Table of Contents

Cover

Title

Introduction

I.1. From the reinforcement of techno-socio-economic issues…
I.2. To the apparition of a topic: PHM…
I.3. To the purpose of this book…

1 PHM and Predictive Maintenance

1.1. Anticipative maintenance and prognostics
1.2. Prognostics and estimation of the remaining useful life (RUL)
1.3. From data to decisions: the PHM process
1.4. Scope of the book

2 Acquisition: From System to Data

2.1. Motivation and content
2.2. Critical components and physical parameters
2.3. Data acquisition and storage
2.4. Case study: toward the PHM of bearings
2.5. Partial synthesis

3 Processing: From Data to Health Indicators

3.1. Motivation and content
3.2. Feature extraction
3.3. Feature reduction/selection
3.4. Construction of health indicators
3.5. Partial synthesis

4 Health Assessment, Prognostics and Remaining Useful Life - Part A

4.1. Motivation and content
4.2. Features prediction by means of connectionist networks
4.3. Classification of states and RUL estimation
4.4. Application and discussion
4.5. Partial synthesis

5 Health Assessment, Prognostics, and Remaining Useful Life - Part B

5.1. Motivation and object
5.2. Modeling and estimation of the health state
5.3. Behavior prediction and RUL estimation
5.4. Application and discussion
5.5. Partial synthesis

Conclusion and Open Issues

C.1. Summary
C.2. The international PHM: open questions

Bibliography

Index

End User License Agreement

List of Tables

2 Acquisition: From System to Data

Table 2.1. Main methods used for risks evaluation and management
Table 2.2. Examples of physical quantities to observe [CHE 08a]
Table 2.3. Failure distribution of asynchronous motors
Table 2.4. Examples of experiments performed on Pronostia platform

3 Processing: From Data to Health Indicators

Table 3.1. Examples of temporal features
Table 3.2. C-MAPSS features
Table 3.3. TURBOFAN–parameterization of predictive tools
Table 3.4. TURBOFAN - predictability of features from F1 to F8
Table 3.5. TURBOFAN - RUL estimation error with ANFIS

4 Health Assessment, Prognostics and Remaining Useful Life - Part A

Table 4.1. TURBOFAN – Synthesis of prediction performances
Table 4.2. Datasets for SW-ELM performance tests
Table 4.3. Comparison between the performances of the models
Table 4.4. Characteristics of cutting tools
Table 4.5. Performances of robustness and applicability for a cutting tool
Table 4.6. Performances of reliability and applicability for three cutting tools
Table 4.7. Performances of reliability and applicability for unknown data
Table 4.8. Performances of reliability and applicability with an ensemble of SW-ELM
Table 4.9. PHM issues and S-MEFC algorithm
Table 4.10. TURBOFAN - prognostics by dynamic thresholding - results of tests

5 Health Assessment, Prognostics, and Remaining Useful Life - Part B

Table 5.1. Classification of Markov processes [SOL 06a]
Table 5.2. Algorithms for learning the parameters of a DBN [BEN 03a, MUL 05]
Table 5.3. Inference methods for a DBN (see [MUR 02] for further details)
Table 5.4. Parameters estimated within 10^-3 mm

Conclusion and Open Issues

Table C.1. Towards a new generation of PHM modules?

List of Illustrations

Introduction

Figure I.1: Publications with PHM as a topic (Web of Sciences, February 2016)

1 PHM and Predictive Maintenance

Figure 1.1 Forms of maintenance according to the standard EN 13306 (2001) [EN 01]. For the color version of this Figure, see www.iste.co.uk/zerhouni1/phm.zip
Figure 1.2. Illustration of prognostic process. For the color version of this Figure, see www.iste.co.uk/zerhouni1/phm.zip
Figure 1.3. Taxonomy of prognostic approaches
Figure 1.4. Hybrid prognostic approaches
Figure 1.5. Complementarity of detection, diagnostic, and prognostic activities [GOU 11]
Figure 1.6. PHM cycle as an adaptation of the OSA-CBM architecture. For the color version of this Figure, see www.iste.co.uk/zerhouni1/phm.zip

2 Acquisition: From System to Data

Figure 2.1. From system to PHM data - steps
Figure 2.2. Choice of critical components - approach
Figure 2.3. Structure of an acquisition process [ASC 03]
Figure 2.4. Examples of sensors of force and quartz accelerometers
Figure 2.5. Simplified and illustrated structure of an acquisition chain
Figure 2.6. Acquisition card for vibrations (a), temperature (b), and external frame for the computer (c)
Figure 2.7. Subsystems of a train
Figure 2.8. Critical components of a train’s motor
Figure 2.9. Experimental bench Pronostia and NSK 6804DD bearings used
Figure 2.10. Bench elements applying a radial force: pneumatic jack and proportional pressure regulator. For the color version of this Figure, see www.iste.co.uk/zerhouni1/phm.zip
Figure 2.11. Gear system and load transmission
Figure 2.12. Examples of degradation of outer and inner races of a bearing
Figure 2.13. Raw vibration signals obtained for two bearings under test

3 Processing: From Data to Health Indicators

Figure 3.1. Feature extraction, selection and reduction, and construction of health indicators
Figure 3.2. Feature extraction techniques, adapted from [YAN 08]
Figure 3.3. RMS, Kurtosis and fast Fourier transform of vibration signals produced on Pronostia platform
Figure 3.4. Short-time Fourier transform of a vibration signal generated on Pronostia platform. For the color version of this Figure, see www.iste.co.uk/zerhouni1/phm.zip
Figure 3.5. Illustration of wavelet packet decomposition of second order
Figure 3.6. WPD (1-800 Hz) of vibration signals produced on Pronostia platform (vertical and horizontal accelerometers)
Figure 3.7. Signal x(t) and its upper and lower envelopes
Figure 3.8. First extracted IMF
Figure 3.9. Decomposition process to obtain IMFs
Figure 3.10. Variation of the EMD residual as a function of the health state of the component
Figure 3.11. Examples of results obtained from vibration signals gathered on Pronostia platform: EMDhealthy EMD of a new bearing, EMDfaulty EMD of a faulty bearing, THHhealthy HHT of a new bearing, and THHfaulty HHT of a faulty bearing
Figure 3.12. Data reduction methods
Figure 3.13. PCA: reduction of data gathered on Pronostia (three to two dimensions)
Figure 3.14. Principle of kernel PCA, adapted from [BIS 06]
Figure 3.15. Illustration of the kernel trick process
Figure 3.16. Kernel PCA: reduction of data gathered on Pronostia
Figure 3.17. Illustration of ISOMAP steps [TEN 00]: (A) geodesic distance between two points, (B) neighborhood graph and approximation of geodesic distance by the shortest path between two points of the graph, (C) projection of data onto a space of dimension 2, where the geodesic distance is approximated by the straight line.
Figure 3.18. ISOMAP: extraction/reduction of two, then of one feature from vibration data gathered on Pronostia
Figure 3.19. Predictability concept
Figure 3.20. Selection of predictable features
Figure 3.21. Turboreactor [FRE 07]
Figure 3.22. TURBOFAN–prediction of feature F5 and associated predictability
Figure 3.23. TURBOFAN - predictability of features for H = 134 time-units
Figure 3.24. TURBOFAN - RUL estimation on a test sequence
Figure 3.25. Construction of health indicators based on the Hilbert-Huang transform
Figure 3.26. Illustration of IMFs selection
Figure 3.27. Health indicators obtained by using Hilbert-Huang transform

4 Health Assessment, Prognostics and Remaining Useful Life - Part A

Figure 4.1. From data to RUL-prediction and classification
Figure 4.2. Representation and taxonomy of long-term prediction approaches [GOU 12]
Figure 4.3. Schematization of long-term prediction approaches
Figure 4.4. Data sequences from NN3 competition (randomly chosen)
Figure 4.5. NN3–RMSE vs. processing time
Figure 4.6. Data sequences from TURBOFAN application (randomly chosen)
Figure 4.7. TURBOFAN - State classification and RUL estimation
Figure 4.8. Wavelet neural network of the SW–ELM [JAV 14b]
Figure 4.9. SW-ELM set and uncertainties of estimates
Figure 4.10. Wear estimation of cutting tools–methodology
Figure 4.11. Robustness and reliability tests
Figure 4.12. Prediction of wear of tools and confidence intervals at 95%
Figure 4.13. Classification of states and RUL estimation
Figure 4.14. States classification–problems inherent to the learning phase [GOU 13]
Figure 4.15. Classification methods and PHM (adapted from [GOU 13])
Figure 4.16. Prognostics without a priori information about the thresholds–overall synoptic
Figure 4.17. Offline phase: learning of predictors and classifiers
Figure 4.18. Online phase: predictions and estimations of states
Figure 4.19. TURBOFAN - distribution of measurements and lifespans. For the color version of this Figure, see www.iste.co.uk/zerhouni1/phm.zip
Figure 4.20. TURBOFAN – prediction error interval
Figure 4.21. TURBOFAN - RUL estimation by automatic thresholding (Test 1)
Figure 4.22. TURBOFAN - illustration of classes of variable states
Figure 4.23. TURBOFAN - dynamic assignment of failure thresholds
Figure 4.24. TURBOFAN - Estimated and actual RUL (for 100 tests)
Figure 4.25. TURBOFAN - pdf of the RUL: a) proposed approach, b) according to [RAM 13b]

5 Health Assessment, Prognostics, and Remaining Useful Life - Part B

Figure 5.1. Learning and exploitation of degradation models for the prognostics
Figure 5.2. Example of discrete Markov chain
Figure 5.3. Hidden Markov Model (HMM)
Figure 5.4. Representation of a stochastic Markov process by means of BN and DBN: ExRB expanded solution (BN) and ExRBD compact solution, according to [MUL 05]
Figure 5.5. Representation of HMM by means of a DBN developed over 3 instants
Figure 5.6. Representation of a HMM by means of a DBN: DBNHMMdet details of nodes and DBNHMMgen general compact version where P[X₁ = i] = π(i), P[X_t] = P[X_t = x_j|X_t-1 = x_i] = A(i, j), and P[O_t = v_k|X_t = x_i] = B = {b_i(k)}
Figure 5.7. Representation of a MoG-HMM by means of a DBN: MoGHMMRBDder model developed over three instants and MoGDBNHMMgen general compact version
Figure 5.8. Sensitivity study for determination of the number of mixtures M
Figure 5.9. Example of state sequence obtained by Viterbi algorithm
Figure 5.10. Selection of the best model that represents the current observations
Figure 5.11. Definition of short and long path
Figure 5.12. Machining test bench [PHM 10]
Figure 5.13. Clustering of states of the cutting tools
Figure 5.14. State sequence for a record of wear evolution
Figure 5.15. State sequence
Figure 5.16. W6 wear estimation for the tool 6, by using the data from tool 1 and 4 as learning data, and RUL6 predicted RUL for the tool 6

Conclusion and Open Issues

Figure C.1. PHM approach and V&V process

First published 2016 in Great Britain and the United States by ISTE Ltd and John Wiley & Sons, Inc.

Apart from any fair dealing for the purposes of research or private study, or criticism or review, as permitted under the Copyright, Designs and Patents Act 1988, this publication may only be reproduced, stored or transmitted, in any form or by any means, with the prior permission in writing of the publishers, or in the case of reprographic reproduction in accordance with the terms and licenses issued by the CLA. Enquiries concerning reproduction outside these terms should be sent to the publishers at the undermentioned address:

ISTE Ltd

27-37 St George’s Road

London SW19 4EU

www.iste.co.uk

John Wiley & Sons, Inc.

111 River Street

Hoboken, NJ 07030

USA

www.wiley.com

The rights of Rafael Gouriveau, Kamal Medjaher and Noureddine Zerhouni to be identified as the author of this work have been asserted by them in accordance with the Copyright, Designs and Patents Act 1988.

Library of Congress Control Number: 2016947860

British Library Cataloguing-in-Publication Data

A CIP record for this book is available from the British Library

ISBN 978-1-84821-937-3

Introduction

I.1. From the reinforcement of techno-socio-economic issues…

The “RAMS” services (reliability, availability, maintainability, and safety) today are widely applied to perform limited studies or in-depth analysis. Indeed, industrial maintenance appears to be the source and the target of scientific developments, which is reflected into specific actions of partnership “industry research”, or projects of greater scope, such as the one of the IMS center¹. In a more focused way, at a business level, the traditional concepts of predictive and corrective maintenance are being gradually completed by taking into account the failure mechanisms in a more proactive way [HES 08, MUL 08b]; industrialists tend to strengthen their ability to anticipate failures in order to resort to the most correct possible preventive actions with a goal of reducing costs and risks. Therefore, the implementation of solutions of Prognostics and Health Management (PHM) plays a growing role, and the prognostics process is considered today as one of the main levers in the research of global performance.

– First of all, the failure anticipation of critical elements foresees industrial risks and assures the safety of people and goods.
– Then, prognostics assures a continuity of services, and hence increases their quality.
– Additionally, in environmental terms, industrial prognostics is in line with sustainable development principles: it increases the availability and lengthens the life cycle of industrial systems.
– Finally, implementing predictive maintenance (based on prognostics) requires a qualification and contributes to the development of the technical maintenance staff.

I.2. To the apparition of a topic: PHM…

Beyond the reaction that it can encounter among the industrial world, this topic of prognostics or PHM becomes naturally a research framework in its own right, and tends to be more and more visible within the scientific community. Several laboratories are interested in it today (NASA PCoE, Atlanta University, IMS Center and Army Research Lab in the USA, Toronto University in Canada, CityU-PHM Center Hong-Kong University, etc.), and every year, four conferences dedicated to the PHM topic are held², two of which are supported by the IEEE Reliability Society. This is an indicator of the growing awareness of this topic and, moreover, that the research studies in this domain are seeing rapid growth (Figure I.1).

**Figure I.1.** *Publications with PHM as a topic (Web of Sciences, February 2016)*

I.3. To the purpose of this book…

In addition to the evident rise of this topic, PHM solutions are the result of the evolution of methods and technologies of dependability, monitoring and maintenance engineering. This book fits into this context. Our goal is to present the appearance of this PHM topic, to show how it completes the traditional maintenance activities, to highlight the underlying problems, to describe the advantages that can be expected from implementing PHM solutions in industry and, finally, to consider the major problems and challenges which are still relevant today. For this purpose, the book is structured as follows:

– Chapter 1 - PHM and Predictive Maintenance. The first chapter covers the general presentation of the PHM process. Here we highlight the strategic urgency to take the failure mechanisms into account in a more proactive way, and we describe the evolution of related challenges and prerogatives of the maintenance services. In the following, we introduce the PHM activity, and, more specifically, the prognostic process that lies beneath. This chapter defines a coherent set of treatments that are necessary to implement a PHM approach; these different basic building blocks are developed in the following chapters.
– Chapter 2 - Acquisition: from System to Data. In order to deploy PHM, one must be able to observe the behavior of the analyzed system. This is what the second chapter deals with; here we suggest an approach for the generation of monitoring data representative of degradation mechanisms in critical components.
– Chapter 3 - Processing: from Data to Health Indicators. The acquired raw data from the system under consideration (Chapter 2) must generally be preprocessed, in order to extract and select the descriptors that help, over time, to reveal the (mal)function. Chapter 3 deals with this aspect; here we describe the usual tried and tested approaches for creating health indicators.
– Chapters 4 and 5 - Heath Assessment and Prognostics of Residual Lifetime. These two last chapters (excluding the conclusion) focus on the development of models and/or methods for estimating the state of health, and of prognostics of behavior of the monitored system. Thus, we describe how to exploit the information produced in the previous steps in order to estimate the residual lifetime, and to associate to it a confidence measure. We also show how the variability of operating conditions and mission profiles, as well as the physical unawareness of transient states, impact health state modeling and the ensuing performances.

In the conclusion, we deliver a critical view on the maturity of PHM activity, and we open a discussion, on one hand on the problems which remain open to the international context, on the other hand on the decision-making process that stems from these problems. The latter aspect is the subject of the book From PHM Concept to Predictive Maintenance 2 [CHE 16] in which we address the aspects regarding the strategic processes of maintenance decisions, and, more broadly, the life cycle management of product/equipment: the gathered data is traced and transformed into knowledge in order to support the decision-making concerning maintenance, re-design or recycling of equipment.

From Prognostics and Health Systems Management to Predictive Maintenance 1

Monitoring and Prognostics

Introduction

I.1. From the reinforcement of techno-socio-economic issues…

I.2. To the apparition of a topic: PHM…

I.3. To the purpose of this book…