVs. Once the position of the MSV is
imbalanced pressure 𝛥𝑝𝑠𝑝𝑜𝑜𝑙 = 𝑝𝐴 − 𝑝𝐵 on the two-side determined, the flows through its ports can be determined
spool creates a displacement that allows the flows to run considering both laminar and turbulent flow resistances, as
for the EHSV.

5.5. Recentering Device RD

Particularly relevant component in both normal and
unexpected working conditions of the unit. It consists of a
complex mechanical device which moves the main ram to its
null position in case of failure of the current supply to the
EHSV. It is both connected to the main ram and to the jet
deflector system of the EHSV (equivalent to the flapper of
the flapper-nozzles valve), as shown in the diagram of Figure

Figure 5: Magnetic Torque Motor Circuit (Urata, 2007)

6
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2018

Figure 7: Recentering Device Block Scheme

7. One side of the main ram is linked to the RD rotating shaft

hinge on which is mounted the RVDT transducer. The
rotation of this component moves a small eccentric cam that
directly controls the motion of the jet deflector system inside
the EHSV, thanks to two contrasting springs. This
mechanical effect is compensated directly by the external
controller, which provides to the EHSV an additional offset
current both in extension and retraction. By the way, this
offset is not constant for the entire stroke, but the RD works
as nonlinear force that depends on the position of the main
ram. Since its geometry is complex, the effect of this
Figure 8: Pressure drop comparison between simulated and
component is considered inside the model through its zero
experimental results
order characteristic curve. Its offset effect has been
considered with an equivalent force 𝐹𝑟𝑑 on the EHSV spool plots in Figure 8 displays a comparison between measured
(Figure 7). and simulated supply pressure commanding the unit with a
Pseudorandom Binary Sequence (PRBS) in closed loop, with
5.6. Test Bench TB reference to the input extracted flow.
As already mentioned in 5.1, the dynamic model of the TB is
built with an ARX Model, because most of the physical 5.7. Model Validation
parameters are unknown and in order to reduce the The model validation is performed using data acquired during
complexity of the model. Model tuning is performed using the experimental campaign at Lufthansa Technik facilities.
two different sets of measured data: supply pressure and The plots in Figure 9-left) displays a comparison between
outlet flow. Part of these two time-series (70% of the time- measured and simulated position and EHSV current during
span of the command) is used for the model identification, the PRBS sequence. A closer look in the 30th second in Figure
while the remaining part for its validation. The equation at 9-right) shows that the outputs of the model are very close to
the base of an ARX model is: the actual behavior, concerning both the current of EHSV and
𝐴(𝑞)𝑦(𝑡) = 𝐵(𝑞)𝑢(𝑡 − 𝑛𝑘 ) + 𝑒(𝑡) (9) the main ram position. Both the plots show that the model is
able to represent in an accurate way the system behavior and
where 𝑞(𝑡) is a delay operator, 𝑒(𝑡) is a white noise, 𝑛𝑘 that it can be used to study the performance degradation when
represents delay from input to output in terms of number of the parameter values move away from the baseline condition.
samples and 𝐴(𝑞) and 𝐵(𝑞) are two linear functions defined
as follow: 6. DATA ANALYSIS AND EARLY STAGE DEVELOPMENT OF
MACHINE LEARNING ALGORITHM
𝐴(𝑞) = 1 + 𝑎1 𝑞−1 + ⋯ +𝑎𝑛𝑎 𝑞 −𝑛𝑎
(10) As previously anticipated, a crucial point within the
Intelligent Diagnostic frame is the development of a Machine
𝐵(𝑞) = 𝑏1 + 𝑏2 𝑞−1 + ⋯ +𝑏𝑛𝑏 𝑞−𝑛𝑏+1 Learning Algorithm for fault classification. The idea is to
(11) develop an algorithm which provides a unique and accurate
where: indication of where the failure is located in the EHSA, by
using the features extracted from the raw signals acquired
1. 𝑦(𝑡): model output at instant t; during the initial automatic Entry Test. The development of
2. 𝑛𝑎 : number of poles, so the number of past output terms this specific part of the Intelligent Diagnostic presents some
to predict the current output; critical issues:
3. 𝑛𝑏 : number of zeros plus one, so the number of past input 1. Number of entries reported inside historical database;
terms to predict the current output; 2. Some reports could not present all the needed
In this case, the best fitting is obtained fixing 𝑛𝑎 = 𝑛𝑏 = 4 information for this investigation;
and 𝑛𝑘 = 0, with a reached accuracy of more than 95%. The

7
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2018

Figure 9: Left) Model and Experimental Results – PRBS Test; Right) Particular of previous graph on the left side

3. Difficulties in disambiguation of some physical 1. Excessive roundness of control ports orifice due to local
degradations that affect the EHSA response in similar wears, which can have important impact on the control
ways. flows;
As already mentioned in paragraph 3, the analysis of 2. High hysteresis current in the first-stage torque motor;
historical data of repairing reports indicates the EHSV as 3. Increase of radial clearance between spool and sleeve;
most recurrent exchanged sub-assembly. Furthermore, this
4. Stiffness variation of the feedback spring between first
part can be considered as the most complex and crucial
and second stage;
element for the correct behavior of the EHSA. From data-
mining perspective, the large number of entries in the 5. Excessive backlash between feedback spring and spool;
database (70% of the recorded SLEs) guaranties a good These simulated degradations can be directly detected
starting database for fault identification. Despite the large extracting the health-features directly from the EHSA
number of events, the information related to the health status response of six specific portions of the complete excitation
of the EHSV are just two since the EHSV can be entirely signal. In particular, these parts of the signal include all the
replaced or not. As consequence, an initial binary standard certification tests that are strictly influenced by the
classification can be easily achieved as first step to prove the EHSV behavior. These six signals have been used as input of
feasibility of the ML development. According to what the Simulink Model and, from the analysis of its responses,
reported above, the historical information regarding the eighteen Health Features (HF) have been extracted as output.
EHSV were classified in two classes, marked with a binary If either one of them results out of standard test limits, the
label: zero if the sub-assembly needs to be replaced, or one troubleshooting of the related test would be to substitute the
otherwise. failed EHSV, and the entire entry-simulated test would be not
passed. An example of the structure of the built database is
6.1. Database Generation shown in Table 1: the first eighteen columns represent the
extracted health features from the signal. The last column
In order to increase the number of entries, the historical
“Label” reveals the final result of the entry test: one if all the
database has been combined with simulated results, created
health features are within their limits, and zero if at least one
including EHSV degradations. The first step for building this
test is not passed. Finally the historical and simulated results
database is to evaluate how many HF are directly related to
were merged together in one database. The database
the EHSV behavior. From this result, the historical database
structure, as reported in Table 1 constitutes the training
of SLE has been “cleaned” from useless data:
dataset used to implement the supervised ML algorithm for
1. Data loss due to missing information; binary classification between those unserviceable sub-
2. All the features not directly linked to the EHSV. components and the serviceable ones.
In order to increase the database size, several hundred
simulated responses have been performed with different
combinations of EHSV degradation. Five different possible
physical degradations were considered with different levels
of severity and considered in pair:

8
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2018

HF n.1 HF n.2 … HF n.18 Label Both the hidden layers employ a Rectifier Linear function
(ReLu) as activation, whereas the last output layer adopts a
Actuator 1 … … … … 1 sigmoid function because of its two possible binary states.
Actuator 2 … … … … 0 During the training phase of the NN, to improve the way it
… … … … …
learns, it is necessary to choose wisely:

Actuator n-th … … … … 1
 The Training Set of Data from the original database,
where two balanced classes of passed and failed tests are
Table 1: Database Structure of the Simulated Extracted represented equally;
Health Features  Its Loss Function ℒ to better cope with NN learning
slowdown;
6.2. Implementation of Neural Network for Failure
Classification  A Regularization Method, “which make our network
better at generalizing beyond the training data” (Nielsen,
The implemented ML algorithm for this binary classification 2017);
is a supervised Neural Network (NN), which has to be trained
 The Network hyper-parameters, like the learning rate or
in order to detect data representing failure modes of the
the optimization parameters.
dynamic system. Chen and Burrell (2001) have already
highlighted how NN can be particularly suitable for “pattern The Training Set has been extracted from the HF database in
matching and classification problem”. In this first stage of Table 1. However, these features are of different type and
design, the Multilayer Perceptron (MLP) structure has been scale, so before feeding them inside the input layer, they have
developed in order to perform a logistic classification been regularized with the standard scaler. The standard scaler
between Failed and Passed units, as reported in the last regularization assumes that the data are normally distributed
columns of Table 1. This initial step of the ML algorithm has and scale the result in a way that the distribution would be
been fundamental for two main reasons: centered on zero, with a standard deviation of one:

1. It has demonstrated ability to recognize anomalous 𝑥𝑖 − 𝑚𝑒𝑎𝑛(𝑥)

behaviors by learning from the provided data during the 𝑥̃𝑖 = (12)
𝑠𝑡𝑑𝑒𝑣(𝑥)
training stage;
where 𝑥 represents a feature and 𝑖 indexes the 𝑖 𝑡ℎ training
2. It constitutes the first step for a more specific
sample. Once the dataset is normalized it is possible to split
identification of the failure component inside the EHSV,
randomly the entire database into: training set and test set,
while nowadays the entire servo-valve is completely
with a ratio of 4:1. This sub-set of data for training has to be
removed without any further investigation.
as much balanced as possible, in order to prevent the
The architecture of the proposed fault diagnostic algorithm is algorithm from overfitting the dataset, which would then
similar to the one in Figure 10: in this case, the NN presents affect in a negative way the classification performance. The
eighteen input nodes, corresponding to the extracted HF, and loss function used (ℒ) is a particular case of the multinomial
two full dense hidden layers with forty nodes each. The final cross-entropy loss where each example belongs to a single
output layer contains just one node, which can assume binary class, particularly suitable for binary classification:
values as reported in the Label column of Table 1.

Figure 10: Neural Network Structure

9
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2018

subcomponents of the actuator. To improve the accuracy is

then possible both to increase the number of samples with
simulated data and improve the quality of the features used to
perform the classification. Furthermore, as already
mentioned in the beginning of the paragraph, this algorithm
constitutes the first base for a more complex NN architecture
for the localization of the failure inside the EHSV. These
further steps will be achieved in the next future to improve
the accuracy level and ensure a better diagnostic level.

7. CONCLUSION
This paper summarizes the current development for a new
and Intelligent Diagnostic methodology for primary flight
control systems. The combination between an automatic
Figure 11: Normalized Confusion Matrix failure-recognition procedures, simulated results with High-
Fidelity Mathematical Model and Machine Learning
𝑛
1 algorithm for failure localization are already providing
ℒ=− ∑[𝑦𝑖 𝑙𝑜𝑔(𝑝𝑖 ) + (1 − 𝑦𝑖 ) 𝑙𝑜𝑔(1 − 𝑝𝑖 )] (13)
𝑛 excellent results in terms of time and cost savings.
𝑖=1

where 𝑖 indexes the samples, 𝑦 is the sample label and 𝑝 is Further improvements of this research will involve all three
the prediction for a sample. Once the structure of the main topics of the research:
algorithm is defined it is required to tune all the hyper- 1. Extension of the automatic procedure to other different
parameters of the NN. Large NN requires a considerable primary flight control systems, following the same
number of free parameters, which need to be accurately methodology used for the reference EHSA;
selected and evaluated during the training to maximize the 2. Starting from the current results with failures in the
metric score considered but also to make the NN generalize EHSV, the simulation campaign will be widened,
and properly working with a new set of inputs. Dropout is a including more induced local faults in other sub-
regularization technique for addressing this overfitting assemblies. An interesting study will include the analysis
problem. The key idea is to randomly drop neurons (along of the unit behavior combining faults in different sub-
with their connections) from the NN during training to assemblies;
prevent units from co-adapting too much. In this way, every
time that a new training input is set as input, the NN presents 3. Accurate management of all the collected data from real
a different shape, but at the end “all these architectures shares and simulated results, in order ensure a reliable and high
weights” (Krizhevsky, Sutskever, & Hinton, 2012). At test diagnostic level from the Machine Learning algorithm.
time, the drop of the unit is not performed and predictions of The final goal will be to provide a tool which is able not
all these sub-structures compose a NN that has smaller only to classify if the tested unit is serviceable or not, but
weights. In our specific case, the dropout parameter has been also to localize the defected component inside the
settled to 0.6 (likelihood of keeping a specific neuron active). affected sub-assembly of the EHSA.
The training of the NN has been performed with the
backpropagation approach with the Adaptive Moment ACKNOWLEDGEMENT
Estimation optimization algorithm (Kingma & Ba, 2015). This work was supported by Lufthansa Technick AG. We
This optimization method computes the adaptive learning appreciate their contribution to the development and
rates for each parameter and stores the exponential decaying validation of the mathematical model by providing to the
average of the past squared gradients. Finally, the metric used research team important data, actuators and technology
to score the algorithm was the accuracy. The Accuracy applied to test bench.
achieved by using the algorithm is about 80% of correct
classification (Figure 11): NOMENCLATURE
𝐴 Surface of the Head of the Piston
∑ 𝑇𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒 + ∑ 𝑇𝑟𝑢𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒 𝐴𝑔 Cross-Sectional Area Air-Gap
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (14) 𝐴𝑠𝑝𝑜𝑜𝑙
∑ 𝑇𝑜𝑡𝑎𝑙 𝑝𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛 Spool Section Area
𝐵𝑖 Flux Density in i-th Air Gap
𝐶𝑒𝑥𝑡1 Damping Coefficient of Ball-eye of the Actuator
This level of accuracy proves the feasibility of this method 𝐶𝑠𝑎 Damping Coefficient of External Sleeve Structure
that can be extended starting from the EHSV to many other CAGR Compound Annual Growth Rate

10
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2018

EHSA Electro- Hydraulic Servo Actuator 3

EHSV Electro- Hydraulic Servo Valve
Bertolino, A. C., Jacazio, G., Mauro, S., & Sorli, M. (2017).
𝐹𝑓𝑏 Feedback-spring Force
High Fidelity Model of a Ball Screw Drive for a Flight
𝐹𝑓𝑟𝑖𝑐 Coulomb Friction Force Control Servoactuator. In ASME 2017 International
𝐹𝑟𝑑 Recentering Device Force on the EHSV spool Mechanical Engineering Congress and Exposition (pp.
𝐺𝑃 EHSV Pressure Gain 1–10). Tampa, Florida, USA.
𝐺𝑄 EHSV Flow Gain
HF Health – Features Byington, C. S., Watson, M., & Edwards, D. (2004). Data-
𝐾𝑠𝑎 Stiffness of External Sleeve Structure driven neural network methodology to remaining life
𝐿𝑎 Distance between Magnetic Poles predictions for aircraft actuator components. In IEEE
LHT Lufthansa Technik Aerospace Conference Proceedings (Vol. 6, pp. 3581–
LRU Line Replaceable Unit 3589). https://doi.org/10.1109/AERO.2004.1368175
MRO Maintenance Repair and Overhaul Chen, D., & Burrell, P. (2001). Case-Based Reasoning
MSV Mode Switching Valve System and Artificial Neural Networks: A Review.
NARX Nonlinear Autoregressive Exogenous Model Neural Computing & Applications, 10(3), 264–276.
NFF No Fault Founds https://doi.org/10.1007/PL00009897
NN Neural Network
OEM Original Equipment Manufacturer Cooper, T., Smiley, J., Porter, C., & Precourt, C. (2017).
𝑄1 , 𝑄2 Control Flows in the Actuator Chambers Global Fleet & MRO Market Forecast Summary.
𝑄𝑙𝑖 Internal Leakage Flow between Actuator Chambers Retrieved from http://www.oliverwyman.com/our-
𝑄𝑙𝑒 External Leakage Flow of i-th Chamber expertise/insights/2016/apr/mro-survey-2016.html
𝑅𝐶 , 𝑅𝐴 Laminar and Turbulent Resistance in EHSV ports Crowther, W. J., Edge, K. A., Burrows, C. R., Atkinson, R.
RD Recentering Device M., & Woollons, D. J. (1998). Fault diagnosis of a
SLE Shop Load Event hydraulic actuator circuit using neural networks—an
𝑇 Torque applied by the EHSV Torque Motor output vector space classification approach.
TSI Time Since Installation Proceedings of the Institution of Mechanical
𝑉0 Initial Volume of both Actuator Chambers Engineers, Part I: Journal of Systems and Control
𝑖𝑐𝑜𝑚 Command Current Engineering, 212(1), 57–68.
𝑚𝑅 Mass of the Actuator Ram https://doi.org/10.1243/0959651981539299
𝑚𝐶 Mass of the Actuator Sleeve
Diversi, R., Guidorzi, R., & Soverini, U. (2010).
𝑝1 , 𝑝2 Pressures in the Actuator Chambers
Identification of ARX and ARARX models in the
𝑝𝐴 , 𝑝𝐵 Pressures in the EHSV Spool Chambers
presence of input and output noises. European Journal
𝑥𝐶 Displacement of the Actuator Sleeve
of Control, 16(3), 256–257.
𝑥𝑓 Flapper Position
https://doi.org/10.1016/S0947-3580(10)70649-X
𝑥𝑅 Position of the Actuator Ram
𝑥𝑠 Spool Position Garimella, P., & Yao, B. (2004). Fault detection of an electro-
𝛥𝑝𝑠𝑝𝑜𝑜𝑙 Pressure Drop between EHSV Sleeve Chambers hydraulic cylinder using adaptive robust observers. In
β Oil Bulk Module 2004 ASME International Mechanical Engineering
γ Viscous Friction Coefficient Congress and Expo.
µa Air Magnetic Permeability Groenenboom, J. (2017). Aircraft Health Monitoring: The
True Value of Aircraft Health Monitoring and Data
REFERENCES Management. In 13th Maintenanance Cost
Bernieri, A., D’Apuzzo, M., Sansone, L., & Savastano, M. Conference. Panama City.
(1994). A neural network approach for identification Isermann, R. (2005). Model-based fault-detection and
and fault diagnosis on dynamic systems. IEEE diagnosis - Status and applications. Annual Reviews in
Transactions on Instrumentation and Measurement, Control, 29(1), 71–85.
43(6), 867–873. https://doi.org/10.1109/19.368083 https://doi.org/10.1016/j.arcontrol.2004.12.002
Bertolino, A. C. (2016). Modellazione di Sistema Karpenko, M., Sepehri, N., & Scuse, D. (2003). Diagnosis of
Vite/Madrevite in Attuatori Elettromeccanici per process valve actuator faults using a multilayer neural
Comandi di Volo. Politecnico di Torino. Retrieved network. Control Engineering Practice, 11(11), 1289–
from 1299. https://doi.org/10.1016/S0967-0661(02)00245-9
http://www.informaworld.com/openurl?genre=article
&doi=10.1080/03043798408903597&magic=crossref Kingma, D. P., & Ba, J. (2015). Adam: A Method for
%7C%7CD404A21C5BB053405B1A640AFFD44AE Stochastic Optimization. In ICLR: International

11
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2018

Conference on Learning Representations (pp. 1–15). Adaptive Control and Signal Processing, 11(1), 49–63.
San Diego. https://doi.org/10.1002/(SICI)1099-
https://doi.org/http://doi.acm.org.ezproxy.lib.ucf.edu/1 1115(199702)11:1<49::AID-ACS394>3.0.CO;2-X
0.1145/1830483.1830503
Urata, E. (2007). Influence of unequal air-gap thickness. In
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Proceedings of the Institution of Mechanical
ImageNet Classification with Deep Convolutional Engineers, Part C: Journal of Mechanical Engineering
Neural Networks. Advances In Neural Information Science (Vol. 221, pp. 1287–1297).
Processing Systems, 1–9.
https://doi.org/http://dx.doi.org/10.1016/j.protcy.2014.
09.007
BIOGRAPHIES
Le, T. T., Watton, J., & Pham, D. T. (1998). Fault
Oliver Ritter, graduated in Mechanical Engineering at the
classification of fluid power systems using a dynamics
Institute of Aircraft Systems Engineering of the Technical
feature extraction technique and neural networks.
University of Hamburg in 2013. After graduation he directly
Proceedings of the Institution of Mechanical
Engineers. Part I: Journal of Systems and Control started as a Project Engineer at the Lufthansa Technik Base
Engineering, 212(2), 87–97. in Hamburg. As an EASA 21-J Design Engineer Oliver Ritter
has gained relevant experience in research and innovation
Marino, F. (2017). Advanced Fault Detection and projects at the components division of Lufthansa Technik.
Classification on Primary Flight Control Actuators. The focus of his engineering activities is the development and
Polytechnic of Turin. improvement of test equipment and procedures for flight
control systems. For this purpose, Oliver Ritter is the
Martin, A. De, Jacazio, G., & Vachtsevanos, G. (2017).
responsible operator of a hydraulic test laboratory, where
Windings Fault Detection and Prognosis in Electro-
research and development projects focused on electro-
Mechanical Flight Control Actuators Operating in
hydraulic actuators are performed. Currently he is leading a
Active-Active Configuration. International Journal of
research project for the development of new diagnostic
Prognostics and Health Management, (1991), 0–13.
methods and collaborative robotic assistance during repair
Martini, L. J. (1984). Practical Seal Design. (T. & Francis, process of aircraft components in accordance with aviation
Ed.). regulations.
Merritt, H. E. (1967). Hydraulic Control Systems. (I. John Gerko Wende, graduated in Electrical Engineering at the
Wiley & Sons, Ed.). New York, London, Sydney: John University of Hannover in 1996 and received a PhD in 2003
Wiley & Sons, Inc. from the Institute of Flight Guidance & Control at the
Technical University of Braunschweig. He then continued
Mornacchi, A. (2016). Design and Development of
working as a scientist and flight test engineer at the same
Prognostic and Health Management System for Fly-by-
University with a research focus on flight measurement
Wire Primary Flight Control Electrohydraulic
systems and avionics. During this time he initiated an
Servoactuators. Politecnico di Torino.
University Spin-Off in the framework of European
Mornacchi, A., Vachtsevanos, G., & Jacazio, G. (2015). Promotion of economic development. Since 2004 Gerko
Prognostics and Health Management of an Electro- Wende is employed with Lufthansa Technik where he is
Hydraulic Servo Actuator. In Annual Conference of the currently the Head of the Innovation Management in the
Prognostic and Health Management Society (pp. 1– Product Division Component Services. At his previous
12). activities at Lufthansa Technik he has been responsible for
strategic product planning and research at the department of
Narasimhan, S., Roychoudhury, I., Balaban, E., & Saxena, A.
Original Equipment Innovation. From 2009 until 2012 Gerko
(2010). Combining Model-Based and Feature-
Wende gave lectures at the Technical University of Hamburg
DrivenFeature-Driven Diagnosis Approaches – A Case in the topics of Flight Mechanics, Flight Test and Flight
Study on Electromechanical Actuators. In 21st Measurement Systems.
International Workshop on Principles of Diagnosis
(pp. 1–9). Retrieved from Rocco Gentile, born in Barletta (Italy) in 1992, is a
https://ntrs.nasa.gov/search.jsp?R=20110008515 mechanical engineer and Ph.D. student in Mechanical
Engineering at Politecnico di Torino, where he graduated in
Nielsen, M. (2017). Improving the Way Neural Network
2016. His scientific activity is focused in the area of advanced
Learnitle. Retrieved from diagnostic and prognostic of hydraulic servo-actuators. In
http://neuralnetworksanddeeplearning.com/chap3.html particular, his main activity is focused data-analysis and
Nordin, M., Galic’, J., & Gutman, P.-O. (1997). New models implementation of machine learning techniques in diagnostic
for backlash and gear play. International Journal of field. He is member of the servo-systems and mechatronic

12
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2018

research group of the Department of Mechanical and

Aerospace Engineering.
Francesco Marino, graduated in Aerospace Engineering at
Politecnico di Torino works as research assistant since 2017.
His research activities are focused on primary flight servo-
actuators (hydraulic servo-systems), diagnostic, features
engineering, deep learning and collaborative robots.
Currently, he is mainly involved in a project in partnership
with Lufthansa Technik, where leads research activities
related to: data-analysis, predictive maintenance and
automated maintenance procedures. He is member of the
servo-systems and mechatronic research group.
Antonio Carlo Bertolino, born in Cagliari (Italy) in 1992, is
a Ph.D. student in Mechanical Engineering and Research
assistant in Applied Mechanics at Politecnico di Torino. He
graduated in Mechanical Engineering at Politecnico di Torino
in 2016. His research activity is focused in the area of
aerospace applications, such as flight control systems for
fixed and rotary wing aircrafts (hydraulic and electro-
mechanical servo-systems) and their prognostics. In
particular, his main interest is in the development of high-
fidelity model of ball screws for EMA flight controls. He’s a
student member of ASME since 2017.

Andrea Raviola, he got his bachelor’s degree in aerospace

engineering in 2015 at the Polytechnic of Turin, while now
he is attending his master’s in mechanical engineering at the
Polytechnic of Turin focusing in automation and production
processes. Nowadays, he is writing his thesis about
automation of primary flight control actuators maintenance
procedures using collaborative robotics at Lufthansa Technik
AG. His main interests are: industrial and collaborative
robotics, human-robot interaction and cognitive architectures
in robotics.
Giovanni Jacazio, full professor of Applied Mechanics and
lecturer of control systems at Department of Mechanical and
Aerospace Engineering, Politecnico di Torino, Turin, Italy,
until 2015, now retired; he is lecturing courses on fly-by-wire
flight control systems and PHM in the Doctoral School of
Politecnico di Torino and consulting for high tech companies.
His main research activities are in the areas of flight control
systems for aerospace applications and of prognostics and
health management. He has been Research Leader for several
European and national research programs, Coordinator of
research and consulting activities for several engineering
industries. He is member of the SAE A-6 Committee on
Aerospace Actuation Control and Fluid Power Systems, and
fellow of the Prognostics and Health Management Society.

13