10th
International Symposium on NDT in Aerospace
1 License: https://creativecommons.org/licenses/by-nd/4.0/
Machine Learning based Data Driven Diagnostics &
Prognostics Framework for Aircraft Predictive
Maintenance Partha ADHIKARI
1, Harsha GURURAJA RAO
1, Dipl.-Ing. Matthias BUDERATH
2
1Airbus India, Bangalore, 560 048, India;
E-mail: [email protected], [email protected] 2Airbus Defence and Space, Willy-Messerschmitt-Strasse 81663 Munich,Germany
E-mail: [email protected]
Abstract. Recent trends of Digitalization across different industries have led to
generation of massive amounts of data. As a natural consequence, there is a surge in
Advanced Machine Learning techniques being applied to this Big Data.
Simultaneously, there has been growing needs of increased Operational Reliability,
driving down Maintenance costs and increase in Safety, due to which Predictive
Maintenance is quickly becoming the most important strategy across many
industries, especially in Aerospace. As newer Aircrafts are being equipped with
more sensors, Machine Learning based Diagnostics and Prognostics techniques are
becoming increasingly popular compared to conventional approaches in developing
Predictive Maintenance solutions. Building a ML based Diagnostics & Prognostics
Model requires massive amount of Run-to-Fail Sensor data, but the opportunity to
capture this in-service failure related data is very limited in highly reliable & safety
critical aircraft platforms in comparison to other domains. To address this challenge
of lack of adequate & appropriate in-service failure data, Airbus DS has developed a
Simulation Framework in the roadmap of Technology development of ISHM and
Predictive Maintenance. To accelerate the development of Predictive Maintenance
solutions for various aircraft systems, a Data Driven Diagnostics & Prognostics
Framework has been developed. This paper outlines this unique framework and its
validation using the data generated from the ISHM Simulation Framework.
Keywords: Machine Learning, Data Analytics, Diagnostics, Prognostics, Predictive Maintenance,
Simulation, Aircraft Systems, ISHM
1 Introduction
Globally, aircraft operators as well as OEMs are facing major challenges in enhancing,
even meeting specified Mission Success Rate (MSR) and Operational Support Cost targets
due to unscheduled Maintenance leading to Operational Interruptions (OI). Lack of
advanced warning of failure events, fault isolation / troubleshooting overhead and lack of
integrated and optimized planning of fleet, maintenance and logistics are the key
contributors leading to OI. So, Predictive maintenance / ISHM and related technologies are
generally seen as potential strategies to address these problems.
Predictive maintenance provides an integrated solution of assessing present & future
health (diagnostics and prognostics) of complete system and derives planning of operation,
maintenance and logistics. With the rapid growth of technologies such as Internet of Things
(IoT), Machine Learning (ML), Artificial Intelligence (AI) and Big Data technology which
are inherent in Industry 4.0, Predictive maintenance is becoming smarter day-by-day. Thus,
Data-driven and hybrid diagnostics and prognostics solutions are becoming more popular
2
compared to purely model based diagnostics and prognostics. In order to keep up with the
Digital transformation across the enterprise and to speed up the development of integrated
Predictive maintenance solutions across various aircraft systems, a Machine Learning based
Diagnostics and Prognostics (DnP) framework has been developed.
Given the highly reliable and safety critical nature of aircraft systems, it is very
challenging to find sufficient amount of run to fail in-service data. Run-to-fail data is
essential to train purely data-driven predictive model for calculating health grade and
estimating Remaining Useful Life (RUL). Maturity and validation of ML based
Diagnostics and Prognostics framework is supported by simulation strategy realized
through digital twin of integrated predictive maintenance in Airbus DS.
Pundarikaksha Baruah, et. al, 2006 [1], has presented an innovative on-line
approach for autonomous diagnostics and prognostics by developing a “generic”
framework that is relatively independent of the type of physical equipment under
consideration. The framework is based on unsupervised machine learning methods. Honglei
Li, et. al, 2014 [2] presents a diagnostics and prognostics framework tested on data from a
high fidelity equivalent circuit system simulation and empirical component degradation
models. It adapts proven prognostics and health management techniques from machinery to
electronic health management. Hamed Khorasgani, et. al, 2018 [3], claims that both data-
driven and model-based diagnosis methods follow similar procedure and can by brought
under one unifying framework. Their proposed framework for fault detection and isolation
can integrate methods from both data-driven and model based techniques. Zhe Li, 2018 [4],
suggests deep learning based framework for predictive maintenance.
The ML based Diagnostics and Prognostics (DnP) framework discussed within the
scope of this paper (publication) is being developed keeping in mind several different
applications and requirements of various stakeholders. One of them being an array of
options available for conducting advanced ML based techniques for various benchmarking
studies. Since the early stages of development, the framework has been consciously
designed to be generic in nature, to minimize the efforts needed to customize it to different
systems. The DnP framework is based on Python, a popular programming language for
Data Science applications, providing easy deployment options to the Big Data ecosystem.
The subsequent sections describe the proposed ML based Diagnostics and
Prognostics (DnP) framework in detail. Section 2, describes the simulation strategy for
diagnostics and prognostics development. It is followed by section 3, which delves deeper
in to the enablers for diagnostics and prognostics. Section 4 presents the proposed
framework and subsections detail various algorithms incorporated in the diagnostics and
prognostics framework. The framework is applied on NASA engine data and on the
simulation of Electrical and Landing Gear systems within Airbus DS as use-cases and
results are described in section 5.
2 Simulation Strategy for ML based Diagnostics & Prognostics Development
Airbus DS has developed a comprehensive integrated ISHM simulation framework which
contributes in integrated demonstration of Proof of Enablers (PoE), training & maturity of
diagnostics & prognostics algorithm, maturation of ISHM requirements (KPIs) and V&V
of predictive maintenance (ISHM) functions. For end-to-end demonstration of ISHM,
simulation framework hosts simulation of aircraft system with fault injection provision, on-
board health assessment function, off-board analytics related to diagnostics & prognostics,
fleet planning, maintenance/logistics planning, etc. in enterprise level. ISHM instead of
IVHM is used because Health Monitoring and Management is considered as System of
Systems approach.
3
There are three broad categories of condition based diagnostics & prognostics
approaches: Data based, Model based and Hybrid. To develop data driven solution, ML
based Diagnostics & Prognostics (DnP) Framework has been developed. DnP framework
receives data from Big Data environment which is ingested with sensor and BITE data
generated by ISHM simulation framework under different fault scenarios. DnP framework
contains various steps of diagnostics & prognostics algorithms with having provision of
selecting different options for benchmarking study. Initially, study in DnP framework
provides feedback on model fidelity and data generation to ISHM simulation framework.
Key objective of DnP framework is analysing & benchmarking various types of diagnostics
& prognostics algorithms and providing selected trained algorithms for integration and
demonstration of data driven diagnostics & prognostics algorithm in ISHM simulation
framework and in-service deployment (Fig 1).
Fig. 1. ML Based DnP Framework as part of ISHM Simulation Framework
3 Enablers for Diagnostics & Prognostics
Diagnostics & Prognostics are key technology building blocks for predictive maintenance.
Fig. 2 depicts different types of diagnostics & prognostics techniques. There exist the
following three categories of diagnostics and prognostics.
Type I: Reliability Data-based which characterizes the expected lifetime of
the average component operating in historically average conditions (e.g. Weibull Analysis).
Type II: Stress-based which estimates the lifetime of the average component in
a specific environment (e.g. Proportional Hazards Model). Type III: Condition-based which
estimates the lifetime of the specific component in its specific operating environment.
Condition-based maintenance typically falls into three major categories viz. data-
driven, model-based and hybrid. Data-driven Techniques often address only anticipated
fault conditions, where a fault ‘‘model’’ is a construct or a collection of constructs such as
neural networks, expert systems, etc. that must be trained first with known prototype fault
patterns (data) and then employed online to detect and determine the faulty component’s
identity. Model-based Techniques, on the other hand, rely on an accurate dynamic model of
the system and are capable of detecting even unanticipated faults. They take advantage of
the actual system and model outputs to generate a ‘‘discrepancy’’ or residual between the
two outputs that is indicative of a potential fault condition. Hybrid Techniques use
knowledge about the physical process and information from observed data together to
derive advantage of both model based and data driven methods.
4
Detailed enablers for data-driven diagnostics are shown in Fig. 2 and sections to
follow, which constitute the proposed ML based diagnostics and prognostics framework.
Fig. 2.Broad categories of Diagnostics & Prognostic techniques and different steps in Data Driven
Diagnostics & Prognostics
4 ML Based Diagnostics and Prognostics (DnP) Framework
DnP Framework has mainly three functions: data pre-processing & feature engineering,
diagnostics and prognostics, each with the provision of selecting techniques from a variety
of different options. Following sub-sections summarize the landscape of various enablers
for different steps of diagnostics & prognostics.
4.1 Data Pre-processing & Feature Engineering
The accuracy in ML based diagnostics and prognostics are primarily driven by the
appropriate data pre-processing, data exploration and feature engineering which are to be
systematically followed.
Data exploration starts with identification of predictor and target variables. In the
context of predictive maintenance, predictor represents features whereas target represents
the health grade or failure modes. Subsequent steps in data exploration are univariate
analysis, bi-variate analysis, missing value treatment, outlier treatment and regime
identification. Flight regime identification in data exploration stage for predictive
maintenance may be necessary if failure propagation happens differently based on different
flight regime. Flight regime are calculated by clustering operating conditions like take-off,
cruise, landing, etc. Statistics of sensor data under different flight regimes are to be
calculated and sensor data segregated in terms of flight regimes.
Next comes feature engineering (Fig. 3) which is a science (or art) of extracting
more information from existing data. Original sensor data are primary features but in many
situations generation of new derived features (FFT, Kurtosis, mean, etc.) may enhance
failure observability of the model. Subsequent steps are feature benchmarking or dimension
reduction.
5
Fig. 3. Feature engineering overview
4.2 Anomaly Detection
In Diagnostics phase, the first important step is to detect the presence of an incipient failure
or point of transition from normal state to degraded state within a sub-system component.
Conventionally, this phase is referred as Anomaly Detection (AD). Along with Predictive
Maintenance, there are various use cases for anomaly detection in areas of fraud
prevention, fault detection, and monitoring across many industries such as finance, IT,
security, medical, energy, e-commerce, and social media. Naturally, significant amount of
research is ongoing, to develop advanced methods of Anomaly Detection.
4.2.1 Survey of state-or-art Anomaly Detection Techniques
Anomaly detection started out using statistical methods to detect outliers with “simple
threshold” which was then improved to “dynamic thresholds” based on standard deviations,
moving averages, comparison to a least squares model of the data, statistical comparisons
of very recent data to older data, and histogram analysis. Lately the statistical approaches
have been expanded by Machine Learning methods. Based on the way of training, machine
learning based anomaly detection can be broadly classified in three categories: supervised,
semi-supervised and unsupervised.
Supervised Anomaly Detection describes the setup where the data comprises of
fully labelled training and test datasets. An ordinary classifier can be trained first and
applied afterwards. Semi-supervised Anomaly Detection also uses training and test
datasets, whereas training data only consists of normal data without any anomalies. The
basic idea is that a model of the normal class is learned and anomalies can be detected
afterwards by deviating from that model. Unsupervised Anomaly Detection is the most
flexible setup which does not require any labels. Furthermore, there is also no distinction
between training and test dataset
Unsupervised anomaly detection algorithms can be roughly categorized into the
following main groups (Chandola V et al. 2009 [5]): Nearest-neighbour based techniques,
Clustering based, Statistical algorithms, Subspace techniques, Unsupervised Neural
Network, real time and time series analysis based algorithms. Fig. 4 summarizes the
different categories of anomaly detection.
6
Fig. 4. Enablers for state of art anomaly detection
4.2.2 Anomaly Detection in the context of Predictive Maintenance
There are three types of anomalies: point anomalies, sequential anomalies and contextual
anomalies. If a single point deviates from the considered normal pattern it is referred to as a
point anomaly. If a sequence or collection of points is anomalous with respect to the rest of
the data, but not the points themselves, it is referred to as a sequential or collective
anomaly. If a point or a sequence of points is considered as an anomaly with respect to its
local neighbourhood, but not otherwise, it is referred to as a contextual anomaly. In
Predictive maintenance, sequential and contextual anomaly is applicable as it deals with
sensor data under healthy stage of any asset to gradually degraded stages under different
flight regimes.
Almost all types of algorithms as mentioned are applicable for predictive
maintenance. Ramasso1 E., & Saxena A (2014) [6] surveyed extensively on different
technique of anomaly detection related to predictive maintenance. Survey is summarized in
the following table.
Table 1. Survey papers for Anomaly Detection
Category of Anomaly
Detection References
Supervised
Ramasso E. 2009 [7], Ramasso E. et. al. 2010 [8], Ramasso E. et. al.
2013 [9], Sarkar S. et.al 2011 [10], Xue Y. et al. 2011 [11], Zhao D. P.
et al 2011 [12], Yu J 2013 [13], Lin Y. et al. 2013 [14], Tamilselvan P.
et al. 2013 [15]
Semi-Supervised Ramasso E. et al. 2013 [16], Ramasso E. et al 2013 [17], Alestra S. et al
[18]
Unsupervised Wang T. 2010 [19], Wang T. et al 2008 [20], Sarkar S. et al 2011 [10],
Peng Y. et al 2012 [21], Serir L et al 2012 [22], Javed K. et al 2013
[23]
7
4.2.3 Time-series / Real-time Anomaly detection
Appropriate pre-processing, feature extraction and suitable customization of algorithms are
needed to address time series data for predictive maintenance. Holt-Winters, Auto-
Regressive Integrated Moving Average (ARIMA), and Support Vector Regression (SVR)
are examples of statistical models that take into account seasonality. These models can be
used as a time series dynamic threshold for anomaly detection. Alestra et al. used one class
SVM for detection of degradation. For the modeling of multidimensional degradation
behaviour i.e. degradation trending in time Autoregressive Integrated Moving Average
(ARIMA) model is employed.
In recent years, for anomaly detection of time series data, deep learning Akash Sing,
2017 [24] has emerged as one of the most popular machine learning techniques, yielding
state-of-the-art results for a range of supervised and unsupervised tasks due to its ability to
learn high-level representations relevant for the task at hand. These representations are
learned automatically from data with little or no need of manual feature engineering and
domain expertise. For sequential and temporal data, LSTMs, RNNs have become the deep
learning models of choice because of their ability to learn long-range patterns.
The ability to perform predictive analytics in real-time on streaming data is a game
changer. Works Numenta, 2015 [25], Aggarwal C. et al, 2012 [26], Aggarwal C. C., 2005
[27], Aggarwal C. C. et al, 2011 [28] address application of real time anomaly detection.
The most exciting technique is developed by Numenta is Hierarchical Temporal Memory
(HTM) [25] which is a biologically inspired machine intelligence technology that mimics
the architecture and processes of the neocortex. HTM is a promising candidate for real-time
predictive maintenance.
4.3 Fault Identification & Prognostics
After an anomaly is detected, estimation of the degree of degradation is the next natural
step. This is called Fault Identification which is followed by Prognostics. The main
objective of prognostics is to provide an estimation of the Remaining Useful Life (RUL) of
a degrading component or system. Anomaly detection and fault isolation are generally
classification or clustering problem whereas Fault Identification & Prognostics are
regression problems. Various techniques in Fault Identification and Prognostics are
summarized in Fig. 5. Broadly there are three categories of prognostics techniques:
similarity based, extrapolation based, and, model identification & estimation based. Table 2
lists some references for each broad category.
Fig. 5. Enablers for Fault Identification & Prognostics
8
Table 2. List of references for different prognostics methods
Category of Fault
Identification &
Prognostics
References Remarks
Similarity based
Wang T. et al. 2008 [20], Wang
T. 2010 [19], Peng et al. 2012
[21], Ramasso et al. [16],
Ramasso E. (2014) [29]
[20] [19] →HI-based 3 similarity measures
and kernel smoothing
[21] → Similar to 36 and 37 using 1
similarity measure
[16] → Feature-based similarity, 1
similarity measure, ensemble, degradation
levels classification
[29] → HI-based similarity, polygon
coverage similarity, ensemble
Extrapolation based
– Statistical Regression
Hu C. et al. 2012 [30], Coble J.
2010 [31], Coble & Hines, 2011
[32], Siegel 2009 [33], T. Wang
2010 [19], P. Wang et al. 2012
[34], Juan F. L et al. 2017 [35],
Hu et al., 2012 [30], Liu K. et al.
2013 [36]
[31] [32] → Quadratic fit, Bayesian
updating,
[33] → Logistic Regression,
[19] → Karnel regression, RVM
[34] →RVM
[35]→ GPR
[30]→ RVM, SVM, RNN, Exponential and
quadratic fit, Bayesian updating
[36] → Exponential fit
Extrapolation based
– Neural Network
Riad A. et al. 2010 [37], Jie Liu et
al. [38], Marco R. et al. 2016 [39]
[37] → Application of Adaptive RNN
[39] → ESN
Extrapolation based
– Deep Learning
Gugulothu N. et al. 2017 [40],
Jason Deutsch et al. 2017 [41]
[41] → Embed-RUL is used here.
Model Identification Alestra S. et al. [18] [18] → One class SVM, ARIMA
State Estimation Bhaskar Saha et al. 2008 [42] [42] → RVM, fitting exponential growth
model, Particle Filter
In similarity based method, historical instances of trajectories of different features
of all possible failures are used to create a library. For a given test instance similarity with
instances in library is evaluated to generate a set of RUL estimates. Based on how the
notion of similarity is defined and quantified there are three types of models [43]: Pairwise
similarity model, Hash similarity model & Residual similarity model. Pairwise similarity
model computes the distance between different time series, where distance is defined as
correlation. Hash-similarity model transforms historical degradation data from each
member of ensemble into fixed-size condensed information such as the mean, total power,
maximum or minimum values, or other quantities. Residual-based estimation fits prior data
to model such as an ARMA model or a model that is linear or exponential in usage time. It
then computes the residuals between predicted model data and test data.
Unlike the similarity-based approach, the extrapolation-based approach employs the
training data set not for the comparison with the testing data set but rather for obtaining
prior distributions of the degradation model parameters. The testing data set is then used to
update these prior distributions. An RUL estimate can be obtained by extrapolating the
updated degradation model to a predefined failure threshold. Degradation models can be
constructed using statistical regression, neural network and deep learning approach.
RUL can be estimated using dynamic model identification techniques like AR,
ARX, ARMAX, ARIMA, etc. which are trained (i.e. model coefficients are estimated)
using time series data [43]. Using time series model, features or parameters are predicted to
compute RUL. There are also recursive estimators fitting models in real-time and process
the data, such as recursive ARX and recursive AR. RUL estimation with state estimators
such as Unscented Kalman Filter (UKF), Extended Kalman Filter (EKF), and Particle Filter
(PF) works in a similar way. One performs state estimation on some time-varying data, and
predict future state values to determine the time until some state value exceeds some
specified limit or threshold indicating presence of failure.
9
4.4 Enhanced Fault Isolation
ML based component level fault isolation in early stage of degradation is carried out by
supervised classification. In most cases residuals computed from sensor measurement and
estimated measurement using physics based model are considered useful features along
with other features e.g. sensor measurements and derived features from sensors. Health
aggregation from subsystem component to subsystem, system and aircraft level is integral
part of diagnostics. Lopez B. et al 2016 [44] implemented component level fault isolation
and health aggregation using Bayesian network based HLR framework.
5 Use cases to validate D&P Framework
The proposed DnP framework shows a series of systematic steps to implement diagnostics
and prognostics of different aircraft systems. Initially, this framework is used to implement
diagnostics & prognostics of Landing Gear and Electrical systems. This in-turn helps to
validate and mature the DnP framework. Turbofan engine simulated data from NASA [45]
is used as third use case to benchmark the algorithms of the framework due to the
popularity of this dataset which is mentioned in more than 100 publications related to
diverse varieties of Diagnostics & Prognostics solutions. Details of fault and sensor
measurements are mentioned for respective use-cases as shown in the Fig. 6.
Fig. 6. Use-cases for DnP Framework
Various steps for use-case study driven by the DnP framework are shown in fig. 7.
Feature engineering study for all systems are provided in Table 3. For Anomaly detection,
One class SVM is chosen to apply for all use-cases, where model is trained using only
selected features in normal operation. For fault isolation bench-marking study was carried
out and K-NN classifier was selected for Engine whereas SVM was chosen for Electrical
system. Different degradation levels (Healthy, Sub-healthy, Degraded, Fail) were obtained
using K-mean clustering of PCA components of selected features. Time series health index
for all training instances was computed using Euclidean distances from fail centroid. Then
degradation model was generated using RVR. The degradation trajectory of testing data is
compared with that of stored training data and the trajectory with closest match are selected
for RUL calculation [46]. This Similarity based approach is used for Engine based on
benchmarking study. For Landing Gear, ARIMA model was sufficient to predict RUL.
Selected results are provided in Fig. 8 to 11 and Table 3, 4 and 5.
10
Fig. 7. Various sequential steps for use-case study driven by DnP framework
Table 3. Feature Engineering results
Landing Gear Electrical Engine
No. of original
features (sensor) 9 6 24
New features 2 (Settling time, damping
ratio)
66 (FFT – 36, Kurtosis – 6, Peak to
Peak – 6, Lyapunov Exponent – 6,
Entropy – 6, Correlation – 6)
24 (moving
average)
Features after
benchmarking
3 (Settling time, damping
ratio, shock absorber height) 12 (Peak to Peak – 2, FFT – 10)
12, 8 PCA
components
Fig. 8. Univariate analysis of Shock Absorber of Landing Gear
Fig. 9. Anomaly Detection of Landing Gear : degradation of shock absorber due to Oil leakage.
Original - (left), Zoomed (right)
11
Table 4. Benchmarking for fault classification of Electrical System
Sr. No. Algorithm Name Accuracy Precision Recall Training Time (S)
1 Decision Tree 0.99922 0.99922 0.99922 2.80
2 Naive Bayes 0.90429 0.90429 0.90429 0.41
3 SVM 1 1 1 8.54
4 Extra Tree 1 1 1 1.87
5 Random Forest 1 1 0.99929 7.89
6 ADABoost 0.99872 0.99872 0.99872 11.83
7 Logistic 1 1 1 6.54
8 Linear SVC 1 1 1 54.21
Fig. 10. Cluster features with health levels : Engine - (left)
Propagation of health index with cycle number : Engine - (right)
Fig. 11. Health degraded model with fitted RVR : Engine (left), Matched degraded pattern and RUL (right)
Table 5. Benchmarking of various algorithms for RUL calculation: Engine
12
6. Conclusion
Extensive survey has been done to find the state-of-the-art techniques of diagnostics and
prognostics and selected algorithms are implemented in this ML based Diagnostics and
Prognostics framework. Key highlights of this framework include its generic nature,
options for benchmarking and ability to be customized for specific systems. This has been
validated and matured with use-cases of diverse systems. However, this framework will
undergo continuous maturity with experience gained in Predictive maintenance across
different aircraft systems.
7 Acknowledgments
The authors are grateful to Gaurav Kumar, Elijah Toppo, Bibhudatta Patnaik, Manish
Patwari, Pandiyaraj Subramani, Ajith Shenoy and Amey Muley for their support. We are
also grateful to reviewers for many of the improvements to the document.
Nomenclature
ABOD Angle Based Outlier Detection
AC Advisory Circular
AD Anomaly Detection
AI Artificial Intelligence
aLOCI Approximate Local Correlation
Integral
AMC Acceptable Means of Compliance
ARMAX Auto-Regressive Integrated Moving
Average
ARIMA Auto-Regressive Integrated Moving
Average
ARP Aerospace Recommended Practice
ARTMAP Adaptive Resonance Theory
Mapping
AWR Airworthiness Report
BIT Built-in Test
BITE Built-in Test Equipment
BN Bayesian Network
uCBLOF Cluster-Based Local Outlier Factor
CBM Condition based maintenance
CC Certification Coordinator
CMGOS Clustering-based Multivariate
Gaussian Outlier Score
COF Connectivity-Based Outlier Factor
CS Certification Specification
DBSCAN Density-Based Spatial Clustering of
Applications with Noise
DnP Diagnostics and Prognostics
DT Decision Tree
EAI Enterprise Application Integration
EKF Extended Kalman Filter
ELM Extreme Learning Machine
EMD Empirical Mode Decomposition
ESN Echo state network
EWPCA Exponentially Weighted Principal
Component Analysis
FFT Fast Fourier Transform
FHA Functional Hazard Analysis
IVHM Integrated Vehicle Heath Monitoring
K-NN k – Nearest Neighbour
KPI Key Performance Index
LDA Linear Discriminant Analysis
LDBSCAN Local-Density-Based Spatial
Clustering of Applications with Noise
LDCOF Local Density Cluster-based Outlier
Factor
LOCI Local Correlation Integral
LOF Local Outlier Factor
LoOP Local Outlier Probability
LRU Line Replaceable Unit
LSTM Long Short Term Memory
LSTM-ED Long Short Term Memory Encoder
Decoder
MAD Mean Absolute Deviation
MAE Mean Absolue Error
MAPE Mean Absolute Percentage Error
ME Mean Error
ML Machine Learning
MLP Multi-Layer Perceptron
MSE Mean Square Error
NFLO Influenced Outlierness
NLARX Non-linear Auto Regressive
Exogenous
NLDR Nonlinear dimensionality reduction
NN Neural Network
OCSVM One Class SVM
OEM Original Equipment Manufacturer
OI Operational Interruptions
OPCAAD Over-sampling PCA on-line anamoly
detection
OR Operational Reliability
ORA Operational Risk Assessment
OSA Open System Architecture
PCA Principal Component Analysis
PF Particle Filter
PHM Prognostic Health Management
13
FMECA Failure Modes, Effects, and
Criticality Analysis
FP/FN False Positive / Negative
GMM Gaussian Mixture Model
GPR Gaussian Process Regression
GRU Gated Recurrent Units
GUI Graphical User Interface
HBOS Histogram-based Outlier Score
HILS Hardware in Loop Simulation
HLR High Level Reasoning
HTM Hierarchical Temporal Memory
HUMS Heath Usage Monitoring System
IA Integrity Assessment
ICA Instruction for Continued
Airworthiness
INFLO Influenced Outlierness
IoT Internet of Things
ISHM Integrated System Health
Monitoring
PoE Proof of Enablers
RCM Reliability Centered Maintenance
RMS Root mean square
RNN Recurrent Neural Network
RNN-ED Recurrent Neural Network Encoder
Decoder
RPCA Real Time PCA
RUL Remaining Useful Life
RVM Relevance Vector Machine
RVR Relevance Vector Regression
SHM Structural Health Monitoring
SOM Self-organizing Map
STFT Short time fourier transform
SVM Support Vector Machine
SVR Support Vector Regression
TRL Technology Readiness Level
UKF Unscented Kalman Filter
V&V Verification and Validation
References
[1] P. Baruah, R. B. Chinnam and D. Filev, “An Autonomous Diagnostics and Prognostics Framework for
Condition-Based Maintenance” in International Joint Conference on Neural Networks Sheraton
Vancouver Wall Centre Hotel, BC, 2006.
[2] Honglei Li, Xuerong Ye, Cen Chen and G. Vachtsevanos, A Framework for Model-Based Diagnostics
and Prognostics of Switched-Mode Power Supplies, ANNUAL CONFERENCE OF THE
PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY, 2014.
[3] H. Khorasgani, Ahmed Farahat, Kosta Ristovski, C. Gupta and Gautam Biswas, Framework for Unifying
Model-based and Data-driven Fault Diagnosis, ANNUAL CONFERENCE OF THE PROGNOSTICS
AND HEALTH MANAGEMENT SOCIETY, 2018.
[4] Z. Li, “Deep Learning Driven Approaches for Predictive Maintenance: A Framework of Intelligent Fault
Diagnosis and Prognosis in the Industry 4.0 Era,” 2018.
[5] V. Chandala, A. Banerjee and V. Kumar, “Anomaly Detection: A Survey, ACM Computing Surveys”
University of Minnesota, 2009.
[6] E. Ramasso and A. Saxena, “Performance Benchmarking and Analysis of Prognostic Methods for
CMAPSS Datasets” International Journal of Prognostics and Health Management, vol. 5, pp. 1-15,
2014.
[7] E. Ramasso, “Contribution of belief functions to hidden Markov models with an application to fault
diagnosis” Machine Learning for Signal Processing, vol. 2009, 2009.
[8] E. Ramasso and R. Gouriveau, “Prognostics in switching systems: Evidential Markovian classification of
real-time neuro-fuzzy predictions” Prognostics and Health Management Conference, vol. 2010, 2010.
[9] E. Ramasso and R. Gouriveau, “RUL estimation by classification of predictions: an approach based on a
neuro-fuzzy system and theory of belief functions” IEEE Transactions on Reliability, Accepted, 2013.
[10] S. Sarkar, X. Jin and A. Ray, “Data-driven fault detection in aircraft engines with noisy sensor
measurements” Journal of Engineering for Gas Turbines and Power, vol. 133, p. 8, 2011.
[11] Y. Xue, D. P. Williams and H. Qiu, “Classification with imperfect labels for fault prediction”
Proceedings of the first international workshop on data mining for service and maintenance, 2011.
[12] D. Zhao, R. Georgescu and P. Willett, “Comparison of data reduction techniques based on SVM
classifier and SVR performance” in Signal and Data Processing of Small Targets 2011, 2011.
[13] J. Yu, “A nonlinear probabilistic method and contribution analysis for machine condition monitoring”
Mechanical Systems and Signal Processing, vol. 37, pp. 293-314, 2013.
[14] Y. Lin, M. Chen and D. Zhou, “Online probabilistic operational safety assessment of multi-mode
engineering systems using Bayesian methods” Reliability Engineering & System Safety, vol. 119, pp.
150-157, 2013.
[15] P. Tamilselvan and P. Wang, “Failure diagnosis using deep belief learning based health state
14
classification” Reliability Engineering & System Safety, vol. 115, pp. 124-135, 2013.
[16] E. Ramasso, M. Rombaut and N. Zerhouni, “Joint prediction of observations and states in time-series
based on belief functions” IEEE Transactions on Systems, Man and Cybernetics - Part B: Cybernetics,
vol. 43, pp. 37-50, 2013.
[17] E. Ramasso and T. Denoeux, “Making use of partial knowledge about hidden states in hidden Markov
models: an approach based on belief functions” IEEE Transactions on Fuzzy Systems, vol. 10, p. 1109,
2013.
[18] S. Alestra, C. Bordry, C. Brand, E. Burnaev, P. Erofeev, A. Papanov and C. Silveira-Freixo, “Rare Event
Aanicipation and Degradtion Trending for Aircraft Predictive Maintenance?” in 11th World Congress on
Computational Mechanics (WCCM XI), 6th European Conference on Computational Fluid Dynamics
(ECFD VI.
[19] T. Wang, “Trajectory similarity based prediction for remaining useful life estimation” Diss. University of
Cincinnat, 2010.
[20] T. Wang and others, “A similarity-based prognostics approach for remaining useful life estimation of
engineered systems” Prognostics and Health Management, vol. 2008, 2008.
[21] Y. Peng, H. Wang, J. Wang, D. Liu and X. Peng, A modified echo state network based remaining useful
life estimation approach, In Ieee phm conference, 2012.
[22] L. Serir, E. Ramasso and N. Zerhouni, An evidential evolving multimodeling approach for systems
behavior prediction, In Annual conference of the phm society, 2012.
[23] K. Javed, R. Gouriveau and N. Zerhouni, Novel failure prognostics approach with dynamic thresholds
for machine degradation, In Ieee industrial electronics conference, 2013.
[24] A. Sing, Anomaly Detection for Temporal Data using Long Short-Term Memory (LSTM),
STOCKHOLM, SWEDEN: Information and Communication Technology, 2017.
[25] Numenta, “The science of Anomaly Detection - How HTM Enables Anomaly Detection in Streaming
Data” White Paper, 2015.
[26] C. C. Aggarwal and K. Subbian, “Event detection in social streams” in Proceedings of the 2012 SIAM
international conference on data mining, 2012.
[27] C. C. Aggarwal, “On abnormality detection in spuriously populated data streams” in Proceedings of the
2005 siam international conference on data mining, 2005.
[28] C. C. Aggarwal, Y. Zhao and S. Y. Philip, “Outlier detection in graph streams” in 27th International
Conference on Data Engineering (ICDE), 2011.
[29] E. Ramasso, “Investigating computational geometry for failure prognostics in presence of imprecise
health indicator: Results and comparisons on c-mapss datasets” PHME, vol. 14, 2014.
[30] C. Hu, B. D. Youn, P. Wang and J. T. Yoon, “Ensemble of data-driven prognostic algorithms for robust
prediction of remaining useful life,” Reliability Engineering & System Safety, vol. 103, pp. 120-135,
2012.
[31] J. Coble, Merging data sources to predict remaining useful life - an automated method to identify
prognostic parameters, University of Tennessee, Knoxville: Unpublished doctoral dissertation, 2010.
[32] J. Coble and W. Hines, “Applying the general path model to estimation of remaining useful life”
International Journal of Prognostics and Health Management, vol. 2, pp. 1-13, 2011.
[33] D. Siegel, Evaluation of health assessment techniques for rotating machinery (Doctoral dissertation,
University of Cincinnati, 2009.
[34] P. Wang, B. Youn and C. Hu, “Generic probabilistic framework for structural health prognostics and
uncertainty management” Mechanical Systems and Signal Processing, vol. 28, p. 622, 2012.
[35] J. F. L_pez-L_pez, J. J. Cuesta-Ramirez and H. F. Quintero-Riaza, “Gaussian processes in ball bearing
prognostics,” Dyna, vol. 84, pp. 214-219, 2017.
[36] K. Liu, N. Z. Gebraeel and J. Shi, “A data-level fusion model for developing composite health indices
for degradation modeling and prognostic analysis” IEEE Transactions on Automation Science and
Engineering, vol. 10, pp. 652-664, 2013.
[37] A. Riad, H. Elminir and H. Elattar, “Evaluation of neural networks in the subject of prognostics as
compared to linear regression model” International Journal of Engineering & Technology, vol. 10, pp.
52-58, 2010.
[38] J. Liu, A. Saxena, K. Goebel, B. Saha and W. Wang, An adaptive recurrent neural network for remaining
useful life prediction of lithium-ion batteries, NATIONAL AERONAUTICS AND SPACE
ADMINISTRATION MOFFETT FIELD CA AMES RESEARCH CENTER, 2010.
15
[39] M. Rigamonti, P. Baraldi and E. Zio, “Echo state network for the remaining useful life prediction of a
turbofan engine” annual conference of the prognostics and health management society, vol. 2015, 2016.
[40] N. Gugulothu, T. V. Vishnu, P. Malhotra, L. Vig, P. Agarwal and G. Shroff, “Predicting Remaining
Useful Life using Time Series Embeddings based on Recurrent Neural Networks” 2017.
[41] 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, pp. 11-20, 2018.
[42] B. Saha, S. Poll, K. Goebel and J. (. Christophersen, “September). An integrated approach to battery
health monitoring using Bayesian regression and state estimation” In autotestconIEEE (p, pp. 646-653,
2007.
[43] Mathworks, Predictive Maintenance Toolbox - User Guide, Natick, MA: Mathworks, 2018.
[44] B. S. Lopez, A. M. Siddiolo, P. P. Adhikari and M. A. Buderath, "Bayesian paradigm for aircraft
operational capability assessment and improved fault diagnostics", EUROPEAN CONFERENCE OF
THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY, 2016.
[45] “Prognostics Center of Excellence Data Repository” NASA, [Online]. Available:
https://ti.arc.nasa.gov/tech/dash/groups/pcoe/prognostic-data-repository/. [Accessed September 2018].
[46] Y. L. G. W. X. L. Jianming Shi, “Health index synthetization and remaining useful life estimation”
Maintenance and Reliability, vol. 18 (4), p. 621–631, 2016.
Biographies
Partha Adhikari – Aerospace professional with more than 20 years of experience in
Integrated Vehicle Health Management (IVHM) / Predictive Maintenance, Predictive Data
Analytics, Navigation, Simulation & Modelling of Aircraft Systems and Avionics
integration. He has served in different techno-management roles in various Aerospace
organisations. Presently he is the IVHM domain lead in Airbus India, Bangalore and is
working on design and development of Data driven Diagnostics and Prognostics, High
level reasoning and tools related to decision support. He has published over 15 papers in the
field of estimation, signal processing and IVHM in national as well as international
conferences and journals.
Harsha Gururaja Rao – Software Engineer with more than 7 years of experience in
designing & developing enterprise level Software applications and Predictive Maintenance
solutions for Aircraft systems. Harsha has a Bachelor of Engineering Degree in Computer
Science from the Visvesaraya Technological University. Harsha, in his current role as an
Engineer at Airbus India, Bangalore is working on developing Data driven solutions to
meet Aerospace & Defense business systems requirements in the areas of IVHM and
Predictive Maintenance.
Matthias Buderath - Aeronautical Engineer with more than 30 years of experience in
structural design, system engineering and product and service support. His main expertise
and competency is in system integrity management, service solution architecture and
Integrated System Health Monitoring and Management. Today he is head of Airbus
Defence and Space R&T Strategy and Senior Expert Integrated System / Aircraft Health
Monitoring and Management. He is member of international Working Groups covering
Through Life Cycle Management, Integrated System Health Management and Structural
Health Management. He has published more than 80 papers in the field of Structural Health
Management, Integrated Health Monitoring and Management, Structural Integrity
Programme Management and Maintenance and Fleet Information Management Systems.