Remaining Useful Life Prediction of a Turbofan

Engine Using Deep Layer Recurrent Neural

Networks

by

Unnati Thakkar

A thesis submitted to the Faculty of Graduate and Postdoctoral Affairs in

partial fulfillment of the requirements for the degree of

Master of Applied Science

in

Electrical and Computer Engineering

Faculty of Engineering and Design

Department of Electronics
Carleton University
Ottawa, Ontario, Canada
Copyright © 2021 Unnati Thakkar
Abstract

Turbofan engine is a pivotal component of the aircraft. Engine components are susceptible to

degradation over the life of their operation which affects the reliability and performance of an

engine. In order to direct the necessary maintenance behavior, remaining useful life prediction is

the key. This thesis presents a prediction framework for the Remaining Useful Life (RUL) of an

aircraft engine using the whole life cycle data and deterioration parameter data based on a machine

learning (ML) approach. In specific, a Deep Layer Recurrent Neural Network (DL-RNN) model

is proposed to address the problem of prognostic instability based on deep learning. In addition,

for the aircraft engine, a new health indicator (HI) measure is implemented based on the pre-

processing of raw data. The proposed method is compared against Multilayer-Perceptron (MLP),

Non-linear Auto Regressive Network with Exogenous Inputs (NARX), Cascade Forward Neural

Network (CFNN) and validated through the IEEE 2008 Prognostics and Health Management

(PHM) conference Challenge dataset and Commercial Modular Aero-Propulsion System

Simulation (C-MAPSS) dataset provided by NASA results reveal a better predictive precision with

respect to other ML algorithms.

i
Acknowledgements

First and foremost, praises and thanks to the Almighty God, for the showers of blessings

throughout my research work to complete the research successfully.

I would like to express my deep and sincere gratitude to my research supervisor Dr. Hicham

Chaoui for giving me the opportunity to do research and providing invaluable guidance throughout

this research. His dynamism, vision, sincerity and motivation have deeply inspired me. It was a

great privilege and honor to work and study under his guidance. He is a great teacher, and his doors

were always open when I had any question. His prompt replies to the emails even during holidays

helped in the rapid progression of this work.

I would like to extend my profound gratitude to Department of Electronics and Prof. Chaoui

for the financial assistance in the form of TA and RA funding.

I am extremely grateful to my family for their unconditional, emotional and financial support

throughout my life and during this project. Without their love, sacrifice, patience and motivation,

this study would not have been feasible.

ii
Table of Contents

Abstract ........................................................................................................................................... i
Acknowledgements ....................................................................................................................... ii
Table of Contents ......................................................................................................................... iii
List of Tables ................................................................................................................................. v
List of Figures ............................................................................................................................... vi
List of Abbreviations .................................................................................................................. vii
Chapter 1 ....................................................................................................................................... 1
1. Introduction ............................................................................................................................... 1
1.1 Motivation .............................................................................................................................. 2
1.2 Objectives and Contribution .................................................................................................. 3
1.3 Organization of the Thesis ..................................................................................................... 4
Chapter 2 ....................................................................................................................................... 5
2. Literature Review ..................................................................................................................... 5
2.1 Machine Learning .................................................................................................................. 5
2.1.1 Deep Learning .................................................................................................................. 6
2.1.2 Artificial Neural Networks .............................................................................................. 7
2.2 RUL Prediction Techniques ................................................................................................. 10
2.2.1 Model-Based Technique ................................................................................................ 11
2.2.2 Data-Driven Based Technique ....................................................................................... 12
2.2.3 Hybrid based Techniques ............................................................................................... 12
Chapter 3 ..................................................................................................................................... 14
3.1 Application: Gas Turbine Engine ........................................................................................ 14
3.1.1 Components of Gas Turbine Engine ............................................................................................ 15
a) Air inlet ............................................................................................................................... 15
c) Combustor ........................................................................................................................... 15
d) Turbine ................................................................................................................................ 16
e) Exhaust nozzle .................................................................................................................... 16
3.1.2 Failure of Turbofan Engine ............................................................................................ 16
3.2 C-MAPSS Dataset for Turbofan Engines RUL Prediction.................................................. 17
3.3 Challenge Dataset for Turbofan Engines RUL Prediction .................................................. 20
Chapter 4 ..................................................................................................................................... 23
4. Deep Learning Framework .................................................................................................... 23

iii
 4.1 Deep Learning in Prognostics for Mechanical Components ............................................... 23
4.1.1 Multilayer Perceptron Neural Networks ........................................................................ 23
4.1.2 Nonlinear Autoregressive Network with Exogenous Inputs ......................................... 25
4.1.3 Cascade Forward Neural Network ................................................................................. 26
4.2 Data Preprocessing............................................................................................................... 27
a) Regime Clustering .............................................................................................................. 29
b) Regime Normalization ........................................................................................................ 30
c) Sensor Selection .................................................................................................................. 31
4.2.1 Health Indicator Assessment.......................................................................................... 33
4.3 Deep Layer-Recurrent Neural Network ............................................................................... 35
4.4 Neural Network Training ..................................................................................................... 37
Chapter 5 ..................................................................................................................................... 41
5.1 RUL Prediction .................................................................................................................... 41
Chapter 6 ..................................................................................................................................... 48
6.1 Conclusions .......................................................................................................................... 48
6.2 Future Work ......................................................................................................................... 49
References .................................................................................................................................... 50

iv
List of Tables

Table 3.1: C-MAPSS Dataset ....................................................................................................... 19

Table 3.2: Variables C-MAPSS Dataset ....................................................................................... 19

Table 3.3: Sensor Measurements .................................................................................................. 21

Table 3.4: Challenge Dataset ........................................................................................................ 22

Table 5.1: Prognostic Metrics ....................................................................................................... 42

Table 5.2: Performance comparison of different methods (FD001 dataset) ................................. 43

Table 5.3: Results of testing (SCG algorithm, FD001 dataset) .................................................... 44

Table 5.4: Performance comparison of different methods (Challenge dataset)............................ 45

Table 5.5: Results of testing (SCG algorithm, Challenge dataset) ............................................... 46

v
List of Figures
Fig. 2.1: Artificial Neural Network Architecture............................................................................ 9

Fig. 2.2: RUL prediction techniques ............................................................................................. 11

Fig. 3.1: Simplified diagram of engine simulated C-MAPSS ...................................................... 17

Fig. 3.2: A layout showing various modules and their connections ............................................. 17

Fig. 4.1: Raw data from the engine source ................................................................................... 30

Fig. 4.2: Operational regimes........................................................................................................ 30

Fig. 4.3: Result of normalization for first five sensors ................................................................. 31

Fig. 4.4: Degradation pattern of 5 useful sensors ......................................................................... 33

Fig. 4.5: Health Indicator of engine of FD001 dataset ................................................................. 35

Fig. 4.6: Health Indicator of engine of Challenge dataset ............................................................ 35

Fig. 4.7: Basic architecture of DL-RNN ....................................................................................... 37

Fig. 4.8: Neural Network validation for FD001 dataset (same input) .......................................... 38

Fig. 4.9: Neural Network validation for FD001 dataset (different input) ..................................... 38

Fig. 4.10: Neural Network validation for Challenge dataset (same input) ................................... 40

Fig. 4.11: Neural Network validation for Challenge dataset (different input).............................. 40

Fig. 5.1: The error plot of DL-RNN (FD001 dataset)................................................................... 41

Fig. 5.2: The error plot of MLP (FD001 dataset) ......................................................................... 42

Fig. 5.3: The error plot of DL-RNN (Challenge dataset) ............................................................. 44

Fig. 5.4: The error plot of NARX (Challenge dataset) ................................................................. 45

Fig. 5.5: RUL of 100 test engines ................................................................................................. 46

Fig. 5.6: Sorted prediction for 100 testing engine instances. ........................................................ 47

vi
List of Abbreviations
Abbreviations Definition

AI Artificial Intelligence

ANN Artificial Neural Network

CFNN Cascade Forward Neural Network

C-MAPSS Commercial Modular Aero-Propulsion System Simulation

CNN Convolutional Neural Network

DDA Data-driven Approaches

DL Deep Learning

DL-RNN Deep Layer-Recurrent Neural Network

GB Gradient Boosting

GA Genetic Algorithm

HI Health Indicator

HMMs Hidden Markov Models

HPC High Pressure Compressor

HPT High Pressure Turbine

ISE Integrated Square Error

LSTM Long-short term memory

LPC Low Pressure Compressor

LPT Low Pressure Turbine

ML Machine Learning

MLP Multi-Layer Perceptron

MAE Mean Absolute Error

vii
MSE Mean Square Error

NARX Non-linear Auto Regressive Network with Exogenous Inputs

PHM Prognostics and Health Management

ReLU Rectifier Linear Unit

RMSE Root Mean Square Error

RUL Remaining Useful Life

RF Random Forest

RVM Relevance Vector Machine

RVR Relevance Vector Regression

SRU Statistical Recurrent Unit

SVM Support Vector Machine

SVR Support Vector Regression

viii
Chapter 1

Introduction
Humans have had to deal with a variety of degradations and failures in their creations over the

years due to the corrosion of materials, structural fatigue, or even just bad design choices. A

washing machine, for example, may be used indefinitely before it stops working. Since a

malfunction of such an asset rarely results in a substantial financial loss or a fatal outcome, almost

no maintenance is performed beforehand. This type of maintenance is called Corrective

Maintenance. On the other hand, an automobile, for example, must be maintained on a regular

basis according to either time or mileage to ensure that it is in good working order; without proper

maintenance, a vehicle may be unsafe. This type of maintenance is called Preventive Maintenance

or Condition Based Maintenance. Condition-based maintenance helps an asset's availability and

protection to be maximized while reducing maintenance costs. Condition-based maintenance

considers not only the asset's past and current health problems, but also its predicted future health.

Predictive maintenance systems and/or applications are commonly referred to as "Prognostics and

Health Management," or "PHM." The findings of PHM strategies have been shown to be accurate

and consistent. However, one of their major drawbacks is that each part studied must be

approached as a unique task, as not all devices work in the same way. As a result, expert’s skill is

needed to properly collect and interpret data from reliability tests for a given system in order to

gain a broad understanding of its health status. Reliability engineers have recently tackled many

of the drawbacks of conventional PHM approaches from a data-driven standpoint. Indeed,

advances in computational hardware and the ongoing development of Machine Learning (ML) and

1
Deep Learning (DL) techniques have allowed engineers to research and implement such

algorithms in a variety of fields.

This thesis presents a DL method for estimating the state of health of any system under

uncertainty using big machinery data. The approach consists of Layer-Recurrent Neural Network

as a novel topology to predict the RUL of the turbofan engine. Extensive comparison is also carried

out against three topologies, i.e., Multilayer Perceptron, Non-linear auto regressive network with

exogenous inputs and Cascade forward neural network. Two different datasets are used to train,

test and validate the proposed architecture. These consist of a C-MAPSS dataset and a Challenge

dataset of turbofan engine.

1.1 Motivation
The need to reliably forecast potential behaviour of variables that control complex machines has

prompted the adoption of data-driven computer health monitoring systems approaches in the

reliability community over the last decade [1] [2], [3]. The advancement of Machine Learning

(ML) and Deep Learning (DL) techniques has increased the usage of Data-Driven Approaches

(DDA) as a supplement to standard methods for estimating the RUL, with Neural Networks

becoming the most widely employed methodology for this purpose. Nonetheless, as it will be

discussed later, there are still many DL implementations in this field that have yet to be tested. The

engine is a heart of an airplane. System faults and the deterioration process will eventually occur

with the emergence of action cycles. It is also crucial to know how often such failures will occur

to enhance system’s performance and reliability. Prognostics and Health Management (PHM) is a

framework that delivers integrated yet individualized approaches for health system management

[4]. Aircraft engine prognostics determine an engine's actual state of health and is the most

complex and technically demanding comprehensive technology in prognosis and health

2
management. Remaining useful life (RUL) is defined as the length from the current time to the

failure of a system. The RUL prognostics is now a universal practice for every system, and it is

the primary concern of many organizations, including NASA and the Prognostics and Health

Management Institute (PHM). These organizations promoted the field by publishing several

datasets and amongst all the two most famous datasets are Challenge dataset and C-MAPSS dataset

for RUL prediction of turbofan engine.

1.2 Objectives and Contribution

The main objective of this thesis is to develop and implement a deep learning-based framework

for estimating the health status of physical assets using data from multisensorial measurements.

We suggest a novel data-driven method for assessing remaining useful life. To predict the RUL,

the approach learns the relationship between acquired sensor data and health indicator. This thesis

focused on two separate datasets provided by NASA namely C-MAPSS dataset and Challenge

dataset. The challenge dataset consists of two datasets: training and testing whereas the C-MAPSS

dataset consists of four datasets and each dataset is further divided into training and testing. Each

row is a snapshot of data acquired during a single cycle of operation; each column is a different

variable. Each engine has 21 sensors which collects data for different measurements including the

temperature, pressure, speed, bypass ration, cooling flow of each module and so on. The first two

columns are the unit number and operational time in cycles whereas the column number three, four

and five are operational settings. Rest of the columns are sensor measurements. The aim of this

thesis is to propose and validate a deep learning framework for producing accurate prognostics on

a system's health. Two separate datasets are used to test the proposed system, and the output results

are correctly described in tables and figures.

3
1.3 Organization of the Thesis
The rest of the thesis is organized as follows. Chapter 2 presents the quick review on the definition

and purposes of Machine Learning and Deep Learning. In this context, DL techniques are

thoroughly explained, addressing their advantages and applications and it also reviews various

techniques for RUL prediction and their classification in detail. Following this, the architecture

and working procedures of the proposed techniques and data description are described in detail in

Chapter 3. Chapter 4 reviews the proposed deep learning framework along with the data

preprocessing in detail. Next, Chapter 5 presents the RUL prediction results obtained in the

MATLAB software along with a comparison against different methods. Finally, Chapter 6

provides the concluding remarks and insights on future research opportunities.

4
Chapter 2

Literature Review
The following chapter details the necessary background to successfully achieve the objectives

stated for this thesis wok. First, a quick review is given on the definition and purposes of Machine

Learning and Deep Learning. In this context, DL techniques are thoroughly explained, addressing

their advantages and applications. This chapter also reviews various techniques for RUL prediction

and their classification in detail.

2.1 Machine Learning

Machine learning became popular in the early 90’s. Before that, in order to describe a process or

phenomenon it was necessary to study the behavior of a system under certain boundary conditions.

Also, it was necessary to analyze different effects over the system functionality and the properties

of the system on which the model is dependent. It could take months, years, or even decades to

identify certain interactions that can reliably model measurable phenomena. Machine learning can

overcome all these issues. Machine Learning is the branch of science that studies how machines

can learn to perform a task without being directly trained to do so. ML algorithms use a variety of

approaches to discover correlations in data based on a given dataset, which are then used to forecast

potential values of determined variables or perform other decision-making tasks [5]. In a more

thorough manner, a ML computer program (model) learns to complete a given task T (output)

from experience E (data) if its success on T improves with experience E, where the performance

5
is assessed by a chosen metric measure P [6]. That is, if the model can learn how to model or

interpret the targeted output variable from the raw data by training itself with new data and

producing better results for the estimation metric P, then the ML algorithm can learn how to model

or interpret the preferred output variable from the input data. As a result, the success of any

machine learning algorithm is highly dependent on the representation of the data it is given. There

are plenty of machine learning methods developed for a variety of applications. However, the core

structure of how the algorithm works remains same for all the models i.e., firstly a parametric ML

algorithm is said to be learned to learn from data and perform a particular task, indicating that the

software has a set of parameters that must be tuned based on the training data. To train the data by

itself, ML uses the training dataset given each sample in the data corresponds to the training

instance. Further the performance of a model is evaluated through the metrices such as Root mean

square error or mean square error depending on the problem the algorithm is intended to solve.

The main benefit of machine learning approaches is that they efficiently handle to analyze

data without requiring specific knowledge about the problem to be solved or specifically writing

down the basic rules that control the algorithm. As a result, the framework is considerably shorter,

simpler to operate, and certainly more reliable.

2.1.1 Deep Learning

Nowadays, machine learning methods can be used in a variety of implementations, ranging from

a speech recognition device in a smartphone that converts a user's voice into text to advertisements

served up by webpages to each specific user depending on their previously searched items.

Processing data in its raw form, on the other hand, necessitates learning skills that are still

constrained by traditional machine learning techniques. This is because it requires designing a

feature extractor that can translate raw data into a suitable internal representation or feature vector

6
for the ML and constructing the features that feed an ML algorithm takes more time, expertise,

and effort than training the ML model itself [7]. To overcome all these issues, DL techniques are

designed. Deep learning is a subset of machine learning which are created by composing simple

non-linear modules that transform the starting raw input into a higher abstract level representation

with multiple level of representation. One of the many approaches to Artificial Intelligence is Deep

Learning (DL) that allows computer system to learn and improve with experience and data [8].

DL algorithms have been around for over two decades, but their deployment has been

hampered by the high computing capacity needed to tune the large number of parameters in a DL

architecture. In addition, in order to effectively train a model, vast datasets must normally be

evaluated, resulting in extensive training and testing procedures. However, in the last decade, DL

algorithms have increased in popularity as the cost of computational components has declined,

allowing students, academics, and virtually all in the science world to access fast computers. This,

along with the advancement of GPU-based computing, has accelerated the adoption of DL

techniques.

2.1.2 Artificial Neural Networks

Artificial Neural Networks have been popular in recent years with their strong ability to

approximate functions [9]. ANNs are biologically inspired networks that are closely equivalent to

the behavior of neural networks in the brain and are used as machine learning applications that are

made up of data processing neurons in the brain, which are neural network components [10] [11].

A biological neuron comprises of many branches classified as dendrites and axon [12]. The

impulse signals are received by the dendrites, and when a certain threshold limit is reached, an

acceptable response signal is produced [12]. With the aid of axon, this response signal is

transmitted to other neurons through a complex branching mechanism [12]. After attaining the

7
neuron body, the incoming signals (inputs) are weighted and summed, and then transformed into

outgoing signals after passing through the transfer function [12]. The input layer, hidden layer/s,

and output layer are the three layers that make up an ANN [13]. The input layer receives data from

the network and transfers it as raw information to the hidden layer/s [13].. This data is processed

by the hidden layer/s and transferred to the output layer [13]. There is an interconnected link

established by a neuron between an input and a desired output. Inside an ANN, each layer performs

a linear transformation by multiplying the data from the previous layer by a weights matrix and

adding a bias term b. A non-linear activation function, such as the Rectifier Linear Unit (ReLU)

or the hyperbolic tangent, is used to determine the output of this linear transformation (tanh) [14].

ANN uses data from continuous monitoring programs which includes samples from preparation

[15]. Fig. 2.1 illustrates an example of an ANN architecture, where a three-layer ANN is presented.

The number of nodes in the NN structure's hidden layer is determined by trial and error. The input

layer feeds the signals to the hidden layer nodes, and the output layer produces the output after

processing them with hyperbolic tangent activation functions. There are various training functions

to train the NN and amongst all Levenberg-Marquardt is most used as it takes less computing time.

Fig. 2.1: Artificial Neural Network Architecture [16]

8
Since deep learning architectures typically offer a lot of flexibility, it is simple for overfitting to

happen during the training phase. That is, it is possible to achieve an over adjustment of the model's

weights and prejudices to the training data as a result of the training phase, resulting in poor

generalization results to unseen data (test data). To overcome this issue, regularization techniques

are applied. Regularization is a strategy for enhancing the generalization of a learning algorithm

by making small changes to it. One more common regularization technique is early stopping which

ends the training loop when training and validation failures begin to diverge.

ANN is the most popular approach for estimating useful life of turbofan engine. An

advanced neural network training is performed in [17] to classify the difference between a good

and a bad system. Recurrent Neural Network (RNN) is proposed in which to create an accurate

and compact algorithm, the recurrent neural network is trained with back-propagation by time

gradient calculations, an Expanded Kalman Filter training process, and evolutionary algorithms.

Initially, a Multi-Layer Perceptron was used to train the algorithm, but the output was very noisy

and the MLP network is reasonably accurate for predicting remaining useful life at the end of the

run when the machine was close to failure. However, the job of forecasting RUL is much more

challenging during the early and middle sections of the race, and the performance is not as good.

Long Short-Term Memory (LSTM) which is an artificial neural network architecture used in the

field of deep learning is proposed in [18] for RUL prediction of turbofan engine. LSTM can learn

how long the old data should be memorized, when it should be forgotten and when new data should

be used. An initial RUL was set to improve the estimation accuracy and the initial RUL of 97

showed better performance compared to initial RUL of 130. The performance of LSTM was

compared with other methods such as Multilayer Perceptron (MLP), Support Vector Regression

(SVR), Relevance Vector Regression (RVR) and Convolutional Neural Network (CNN) and it

9
outperformed all the methods. LSTM performs well with time series data, but it takes a very long

time to converge [19] [20]. In order to predict the RUL of the same engine, Convolution Neural

Networks (CNN) is proposed [21]. For sample processing, it adopts a time window strategy, which

may provide more detail on deterioration. As a result, the dimension of model inputs becomes

greater, so the creation of a neural network model is a daunting task: how to set up network nodes.

A Neural Network based on Statistical Recurrent Unit (SRU) is proposed in [22] for RUL

prediction of turbofan engine. This method can derive hidden partners from multivariate time

series sensor data with various operation state faults and errors and outperforms other dynamic

methods of deep learning, since it obtains a plurality of historical perspectives by linear

combinations of few averages.

2.2 RUL Prediction Techniques

Remaining useful life (RUL) is defined as the length from the current time to the failure of a system

and its prediction techniques are classified in the following ways: model-based technique (also

called physics of failure technique), data driven based technique and hybrid-based technique [23].

The methods are listed in Fig. 2.2. The detailed explanation of each technique is given below.

10
 Fig. 2.2: RUL prediction techniques

2.2.1 Model-Based Technique

Model-based technique also known as physics of failure technique relies on accurate physical

model that represents the system’s dynamics and integrates the physical model with calculated

data to define the model parameters and forecast potential behavior [24]. It is derived from the

model and uses the data on "run-to-failure" and is applicable to decision-making for maintenance

[25]. RUL is often estimated using a model-driven approach, which guides maintenance decisions

based on failure threshold. Model based technique includes Corrosion Model, Abrasion Model,

Taylor Tool Wear Model and Hidden Markov Models (HMMs) [26] [27]. The model-based

technique is easy and fast to implement, and the model can be reused. However, model

development requires a thorough understanding of the system and high-fidelity models can be

11
computationally intensive [28] [29].

2.2.2 Data-Driven Based Technique

The data-driven method uses data from previously acquired data (training data) to analyze the

characteristics of the damage situation currently analyzed and to model the future trend. Data-

driven methods are determined by analyzing frequent condition-monitoring data from device

metrics on daily basis [30] [31] [32]. Learning approaches from machine learning and pattern

analysis are used in data-driven prognostics. Examples of the data driven techniques are

Regression Analysis, Neural Networks, Fuzzy Logic, Bayesian, Random Forest (RF), Gradient

Boosting (GB), Support Vector Machine (SVM), Genetic Algorithm (GA), and Relevance Vector

Machine (RVM) [33] [34]. Data driven technique is easy and fast to implement and it may identify

relationships that were not previously considered. On the other hand, it requires lots of data, a

balanced approach and there is a risk of overlearning and over-generalizing the data [35]. Bayesian

networks also called as belief networks is a probabilistic model that depicts a series of random

variables as well as their conditional interdependencies [36]. In different modes, the Markov model

is represented as a Markov mechanism with unobserved (hidden) state identification. For turbofan

engines with powerful nonlinear and time-dependent characteristics, the SVM and RF are not

appropriate, because the dynamic interaction between sensor data and RUL cannot be captured

within a comparatively simple framework. Among all the data driven techniques, Artificial Neural

Networks are commonly used for prognostics.

2.2.3 Hybrid based Techniques

Hybrid based technique also known as fusion-based technique is the combination of model based

and data-driven technique. It uses the data to learn the model parameters and the knowledge about

the physical process to determine the type of regression analysis to apply (linear, polynomial,

12
exponential, etc.). It forecasts RUL independently and, using probability theory-based approaches,

allows the combination of two or more RUL prediction outcomes to produce a new RUL. The

hybrid-based technique combines the strength of each approach and it uses the data where system

knowledge is lacking, and it uses physics where data is lacking. On the other hand, it requires in-

depth knowledge of the system and meaningful data is still needed [37] [38]. The examples of

hybrid-based technique are Particle filters, Kalman filters etc. [39].

Next chapter provides the architecture and working procedures of the proposed techniques and

theories are described in detail. It also provides the data description in detail.

13
Chapter 3

3.1 Application: Gas Turbine Engine

A gas turbine engine, also known as a jet engine, is a type of turbine engine that spins pressurized

gas to generate electricity or to provide an airplane with kinetic energy [40]. There are mainly five

types of gas turbine engines, namely, turbojet, turbofan, turboprop, turboshaft and ramjet engine

[40]. The turbofan engine or fanjet is the modern variant of gas turbine engine that is widely used

for aircraft propulsion. Propulsion is the net force produce by the thrust that push the object

forward [40]. So, because of their high thrust and good efficiency, almost all airliners use turbofan

engines. The engine refers to the mechanism that transforms the energy of the fuel into shaft power

and the power of the shaft power to drive it. The engine is composed of air intake, a compressor

(low and high pressure), a combustor or combustion chamber, a turbine (high and low pressure)

and an outlet or exhaust nozzle. The air enters and then travels through the compressor to reach

high pressure through the air inlet. Further, increases the temperature of the high-pressure air

burning (mixing) it in the combustor with the gasoline. The combustion results in high levels of

hot velocity gases that move through the turbines, producing energy to power the turbines from a

compressor. Such high-speed hot gases are exhausted from the compressor to the turbine. The high

velocity gas coming from turbine is drained out through the jet nozzle which produces a thrust to

drive the engine forward outside through the outlet. Some common advantages of turbofan engine

are it is fuel efficient and it produces less noise than turbojet engine. On the other hand, being

massive than turbojet and inefficient at very high altitudes are the two disadvantages [40].

14
3.1.1 Components of Gas Turbine Engine

The working principal of turbofan engine is explained in detail below. There are five main parts

of the gas turbine engine which plays a very important role in the operation of an engine.

a) Air inlet

The air travels through the engine from left to right passing through the fan. The fan oversees

producing most of the turbofan engine-generated thrust. The fan is interconnected to the low-

pressure compressor and the low-pressure turbine via the shaft. Most of the air passing through the

fan travels around the core of the engine which is known as bypass air and it never interacts with

other components of engine. Infact it is directly accelerated from the rear of an engine [41]. The

leftover air enters an engine's heart and then enters the low-pressure compressor.

b) Compressor

The main function of the compressor is to add energy to the air in the form of pressure and heat

which makes it ready for combustion. As the pressure increases, temperature of the air also

increases. There are primarily two types of compressor: the low-pressure compressor (LPC) and

the high-pressure compressor (HPC). The low-pressure compressor is coupled with the fan and the

low-pressure turbine via the low-pressure shaft. Whereas, directly downstream of the LPC and

directly upstream of the combustor is the high-pressure compressor located which is connected to

high-pressure turbine through high pressure shaft [41].

c) Combustor

The hot pressure air travelling through the compressor mixes with the fuel and burned in

combustion chamber. This is where the reaction takes place, and it results in very high temperature

and high velocity gases which is used to run the turbine [41]. Combustion chamber is placed just

after the compressor and they are made with the materials that can withstand high temperature

15
resulting from combustion. It is found in an engine's heart.

d) Turbine

Turbine consists of row of blades which spins. The high temperature and high velocity gas from

the combustor pass through the turbine which extracts the energy and transfers it to the compressor

and fan through connecting shaft. This energy helps compressor and fan to spin. The temperature

and pressure of the air leaving the turbine is relatively low due to the energy extracted by the

turbine. Two types of turbines primarily exist: the high-pressure turbine (HPT) and the low-

pressure turbine (LPT). The high-pressure turbine is located to the downstream of the combustor

or burner and to the upstream of low-pressure turbine. Whereas the low-pressure turbine is located

to the downstream of high-pressure turbine and to the upstream of exhaust nozzle [41].

e) Exhaust nozzle

Exhaust nozzle is the component located at the right most of the engine and just after the turbine.

Also, it is the last component the air passes through before leaving the engine. It is a tube-shaped

component which helps to produce additional thrust which helps the engine move forward [41].

Moreover, it also helps regulating pressure within the engine which helps other components

functioning properly

3.1.2 Failure of Turbofan Engine

Just like the life of human being is affected by their health same as all systems deteriorate as a

result of time, usage, and environmental conditions. Gas-turbine engines may be subjected to

severe environmental and operating conditions such as corrosion, wear, buckling, erosion etc.

which eventually lead to costly and catastrophic failures if a run-to-failure philosophy is adopted

[42]. Hence, in order to monitor the state of gas turbine engine more than two hundred sensors are

installed in each gas turbine. An engine health deteriorates substantially over time. Certain engine

16
variables which are also known as health parameters have significant impact on engine health.

Different engines have different health parameters. Hence it becomes important to monitor and

evaluate these health parameters in order to improve life, performance, reliability, etc. of engine.

However, due to the complexity of aircraft and sometimes due to unavailability of devices to

measure certain parameters it is not possible to directly measure the health parameters [42] .Hence

it is of great importance to have fault diagnosis and prognostics and health management (PHM)

on engine. The prognostics on remaining useful life (RUL) of an engine is the most difficult and

challenging part among PHM.

3.2 C-MAPSS Dataset for Turbofan Engines RUL

Prediction
The first dataset used to test the proposed deep learning architecture is the C-MAPSS data

competition benchmark data [43].

Fig. 3.1: Simplified diagram of engine Fig. 3.2: C-MAPSS modular components [44]
simulated C-MAPSS [44]

17
The engine diagram in Fig. 3.1 shows the main elements of the engine model and the modular

components of the simulation software and their connections are shown in Fig. 3.2. Commercial

Modular Aero-Propulsion System Simulation (C-MAPSS) simulates the 90000lb thrust level

commercial high bypass ratio turbofan engine by MATLAB/Simulink, which built the engine and

relative control system model [44]. The PHM08 Challenge dataset and Turbofan Engine

Degradation Simulations were implemented using the C-MAPSS software [44] .Training and test

trajectories are two types of datasets, with the latter being a subset of the former. Operational cases

of full run-to-failure data are generated in the training trajectories, and these are intended to be

used to train the algorithms. Test trajectories, on the other hand, can only be set up using shorter

instances of data up to a certain point before the method fails. Four different operating conditions

are simulated using a collection of trajectories, as shown in Table 3.1. To model uncertainty in

parameter readings during operation, the sensors have been polluted with operating regimes and

noise. In addition, each trajectory has its own degree of initial wear and manufacturing variance,

which is considered natural and unnoticed by the user. Each dataset is stored in a CSV file, with

each row representing one time-step (measured in cycles) and containing 21 sensor measurements,

three operating conditions, and other useful information (see Table 3.2). Since each time series is

from a different engine, the data from the C-MAPSS dataset may be considered to come from a

fleet of similar engines (Unit Number).

18
 Table 3.1: C-MAPSS Dataset

Dataset Train Trajectories Test Trajectories Conditions Fault Mode

FD001 100 100 One (sea level) One (HPC)

FD002 260 259 Six One (HPC)

FD003 100 100 One (sea level) Two (HPC, Fan)

FD004 248 249 Six Two (HPC, Fan)

Table 3.2: Variables C-MAPSS Dataset

Index Data description

1 Unit number
2 Time (in cycles)
3 Operational setting 1
4 Operational setting 2
5 Operational setting 3
6 Sensor measurement 1
7 Sensor measurement 2
8 …
9 Sensor measurement 21

The engine's performance is greatly affected by the three operational settings used in the dataset.

The engine begins working normally with an unknown initial deterioration for each trajectory in

the training sets before failure is detected. Sensor noise has also contaminated the data. The time

series in the test set ends just before the device fails, and the RUL for each test engine is given in

a separate CSV file. In this thesis, only the first sub-dataset is evaluated i.e., FD001.

19
3.3 Challenge Dataset for Turbofan Engines RUL
Prediction
Over two hundred multivariate time-series datasets were provided as a part of PHM 2008

prognostics data challenge [45]. Each dataset is unique with unknown initial wear and

manufacturing variation. The datasets consist of three operational settings and twenty-one sensor

measurements that are contaminated with sensor noise. Three datasets are provided namely

“training” and “testing” and “final”. Training trajectories are used to train the algorithm to predict

remaining useful life and is formed of complete run-to-failure data. Each time series is from a

different engine, i.e., the information can be of the same kind from a fleet of engines. The trained

algorithm is then feed to test trajectories which are formed of short instance time-series that ends

some time prior to degradation. Users will implement their methods to the final test dataset and

submit the vector of expected RULs to the Prognostics Center of Excellence for review after their

algorithms have been trained to their satisfaction. There is a web-based framework for uploading

test data and calculating an overall score feedback. Each row is a snapshot of data acquired during

a single cycle of operation; each column is a different variable. Each engine has 21 sensors which

collects data for different measurements of engine. The minimum unit length is 127 cycles whereas

maximum unit length is 356 cycles. The number of trajectories for training, testing and final test

are 218, 218 and 435, respectively. The first two columns are the unit number and operational time

in cycles whereas the column number three, four and five are operational settings. Rest of the

columns are sensor measurements as shown in Table 3.3.

20
 Table 3.3: Sensor Measurements
Symbol Description Units
◦
Sensor 1 T2 Total temperature at fan inlet R

◦
Sensor 2 T24 Total temperature at LPC outlet R

◦
Sensor 3 T30 Total temperature at HPC outlet R

◦
Sensor 4 T50 Total temperature at LPT outlet R

Sensor 5 P2 Pressure at fan inlet psia

Sensor 6 P15 Total pressure in bypass-duct psia

Sensor 7 P30 Total pressure at HPC outlet psia

Sensor 8 Nf Physical fan speed rpm

Sensor 9 Nc Physical core speed rpm

Sensor 10 epr Engine pressure ratio -

Sensor 11 Ps30 Static pressure at HPC outlet psia

Sensor 12 phi Ratio of fuel flow to Ps30 pps/psi

Sensor 13 NRf Corrected fan speed rpm

Sensor 14 NRc Corrected core speed rpm

Sensor 15 BPR Bypass ratio -

Sensor 16 farB Burner fuel-air ratio -

Sensor 17 htBleed Bleed Enthalpy -

Sensor 18 Nf_dmd Demanded fan speed rpm

Sensor 19 PCNfR_dmd Demanded corrected fan speed rpm

Sensor 20 W31 HPT coolant bleed lbm/s

Sensor 21 W32 LPT coolant bleed lbm/s

21
Table 3.4 below shows the number of trajectories, faults modes and unit length of all three

trajectories.

Table 3.4: Challenge Dataset

Train Trajectories Test Trajectories Final Test Trajectories

Number of Trajectories 218 218 435

Fault Modes One (HPC) One (HPC) One (HPC)

Maximum Unit Length 357 364 298

Minimum Unit Length 128 15 20

22
Chapter 4

Deep Learning Framework

Many deep learning techniques have been used for RUL prediction of turbofan engine. Recurrent

Neural Network trained with back-propagation through time gradient calculations showed a great

performance for RUL prediction. On the other hand, LSTM neural network outperformed all other

methods such as MLP, SVR, RVR and CNN and the prediction accuracy was improved by setting

an initial RUL of 97. This chapter will discuss the deep learning techniques used in this thesis.

4.1 Deep Learning in Prognostics for Mechanical

Components

4.1.1 Multilayer Perceptron Neural Networks

The multilayer perceptron neural network is made up of several layers, as its name suggests.

Multilayer perceptron (MLP) is a supplement of feedforward artificial neural networks having one

or more hidden layers. They are widely used for pattern recognition, input pattern classification,

prediction based on input data, and approximation [46].The connection structure of a feedforward

neural network is unidirectional multi-layer, with each layer consisting of many neurons. Each

neuron receives the output of the previous layer's neurons and produces the output for the next

layer's neurons. Generally Multilayer Perceptron Neural Network consists of input layer, hidden

layer and an output layer [47]. Every perceptron sends outputs to all perceptron’s on the second

layer i.e., the hidden layer from the first layer on the left i.e., the input layer and all perceptrons on

23
the second layer transmit outputs to the final layer on the right i.e., the output layer. Every

perceptron transmits multiple signals which passes through the next layer. Moreover, the

perceptron sets unique weights for every signal it transmits. In order to train the algorithm for

prediction remaining useful life, multilayer perceptron was used. Multilayer perceptron algorithm

was implemented using MATLAB with ten inputs and one target for the challenge data and

fourteen inputs and one target for the C-MAPSS dataset. The inputs and the output (target) were

all normalized between the range of [0,1] using min-max. The inputs were the normalized raw

sensor data of useful sensors and the target was HI. Four hidden nodes were used in first layer

whereas five hidden nodes were used in second layer. The training function used to train the

algorithm was Levenberg-Marquardt which takes less time compared to Bayesian regularization

but more memory to train and it avoids overfitting. The performance function was set to mean

square error (MSE). The algorithm was trained for around 58 seconds to 1 minute 26 seconds with

18000 epochs and then the same training raw data was incorporated into the received network

feature to verify that the learned neural network precision is adequate. The validation resulted into

the similar pattern but when compared to the original target the results were not satisfactory. The

MLP network is reasonably accurate for predicting remaining useful life at the end of the run when

the machine was close to failure. However, the job of forecasting RUL is much more challenging

during the early and middle sections of the race, and the performance is not as good. Various

changes were made to improve the performance of the network such as different training functions

were used namely Bayesian regularization, Scaled conjugate gradient backpropagation, RPROP

backpropagation and so on. Also, different performance functions and transfer functions were

implemented in order to improve accuracy of the function, but the performance did not improve.

24
4.1.2 Nonlinear Autoregressive Network with Exogenous Inputs

The non-linear autoregressive network with exogenous inputs is a persistent hierarchical network,

with feedback relations encompassing many layers of the network. A NARX network consists of

a multilayer perceptron taking the variables of the input state as a window for past inputs and the

values of the output and determines the current output. This implies that the current value of a time

series is related by the model to both: values from the history of the same series; and present and

historical values of the (exogenous) driving sequence [48] [49]. The NARX model can be

considered as a dynamic recurrent neural network and is based on the ARX linear model, which is

widely used in the simulation of time series [48] [49]. In several dynamic systems with complex

nonlinearities, NARX has been used. NARX neural networks are not only computationally

efficient in theory, but they also have a range of practical advantages. Gradient-descent learning,

for example, has been shown to be more efficient in NARX networks than in other neural network

architectures with "hidden states" [50]. Despite the NARX network's benefits, its potential as a

nonlinear method for univariate time series modelling and prediction has yet to be thoroughly

investigated [51]. After the poor accuracy of MLP neural network, the NARX algorithm was

implemented using MATLAB. A NARX network can be trained in two ways [52]:

a) Series-parallel mode sequence or open loop mode NARX network in which the actual

output is directly used instead of the approximate output being fed back.

b) The parallel mode of the NARX network or the closed loop network is to feed the expected

performance back to the network.

Input and output sequence initialization is the first step in the training of the network structure.

Inputs were configured as a "10*time" cell of usable raw sensor data for the Challenge data and

"14*time" cell for C-MAPSS dataset, while the target series was configured in the previous

25
segment as a 1*time HI cell. The input and feedback delays were set to two along with two hidden

layers in total. The first hidden layer has four nodes, and the second hidden layer has five nodes.

Both the inputs and the target were mapped in the range of [0,1] using min-max. The algorithm

was trained initially in an open loop mode and then changed to closed loop mode. Levenberg-

Marquardt was used as a training function and the convergence time was around 1 minute 4

seconds to around 1 minute 50 seconds with tansig as a transfer function. After training the

algorithm, the validation was carried out on same raw inputs to verify the accuracy of trained

neural network. The validation resulted a very poor performance of the network. Smooth data was

used to filter the noise and smooth the output.

4.1.3 Cascade Forward Neural Network

Cascade-forward neural networks are similar to feed-forward networks in that they have a

connection from the input and every previous layer to the subsequent layers [53]. In a three-layer

network, the output layer is connected directly to the input layer as well as the secret layer.

Provided enough hidden neurons, a two-or more layer cascade-network, like feed-forward

networks, can learn any finite input-output relationship arbitrarily well [53]. For any form of input

to output mapping, a cascade-forward neural network can be used. This approach has the advantage

of accommodating the nonlinear relationship between input and output while leaving the linear

relationship intact [54]. We use the network in the field of time series in this analysis. In both the

input and hidden units, the optimal architecture was computed using the incremental search

process. The basic one was built first, and then the more complicated one was built by one-by-one

adding the modules. The best option is then selected based on the mean square error criterion. To

make predictions of time series data, a CFNN model that is nonlinear-nonparametric and more

versatile can be used. This allows for more accurate results [54].

26
The issue with CFNN modelling is the input selection process, which must be followed in order to

obtain the best architecture. The initialization of the input and output sequences is the first step in

the training of the network structure. For the Challenge data, inputs were configured as a "10*time"

cell of accessible raw sensor data and "14*time" cell for the C-MAPSS dataset, while the goal

series was configured as a 1*time HI cell. Two input and feedback delays, as well as two hidden

layers, were used in total. With tansig as a transition tool, the convergence period was around 1

minute 8 seconds to around 1 minute 35 seconds with Levenberg-Marquardt as a training function.

After training the algorithm, validation was performed on the same raw inputs to ensure that the

trained neural network was accurate. The network's success was severely harmed as a result of the

validation.

The proposed deep learning framework is based on Layer-Recurrent Neural Network. An

exponential RUL target is proposed as an objective for C-MAPSS dataset. Raw data is normalized

with a Min-Max scaler in the range [0,1] before training the resulting model, and variables from

the dataset that do not contribute useful information to the model or were used for label generation

are dropped. Data pre-processing is performed on Challenge dataset to remove senor noise and to

construct health indicator as a target for neural network training. For FD001, dataset of C-MAPSS

health indicator is constructed through an exponential degradation model by using the initial RUL

value. Once the health indicator is constructed for both the datasets, the proposed architecture is

trained and tested through the validation datasets described in Chapter 3.

4.2 Data Preprocessing

The pre-processing of raw data is a required step in any analysis that uses data-driven techniques

[55]. NASA published a dataset with 21 sensor signal variables. The spectrum of these signals may

27
be vastly different, having a significant effect on model training. Therefore, signals must be

normalized. There are several normalization techniques like scaling to range, clipping, log scaling

and z-score. To normalize the data and prevent overfitting the model over variables with a high

order of magnitude, a scaling to range normalization, also known as Min-Max normalization, is

applied to both training and testing trajectories of C-MAPSS dataset along each variable. Equation

describes how to evaluate the scaled value 𝑋𝑠𝑐𝑎𝑙𝑒𝑑 of a vector 𝑋. The minimum and maximum

values of each column function are represented by 𝑋𝑚𝑖𝑛 and 𝑋𝑚𝑎𝑥, respectively.

𝑋−𝑋𝑚𝑖𝑛
𝑋𝑠𝑐𝑎𝑙𝑒𝑑 = (4.1)
𝑋𝑚𝑎𝑥−𝑋𝑚𝑖𝑛

In addition, variables that do not alter over time and thus could affect the results are removed. In

this thesis fourteen useful sensor measurements 2,3,4,7,8,9,11,12,13,14,15,17,20,21 are selected,

and the irregular/unchanged sensor data are abandoned, as [56] [57] [58].

A similar approach is taken to deal with the challenge dataset. A data analysis is used in

this work to explain feature extraction, data cleaning, and feature collection in order to improve

the multi regime normalization tool and obtain a useful health index for prognosis. A data

processing approach that involves feature extraction, data cleaning and feature selection is

necessary to have an HI that is useful for prognosis. Feature extraction is linked to data cleaning,

which is the way of detecting and likely correcting faulty or inaccurate measurements from data.

It involves gathering field information to locate inaccurate or irrelevant parts for correction or

standardization. Then, to pick the appropriate features for use in RUL prediction construction, a

feature selection procedure is carried out. Without a validation method, it is impossible to find the

best attainable high precision model in real-world data. Such data is often inconsistent, incomplete,

and deficient habits, as well as containing multiple errors. Data preprocessing entails converting

these data into a readable format. As a result, the approach starts by resolving issues related to

28
organizing condition monitoring data, such as reducing data reliability and enhancing data

integrity. The dataset provided by NASA has 21 sensors measuring parameters which includes

temperature, pressure, bypass ratio. The range of the variables may be very different which urge

the need of normalization as it may have great impact on the training of algorithm. Moreover, the

sensor data of training dataset is contaminated with lot of noise. As can be seen from the Fig. 4.1,

the raw data directly comes from the engine source and is not subjected to any processing or

cleaning methods. Also, each trajectory has specific initial wear and is contaminated with noise.

Hence pre-processing of the raw data is needed which includes regime clustering, regime

normalization and sensor selection. The pre-processing was performed on raw data in order to

transform into an understandable data before switching on to the remaining useful life. The steps

of pre-processing include regime clustering, regime normalization, sensor selection, trendability

analysis and constructing health indicator [59]. The pre-processing was performed on both training

and test trajectories. The pre-processing steps are shown below:

a) Regime Clustering

This is the very first step of data pre-processing which is used to identify the operational regimes.

The operational settings in the engine data have a direct effect on the sensors, and several regimes

can be discovered by determining the optimal number of clusters in these settings. It is unlikely

that all operational settings in different data would respond to a clustering approach in the same

way. Some people may react better than others, while others will not respond at all. Clustering, on

the other hand, should not be a one-size-fits-all approach to operational settings, but rather as a

general problem to solve. Clustering is a method used for large datasets to identify and group

similar data points. The number of regimes can be identified by determining the number of clusters.

There are various methods of regime clustering such as k-means clustering, hierarchical clustering,

29
grid-based clustering etc. This thesis uses k-means algorithm. After applying k-means algorithm,

it was found that there are six cluster points which means there are six operational regimes of three

operational settings as shown in Fig. 4.2 [59].

Fig. 4.1: Raw data from the engine source Fig. 4.2: Operational regimes for three
operational settings

b) Regime Normalization

The dataset provided by NASA includes raw values that are multivariate, incompatible with one

another, and work under various regimes. The raw data must be scaled down from their original

scales to something more common. Normalization is the method carried out on raw data so that all

the data are on a common scale and normalized. By returning raw values into a single domain, a

normalization is a method for performing these changes. This includes calculating z-score which

is nothing but subtracting each regime’s population average from each individual raw sensor and

then dividing the difference by the standard deviation of the population. The formula to calculate

z-score normalization is given as:

𝑥− 𝜇
𝑍 − 𝑠𝑐𝑜𝑟𝑒 = (4.2)
𝜎

30
Where 𝑥 is the raw data value of the regimes, 𝜇 is the population average and 𝜎 is the standard

deviation of the population. The population mean is given by:

1
𝜇= ∑𝑛𝑖=0 𝑥 (4.3)
𝑛

Where 𝑛 is the number of observations and standard deviation is given as:

∑𝑛
𝑖=1(𝑥− 𝜇)
2
𝜎=√ (4.4)
𝑛

After applying this equation to each regime separately, the standardized sensors are used to

reassemble the normalized dataset, considering the sensor locations at the outset in order to allocate

them to a common scale while preserving the initial wear characteristics within each trajectory.

The result of normalization for the first five sensors i.e., sensor 1, sensor 2, sensor 3, sensor 4 and

sensor 5 are shown in Fig. 4.3 [59].

Fig. 4.3: Result of normalization for first five sensors.

c) Sensor Selection

After performing regime normalization, we can see the trend of all 21 sensors. It is important to

exclude those sensors that do not accurately reflect continuous normalized values exhibiting a

monotonic exponential trend during the operational lifetime of the training units in order to

31
determine how well the sensors represent the degradation pattern [59]. For doing this, this study

uses the prognostic parameter-choosing measures "Monotonicity," "Prognosability," and

"Trendability" [60] to allocate the useful (and meaningful) sensors to be used in subsequent stages.

Monotonicity is a simple metric for determining if a sensor has the same underlying positive or

negative patterns that are needed for prognosis. This metric defines the general increasing or

decreasing existence of a sensor, showing the system's degradation process. It is defined as

follows:
𝑑 𝑑
#𝑝𝑜𝑠 #𝑛𝑒𝑔
𝑀𝑜𝑛𝑜𝑡𝑜𝑛𝑖𝑐𝑖𝑡𝑦 = 𝑚𝑒𝑎𝑛 (| 𝑑𝑥
− 𝑑𝑥
|) (4.5)
𝑛−1 𝑛−1

where 𝑛 is the number of trajectories. The average difference of the fraction of positive and

negative derivatives for each trajectory is used to determine the monotonicity of a sensor

population. This metric detects whether the sensor has a characteristic underlying positive or

negative trend because device deterioration follows monotonic trends in all trajectories in a dataset.

The deviation of the failure threshold points for each trajectory is divided by the average

variance of the sensor over its entire operating period to measure prognosability [60]. The main

aim of this metric is to quantify the distribution of a sensor's failure value (or threshold point) over

a population of trajectories. To achieve the desired 0 to 1 scale, the calculation is exponentially

weighted using the following equation:

𝜎𝑓𝑎𝑖𝑙𝑢𝑟𝑒
𝑃𝑟𝑜𝑔𝑛𝑜𝑠𝑎𝑏𝑖𝑙𝑖𝑡𝑦 = 𝑒𝑥𝑝 (− ) (4.6)
𝜇𝑓𝑎𝑖𝑙𝑢𝑟𝑒−𝜇ℎ𝑒𝑎𝑙𝑡ℎ𝑦

Prognosability measures near 1 indicate that the failure points in all trajectories are identical

domains and that the sensors are useful for prognosis, while measures near 0 indicate that the

failure thresholds do not fit and that the sensors are unable to provide meaningful prognostic

calculations [60].

32
 The minimum absolute correlation computed among all trajectories is used to determine

trendability. The metric primarily determines if the sensor for a group of trajectories has the same

underlying trend pattern and, as a result, can be represented by the same parametric function.

Trendability analysis includes the selection of most trendable sensors measurements in order to

construct health indicator (HI) and the degradation pattern of each sensor is estimated [59].

Mathematical expression for this is given as:

𝑇𝑟𝑒𝑛𝑑𝑎𝑏𝑖𝑙𝑖𝑡𝑦 = 𝑚𝑖𝑛(| 𝑐𝑜𝑟𝑟𝑐𝑜𝑒𝑓𝑓𝑖𝑗|) (4.7)

where 𝑐𝑜𝑟𝑟𝑐𝑜𝑒𝑓𝑓𝑖𝑗 is the correlation coefficients computed among all the training trajectories.

Hence, by performing the trendability analysis ten sensors are selected that has the same

degradation pattern that are 2,3,4,7,11,12,15,17,20,21. These ten sensors were found useful for the

calculation of health indicator. The degradation pattern of five useful sensors out of ten are shown

in Fig. 4.4.

Fig. 4.4: Degradation pattern of 5 useful sensors

4.2.1 Health Indicator Assessment

The actual RUL values in C-MAPSS dataset is given only for test data hence it is important to

33
label the training data correctly. Setting the RUL to be the form of (4.8) is an easy and convenient

procedure [61]. The health equation can be written as:

ℎ(𝑡) = 1 + 𝑑 − exp {𝑎 ∗ 𝑡 𝑏 } (4.8)

where 𝑑 is the non-zero initial degradation, index 𝑡 is the number of cycles, 𝑎 and 𝑏 are the

coefficients from which 𝑏 is set to 1 as a default value and 𝑎 can be written as (4.9) at the end of

cycle when ℎ(𝑡) is 0.

𝑎 = log(1 + 𝑑) /𝑡 (4.9)

The deterioration of an engine system, on the other hand, is negligible before the engine has been

running for a while and the malfunction starts. As a result, the RUL of the engine system should

start at a constant value and gradually decrease as the failure progresses. But how do you work out

what RUL is worth in the early stages? It was estimated to be about 130 cycles in literature [17].

The reason for selecting 130 as the initial RUL, on the other hand, is seldom clarified. We

suggested a method for selecting an acceptable initial RUL in this thesis. The coefficient 𝑎 is tuned

in such a way that at zero time cycle it gives an initial RUL value of that engine and at the end of

the cycle it gives zero RUL. Fig. 4.5 below shows the RUL of engine no.1 which shows when the

time cycle is zero the RUL value is 112 which is the initial RUL of engine no.1and it degrades as

the time cycles increase and at the end of cycle or when the maximum threshold is reached the

RUL is zero. This method is adopted for all 100 engines in FD001, and the constructed health

indicator is further used as a target to train the algorithm.

For the Challenge dataset initial RUL values are yet not revealed. So, the above method

cannot be applied for challenge dataset for health indicator assessment. Instead, a health indicator

is constructed by fusing all 21 sensors. The health indicator is assumed to degrade over the time

exponentially from 1 to 0. Hence the health situation at the beginning is assumed to be 1 and at

34
failure it is assumed to be 0 [59]. Moreover, a fit linear regression model was constructed for the

ten useful sensors as regressors. The fused health indicator for training data is shown in Fig. 4.6

which is further used as an output of neural network. This step was performed for both training

and test trajectories.

Fig. 4.5: Health Indicator of engine no.1 of Fig. 4.6: Fused health indicator for
FD001 dataset. Challenge dataset.

4.3 Deep Layer-Recurrent Neural Network

A Deep Layer-Recurrent Neural Network (DL-RNN) module uses within-layer recurrence to learn

contextual knowledge adaptively. This approach is inspired by the ReNet architecture of Visin et

al. [62]. A DL-RNN module is a combination of 1D RNNs that can learn contextual knowledge

adaptively, with the successful receptive field able to reach across the entire function map or

picture if necessary, for the task [62]. Recurrent neural networks with layers are analogous to

feedforward networks, except that each layer has a recurrent relation associated with a tap delay.

This enables the network to have an infinite dynamic response to input data from the time series

[63]. In the training phase, DL-RNN used to consider the previous values. It consists of two parts

35
recurrent layers and feedforward part. With just a few parameters, the DL-RNNs can learn long-

range dependencies across a layer efficiently. Fully recurrent networks, recursive neural networks,

Hopfield networks, Elman and Jordan neural networks, continuous-time RNNs, and bidirectional

RNNs are few examples of DL-RNNs. The Fully recurrent network was established in the 1980s

and is a network of neuron-like units. Each unit is linked to the others in a guided manner. Each

connection has a real-valued weight that can be modified. Hopfield neural network is a form of

recurrent artificial neural network that was first developed in 1982. John Hopfield, the inventor,

was given the name. It has an important function in that it ensures that all its dynamics can

converge to the local minimum. It does, however, often converge to a false local minimum. It is a

unique neural network since it does not use sequences of patterns. Same sets of weights are

recursively distributed over a structure in a recursive neural network, as the name suggests [64].

Elman neural network is distinguished by the presence of local memory and feedback connections.

J.L Elman was the one who first suggested it in 1990. It is both a two-layer neural network and a

back propagation neural network [65]. The basic DL-RNN network's architecture is shown in Fig.

4.7. It has an input layer denoted by 𝑢(𝑘), two hidden layers denoted by 𝑣(𝑘) and an output layer

𝑦(𝑘) [66]. In the DL-RNN, certain node’s output is used as an input to the other nodes to construct

an iterated feedback loop. The DL-RNN uses feedback data to recall the values from the previous

step. Therefore, based on the previous and current input, a new output will be produced [66].

36
 Fig. 4.7: Basic architecture of DL-RNN [66]

4.4 Neural Network Training

The first step in training the network is to configure the input-output series after the useful health

features have been extracted from raw data. The dynamic data from 14 different variables of raw

useful sensor data is expressed in the inputs as a “14*t” cell of a matrix for C-MAPSS dataset. The

goal series is a matrix with a “1*t” cell array that only reflects “t” time steps of an output element,

which is the measured HI from the health equation. This configuration's purpose is to provide a

neural network mapping between a collection of raw numeric inputs and a set of measured health

indicators. From the training set, only 70% of the training dataset is used for training, while 15%

is used for validation, and another 15% for testing. The layer-recurrent neural network is designed

in a triple loop design with an increasing number of hidden layers in the outer loop along with a

default value of layer delays and scaled conjugate gradient is used as a training function. In short,

the hidden layers are not set to a default number. During the neural network learning stage, the

inputs are the raw sensors values for the same trajectory and the target is the HI constructed for

37
the same trajectory. The same raw data is applied to the obtained network function to receive a

neural network prediction, ensuring that the trained neural network can achieve high accuracy for

an output estimate. As seen in Fig. 4.8, the results show a pattern that is very similar to the fitted

HI i.e., target used in network training, but it was contaminated with noise which was filtered using

smooth data the network was able to achieve good accuracy after 57 seconds of training. Both

patterns have exponential curves and scales that are identical. As a result, it can be concluded that

the network can efficiently process raw data and standardize outputs without needing all the

dataset's trajectories. However, in the training of multilayer neural networks, due to overfitting and

computational overheads, the network can result in poor network calculations. As a result, the

validate network function must be used on new inputs of test trajectories to ensure that the

algorithm can memorize the samples for the training data and correctly generalize the upcoming

testing cases. The model was able to correctly describe the HI curve of the inserted data despite

differences in duration and characteristic trajectory features (such as initial wear level and

deterioration pattern) as seen from the Fig. 4.9.

Fig. 4.8: Neural Network validation with Fig. 4.9: Neural Network validation with
the same input for FD001 dataset. the different input for FD001 dataset.

38
The training is performed on the entire training set of 100 engines. The raw test input shown in

Fig. 4.9 has shorter instances.

For Challenge dataset, the training is performed with “10*t” cell of a matrix as an input

and the target is a matrix with a “1*t” cell array that only reflects “t” time steps of an output

element, which is the measured HI after multi regime normalization and function extraction. Both

the inputs and a target are normalized using a Min-Max normalization to map in the range of [0,1].

The hidden layers and the layers’ delays are same as the one used for C-MAPSS dataset and the

training function used is scaled conjugate gradient. During the neural network learning stage, the

raw sensor values of same trajectory are taken as an input and HI value as a target of same

trajectory. To obtain a neural network prediction, the same raw data is applied to the obtained

network function, ensuring that the trained neural network can achieve high accuracy for an output

estimate. As can be seen from the Fig. 4.10, a similar trend culminated in confirmation, but it was

contaminated with noise for which smooth data was used to filter noise and the network was able

to achieve good accuracy after 1 minute 19 seconds of training. Now to validate, the network

function is used on new inputs of test trajectories to ensure that the algorithm can memorize the

samples for the training data and correctly generalize the upcoming testing cases. Despite

differences in length and characteristic trajectory features (such as initial wear level and

deterioration pattern), the model was able to correctly describe the HI curve of the inserted data,

as shown in the Fig. 4.11.

39
Fig. 4.10: Neural Network validation with Fig. 4.11: Neural Network validation with
the same input for Challenge dataset. the different input for Challenge dataset.

40
Chapter 5

5.1 RUL Prediction

The prognostic performance of the proposed method is discussed in this chapter. The trained

network functions can be used to filter all HIs for both training and test trajectories within the same

domain once the neural network has provided a competent generalization of the input-output

relationship. The comparisons with other ensemble methods including Multi-Layer Perceptron

(MLP), Nonlinear Autoregressive Network with Exogenous Inputs (NARX) and Cascade forward

neural network (CFNN) are carried out to show the superiority of Deep Layer-Recurrent Neural

Network (DL-RNN). All the experiments are conducted on a PC with Intel Core i5 CPU, 8-GB

RAM and in the environment of MATLAB. For the FD001 dataset the plot of error and a table of

comparison of different methods are shown below.

Fig. 5.1: The error plot showing the variation of the mean,
maximum, and minimum values of DL-RNN (FD001 dataset).

41
Fig. 5.1 shows the randomly selected 15 series and plots their variation of the mean, maximum,

and minimum values of Yt − Rt over time. Where Yt and Rt is the estimated and predicted RUL

at a time point. As can be seen from the plot, the upper bound denotes the maximum error, lower

bound denotes the minimum error and line in black line denotes mean of the error. It can be clearly

seen that the error is within the acceptable range whereas for other three topologies the error plot

is beyond the range. The error plot of one topology out of three i.e., multi-layer perceptron is

shown in Fig. 5.2.

Fig. 5.2: The error plot showing the variation of the mean,
maximum, and minimum values of MLP (FD001 dataset).

For the performance evaluation, the metrics and their formula shown in the table below are used.

Table 5.1: Prognostic Metrics

Metric Formula

Mean Absolute Error 1

MAE = 𝑛 ∑𝑛𝑖=1|𝑌𝑡 − 𝑅𝑡|

1 𝑌𝑡 = Predicted RUL
Mean Square error MSE = 𝑛 ∑𝑛𝑖=1(𝑌𝑡 − 𝑅𝑡)2

42
Root Mean Square Error ∑𝑛
𝑖=1(𝑌𝑡−𝑅𝑡)
2 𝑅𝑡= True RUL
RMSE= √ 𝑛
𝑛 = 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑡𝑟𝑎𝑗𝑒𝑐𝑡𝑜𝑟𝑖𝑒𝑠
∞
Integral Square Error ISE = ∫0 (𝑌𝑡 2
− 𝑅𝑡) 𝑑𝑡

In order to check whether the training and testing is performed properly, the Mean Absolute Error

(MAE) and Root Mean Square Error (RMSE) are used for performance evaluation and is

calculated both for training and testing for all four topologies. The lower the values of MAE and

RMSE, better the performance of a network. As can be seen from the table, MAE and RMSE of

Layer-Recurrent Neural Network is lower compared to other three topologies. The performance

of Cascade Feedforward Neural Network is poor compared to all other methods. We can note from

Table 5.2 that best results were obtained using DL-RNN model.

Table 5.2: Performance comparison of different methods (FD001 dataset)

Training Testing

MAE (%) RMSE (%) MAE (%) RMSE (%)

DL-RNN 0.174 0.199 0.177 0.121

NARX 0.771 1.04 0.775 1.08
MLP 1.724 2.54 1.754 2.295
CFNN 2.32 3.20 2.77 4.54

Experiments were conducted based on testing the four topologies separately using Levenberg-

Marquardt (LM), Bayesian Regularization (BR) and Scaled Conjugate Gradient (SCG). Compared

to the LM and BR, SCG takes less time to converge, and it provides better result as compared to

other two training algorithms.

Table 5.3 shows the MSE and ISE calculation along with the number of iterations when

trained with SCG. As can be seen from the table the MSE and ISE values are lower for DL-RNN

43
compared to NARX, MLP and CFNN. The number of iterations for all four topologies is set to

18000.

Table 5.3: Results of testing (SCG algorithm, FD001 dataset)

Scaled Conjugate Gradient

Iterations MSE (%) ISE (%)

DL-RNN 18000 0.00002 0.0004

NARX 18000 0.0002 0.0038
MLP 18000 0.0814 0.0156
CFNN 18000 0.0155 0.081

For the Challenge dataset, the error plot for DL-RNN is shown below (Fig. 5.3). Similar to

C-MAPSS dataset, the error range for this dataset is also within the acceptable bounds which

proves that the accuracy of the model is good compared to other three topologies.

Fig. 5.3: The error plot showing the variation of the mean,
maximum, and minimum values of DL-RNN (Challenge dataset).

For comparison, the error plot of NARX is shown below (Fig. 5.4). As can be seen from the plot,

the minimum and maximum error is in the range of -2 and 2 whereas the minimum and maximum

44
range of DL-RNN is within -0.3 to 0.3. This clearly shows the superiority of DL-RNN model to

other three topologies.

Fig. 5.4: The error plot showing the variation of the mean,
maximum, and minimum values of NARX (Challenge dataset).

Same as C-MAPSS, the comparison table of training and testing are shown below. MSE and

RMSE are evaluated for all four topologies and are presented in Table 5.4. As can be seen, the

RMSE and MAE values of DL-RNN is lowest compared to other three. Therefore, it can be said

that for both the datasets, DL-RNN topology outperformed other three topologies in terms of

performance.

Table 5.4: Performance comparison of different methods (Challenge dataset)

Training Testing

MAE (%) RMSE (%) MAE (%) RMSE (%)

DL-RNN 0.159 0.203 0.127 0.135

NARX 0.293 0.455 0.514 0.597
MLP 1.459 1.625 1.516 1.772
CFNN 1.44 2.39 1.36 2.45

45
Table 5.5 shows the comparison of SCG algorithm for all four methods. Using SCG algorithm

the network gives better estimation result and it converge fast compared to other algorithms. The

number of iterations is set to 18000 to train the algorithm.

Table 5.5: Results of testing (SCG algorithm, Challenge dataset)

Scaled Conjugate Gradient

Iterations MSE (%) ISE (%)

DL-RNN 18000 0.00008 0.0018

NARX 18000 0.0004 0.0095
MLP 18000 0.0036 0.0804
CFNN 18000 0.0053 0.1187

The remaining useful life of 100 testing engines is estimated after successfully training the model.

The estimated RUL is compared with the actual RUL values and are shown in Fig. 5.5 below

where the blue line indicates the actual RUL, and the red line is the predicted RUL. As can be

seen, the expected values closely match the true values. Most of the prediction errors are within

appropriate bounds.

Fig. 5.5: RUL of 100 test engines

46
For better observation, the real RUL of 100 research engine instances is sorted from small to large.

Fig. 5.6 shows the sorted real RUL outcomes in 100 testing engine cases, along with the estimated

RUL. The RUL decreases as the horizontal axis value decreases. The expected value becomes

increasingly reliable as the engine degrades, as seen in Fig. 5.6.

Fig. 5.6: Sorted prediction for 100 testing engine instances.

47
Chapter 6

Conclusions and Future Work

6.1 Conclusions
Deep learning methods have gained popularity in engineering applications over the last decade,

especially in data analysis for reliability evaluation, which was previously inefficient due to the

requirement of expert expertise on the studied system, as well as the limitations of conventional

PHM techniques. Many challenges still exist for reliability-related, data-driven applications in

order to keep improving the estimation of health state parameters that can guarantee an accurate

diagnosis for systems, facilities. This thesis presents a deep learning method for estimating the

health state of complex systems using big machinery data. The framework is validated through

two different datasets and consists of a novel topology layer-recurrent neural network to train the

algorithm. The C-MAPSS dataset and the Challenge dataset are used to validate the proposed

framework through the training and testing of different models. The proposed framework shows

robustness for prognosis of remaining useful life of both datasets. For the C-MAPSS dataset, the

training and testing is performed on all 100 engines of first subset. The average result obtained

for both the metrics, MAE and RMSE were compared with other three topologies, where the

proposed DL-RNN based model outperformed all three topologies. For the Challenge dataset, error

plots of DL-RNN is presented along with three other topologies, where for the proposed topology

the error is within the acceptable range whereas for other three topologies the error is beyond the

acceptable range.

48
6.2 Future Work
The results presented and discussed above provide convincing evidence of the framework's ability

to extract features and conduct sequential analysis on raw data. The system is adaptable to a variety

of situations and produces reliable results. However, there are still some areas that could be

improved. One of the key disadvantages of the topologies used in this thesis is that the training

process is fully supervised. As a result, while the model can be directly applied to many problems,

there are still many reliability-related phenomena where the required amount of labelled data to

train a fully supervised model cannot be produced. Therefore, a next step for the proposed work

would be to introduce an unsupervised data preprocessing technique capable of creating labels for

the training data, such as clustering.

49
References

[1] L. Yao and Z. Ge, "Deep Learning of Semi supervised Process Data with Hierarchical

Extreme Learning Machine and Soft Sensor Application," IEEE Transactions on Industrial

Electronics, vol. 65, no. 2, pp. 1490-1498, Feb. 2018.

[2] O. Costilla-Reyes, P. Scully and K. B. Ozanyan, "Deep Neural Networks for Learning Spatio-

Temporal Features from Tomography Sensors," IEEE Transactions on Industrial

Electronics, vol. 65, no. 1, pp. 645-653, Jan. 2018.

[3] J. Pan, Y. Zi, J. Chen, Z. Zhou and B. Wang, "Lifting Net: A Novel Deep Learning Network

with Layer wise Feature Learning from Noisy Mechanical Data for Fault Classification,"

IEEE Transactions on Industrial Electronics, vol. 65, no. 6, pp. 4973-4982, June. 2018.

[4] Goebel K, Saha B, Saxena A, "A comparison of three data driven techniques for prognostics,"

in 62nd Meeting of the Society For Machinery Failure Prevention Technology (MFPT),

2008.

[5] K. P. Murphy, Machine Learning: A Probabilistic Perspective, vol. 4, London: The MIT

Press Cambridge, Massachusetts, 1991.

[6] Géron, A., Hands-On Machine Learning with Scikit-Learn and TensorFlow, Sebastopol:

O'Reilly Media,Inc, 2017.

[7] Y. LeCun, B. Yoshua, and H. Geoffrey, "Deep learning," Nature, vol. 521, no. 2015, pp. 436-

444, May. 2015.

[8] A. YoshuaBengio, Ian J. Goodfellow, Deep Learning, Cambridge: MIT Press, 2016.

50
[9] J. Deutsch and D. He, "Using deep learning-based approach to predict remaining useful life

of rotating components," IEEE Transactions on Systems, Man, and Cybernetics: Systems,

vol. 48, no. 1, pp. 11-20, Jan. 2018.

[10] Bishop, C.M., Neural networks for pattern recognition, New York: Oxford University Press,

Inc., 1995.

[11] Kevin Gurney, An introduction to neural networks, London and New York: UCL Press

Limited, 1997.

[12] B.G.M. Vandeginste, D.L. Massart, L.M.C. Buydens, S. De Jong, P.J. Lewi, J. Smeyers-

Verbeke, "Chapter 44 - Artificial Neural Networks, Data Handling in Science and

Technology," Elsevier, vol. 20, pp. 649-699, part 2,1998.

[13] M. Imran, S. A. Alsuhaibani, "A Neuro-Fuzzy Inference Model for Diabetic Retinopathy

Classification,” in Intelligent Data-Centric Systems, Intelligent Data Analysis for Biomedical

Applications, Dammam: Academic Press, 2019, pp. 147-172.

[14] Uhrig, R. E., "Introduction to artificial neural networks," Proceedings of IECON '95 - 21st

Annual Conference on IEEE Industrial Electronics, Orlando, FL, USA, vol. 1, pp. 33-37,

1995.

[15] Zhigang Tian, Lorna Wong, Nima Safaei, "A neural network approach for remaining useful

life prediction utilizing both failure and suspension histories," Mechanical Systems and

Signal Processing, vol. 24, no. 5, pp. 1542-1555, July. 2010.

[16] N. Budhani, C. K. Jha and S. K. Budhani, "Prediction of stock market using artificial neural

network," in International Conference of Soft Computing Techniques for Engineering and

Technology (ICSCTET), Bhimtal, India, 2014.

51
[17] Heimes, F. O., "Recurrent neural networks for remaining useful life estimation," in

International Conference on Prognostics and Health Management,, Denver, CO, USA, 2008.

[18] G. Lan, Q. Li and N. Cheng, "Remaining Useful Life Estimation of Turbofan Engine Using

LSTM Neural Networks," in IEEE CSAA Guidance, Navigation and Control Conference

(CGNCC), Xiamen, China, 2018.

[19] M. Yuan, Y. Wu and L. Lin, "Fault diagnosis and remaining useful life estimation of aero

engine using LSTM neural network," in IEEE International Conference on Aircraft Utility

Systems (AUS), Beijing, China, 2016.

[20] R. Wu and J. Ma, "An Improved LSTM Neural Network with Uncertainty to Predict

Remaining Useful Life," in CAA Symposium on Fault Detection, Supervision and Safety for

Technical Processes (SAFEPROCESS), Xiamen, China, 2019.

[21] X. Li, Q. Ding, J. Q. Sun, "Remaining Useful Life Estimation in Prognostics Using Deep

Convolution Neural Networks," Reliability Engineering and System Safety, vol. 172, pp. 1-

11, 2018.

[22] Ribeiro de Miranda, T. M. G. de Andrade Barbosa, A. G. Scolari Conceição and S. G. Soares

Alcalá, "Recurrent Neural Network Based on Statistical Recurrent Unit for Remaining Useful

Life Estimation," in 8th Brazilian Conference on Intelligent Systems (BRACIS), Salvador,

Brazil, 2019.

[23] Okoh C, Roy R, Mehnen J, et al., "Overview of Remaining Useful Life Prediction Techniques

in Through-life Engineering Services," Procedia CIRP, vol. 16, pp. 158-163, 2014.

[24] Zaidan, M. A., "Bayesian Approaches for Complex System Prognostics", PhD thesis,

University of Sheffield, 2014.

52
[25] Eker, O. F., Camci, F., and Jennions, I. K., "A similarity-based prognostics approach for

remaining useful life prediction," in 2nd European Conference of the Prognostics and Health

Management Society, Nantes, France, France, July. 2014.

[26] H. Chaoui, M. Kandidayeni, L. Boulon, S. Kelouwani, and H. Gualous, "Real-Time

Parameter Estimation of a Fuel Cell for Remaining Useful Life Assessment," IEEE

Transactions on Power Electronics, (in press), vol. 36, no. 7, pp. 7470-7479, July. 2021.

[27] A. El Mejdoubi, H. Chaoui, J. Sabor, and H. Gualous, "Remaining Useful Life Prognosis of

Supercapacitors Under Temperature and Voltage Aging Conditions," IEEE Transactions on

Industrial Electronics, vol. 65, no. 5, pp. 4357-4367, May. 2018.

[28] Jianhui Luo, M. Namburu, K. Pattipati, Liu Qiao, M. Kawamoto and S. Chigusa, "Model-

based prognostic techniques [maintenance applications]," in Proceedings AUTOTESTCON

2003. IEEE Systems Readiness Technology Conference., Anaheim, CA, 2003.

[29] B. Lamoureux, J. Massé and N. Mechbal, "Improving Aircraft Engines Prognostics and

Health Management via Anticipated Model-Based Validation of Health Indicator,"

Prognostics Journal, vol. 2, no. 1, pp. 18-38, 2014.

[30] Schwabacher, M., "A survey of data-driven prognostics," in Proceedings of the AIAA

Infotech@Aerospace Conference: American Institute for Aeronautics and Astronautics, Inc.,

Reston, VA, 2005.

[31] Schwabacher, M. and Goebel, K., “A Survey of Artificial Intelligence for Prognostics.”,

AAAI Fall Symposium: Artificial Intelligence for Prognostics., 2007.

[32] Peng, Y., Dong, M., and Zuo, M. J., "Current status of machine prognostics in condition-

based maintenance: a review," The International Journal of Advanced Manufacturing

53
 Technology, vol. 50, pp. 297-313, 2010.

[33] H. Chaoui, and C. Ibe-Ekeocha, "State of Charge and State of Health Estimation for Lithium

Batteries using Recurrent Neural Networks," IEEE Transactions on Vehicular Technology,

vol. 66, no. 10, pp. 8773-8783, October. 2017.

[34] H. Chaoui, C. Ibe-Ekeocha, and H. Gualous, "Aging Prediction and State of Charge

Estimation of a LiFePO4 Battery using Input Time-Delayed Neural Networks," Electric

Power Systems Research, Elsevier, vol. 146, pp. 189-197, May. 2017.

[35] A. Mosallam, K. Medjaher and N. Zerhouni, "Data-driven prognostic method based on

Bayesian approaches for direct remaining useful life prediction," Journal of Intelligent

Manufacturing, vol. 27, no. 5, pp. 1037-1048, 2014.

[36] D. A. Tobon-Mejia, K. Medjaher, N. Zerhouni and G. Tripot, "A Data-Driven Failure

Prognostics Method Based on Mixture of Gaussians Hidden Markov Models," IEEE

Transactions on Reliability, vol. 61, no. 2, pp. 491-503, June. 2012.

[37] Y. Song, C. Yang, T. Wang, D. Liu and Y. Peng, "Hybrid approach of iterative updating for

lithium-ion battery remaining useful life estimation," in Prognostics and System Health

Management Conference (PHM-Chengdu), Chengdu, China, 2016.

[38] B. Saha, K. Goebel, and J. Christophersen, "Comparison of prognostic algorithms for

estimating remaining useful life of batteries," Transactions of the Institute of Measurement

& Control, vol. 31, pp. 293-308, June. 2009.

[39] A. El Mejdoubi, A. Oukaour, H. Chaoui, Y. Slamani, J. Sabor, and H. Gualous, "Online

Supercapacitor Diagnosis for Electric Vehicle Applications," IEEE Transactions on

Vehicular Technology, vol. 65, no. 6, pp. 4241-4252, June. 2016.

54
[40] V. Fornlöf, D. Galar, A. Syberfeldt, T. Almgren, M. Catelani and L. Ciani, "Maintenance,

prognostics and diagnostics approaches for aircraft engines," in IEEE Metrology for

Aerospace (MetroAeroSpace), Florence, Italy, 2016.

[41] McClung, M. Enright and R., "A Probabilistic Framework for Gas Turbine Engine Materials

With Multiple Types of Anomalies," Journal of Engineering for Gas Turbines and Power,

vol. 133, no. 8, 2011.

[42] NAEEM, M, "Implications of engine deterioration for operational effectiveness of a military

aircraft," Applied Energy, vol. 60, no. 3, pp. 115-152, 1998.

[43] K. Goebel and A. Saxena , "Turbofan Engine Degradation Simulation Dataset," NASA Ames

Prognostics Data Repository, 2008.

[44] D. Frederick, DeCastro, and Litt, "User's Guide for the Commercial Modular Aero-

Propulsion System Simulation (C-MAPSS)," NASA/ARL, 2007.

[45] Saxena A, Goebel K, "PHM08 challenge dataset.," NASA Ames Prognostics Data Repository,

NASA Ames Research Center, 2008.

[46] F. Smach, "Design of a Neural Networks Classifier for Face Detection," Journal of Computer

Science, vol. 2, no. 3, pp. 257-260, 2006.

[47] B. Zhang, D. Wang, W. Song, S. Zhang and S. Lin, "An Interval-Valued Prediction Method

for Remaining Useful Life of Aero Engine," in 2020 39th Chinese Control Conference

(CCC), Shenyang, China, 2020.

[48] A. Tatli, S. Kahveciolu, "NARX neural networks-based time series prediction for amount of

airworthiness time, in electrical, electronics and biomedical engineering (Elec)," Elektronika

Ir Elektrotechnika, vol. 26, no. 5, pp. 28-32, 2020.

55
[49] H. Tang. H. Xie, Y. H. Liao, "Time series prediction based on NARX neural networks: an

advanced approach, in: machine learning and cybernetics," in 2009 International Conference

on Machine Learning and Cybernetics, Baoding, China, 2009.

[50] S. Chen, S. A. Billings, and P. M. Grant, "Non-linear system identification using neural

networks," International Journal of Control, vol. 51, no. 6, pp. 1191-1214, 1990.

[51] Parthasarathy, K. S. Narendra and K., "Identification and control of dynamical systems using

neural networks," IEEE Transactions on Neural Networks, vol. 1, no. 1, pp. 4-27, 1990.

[52] S. M. Zainorzuli, S. Afzal Che Abdullah, R. Adnan and F. A. Ruslan, "Comparative Study

of Elman Neural Network (ENN) and Neural Network Autoregressive with Exogenous Input

(NARX) For Flood Forecasting," in IEEE 9th Symposium on Computer Applications &

Industrial Electronics (ISCAIE), Malaysia, 2019.

[53] Serackis, M. Tamulionis and A., "Comparison of Multi-Layer Perceptron and Cascade Feed-

Forward Neural Network for Head-Related Transfer Function Interpolation," in Open

Conference of Electrical, Electronic and Information Sciences, Vilnius, Lithuan, 2019.

[54] B. Warsito, R. Santoso, Suparti and H. Yasin, "Cascade Forward Neural Network for Time

Series Prediction," in Journal of Physics: Conference Series, Semarang, Indonesia, 2017.

[55] Diallo, O. N., "A data analytics approach to gas turbine prognostics and health

management.," 2010.

[56] C. Zhang, P. Lim, A. K. Qin, K. C. Tan., "Multiobjective Deep Belief Networks Ensemble

for Remaining Useful Life Estimation in Prognostics," IEEE Transactions on Neural

Networks and Learning Systems, vol. 28, no. 10, pp. 1-13, Oct. 2017.

[57] P. Lim, C. K. Goh, K. C. Tan, "A Time Window Neural Network Based Framework for

56
 Remaining Useful Life Estimation," in 2016 International Joint Conference on Neural

Networks (IJCNN), Vancouver, BC, Canada, 2016.

[58] S. Zheng, k. Ristovski, A. Farahat, C. Gupta, "Long Short-Term Memory Network for

Remaining Useful Life Estimation," in 2017 IEEE International Conference on Prognostics

and Health Management (ICPHM), Dallas, TX, USA, 2017.

[59] T. Wang, Jianbo Yu, D. Siegel and J. Lee, "A similarity-based prognostics approach for

Remaining Useful Life estimation of engineered systems," in 2008 International Conference

on Prognostics and Health Management, Denver, CO, USA, 2008.

[60] Coble, J. and Hines, J. W., "Identifying optimal prognostic parameters from data: a genetic

algorithms approach," in In Annual conference of the prognostics and health management

society, San diego, September. 2009.

[61] A. Saxena, K. Goebel, D. Simon and N. Eklund, "Damage propagation modeling for aircraft

engine run-to-failure simulation," in 2008 International Conference on Prognostics and

Health Management, Denver, 2008.

[62] G. Li, Z. Yan and J. Wang, "A one-layer recurrent neural network for constrained

nonsmoothed invex optimization," Neural networks : the official journal of the International

Neural Network Society, vol. 50, pp. 79-89, 2014.

[63] X. Zhang, Y. Dong, L. Wen, F. Lu and W. Li, "Remaining Useful Life Estimation Based on

a New Convolutional and Recurrent Neural Network," in 2019 IEEE 15th International

Conference on Automation Science and Engineering (CASE), Vancouver, BC, Canada, 2019.

[64] M. Kobayashi, "Noise Robust Projection Rule for Hyperbolic Hopfield Neural Networks,"

IEEE Transactions on Neural Networks and Learning Systems, vol. 31, no. 1, pp. 352-356,

57
 Jan. 2020.

[65] Guanghua Ren, Yuting Cao, Shiping Wen, Tingwen Huang, Zhigang Zeng, "A modified

Elman neural network with a new learning rate scheme," Neurocomputing, vol. 286, pp. 11-

18, 2018.

[66] H. Turabieh, A. A. Salem, N. Abu-El-Rub, "Dynamic L-RNN recovery of missing data in

IoMT applications," Future Generation Computer Systems, vol. 89, pp. 573-583, Dec. 2018.

58