1. Trang chủ
  2. » Tất cả

Diagnosis of the powertrain systems for autonomous electric vehicles

144 1 0
Tài liệu đã được kiểm tra trùng lặp

Đang tải... (xem toàn văn)

Tài liệu hạn chế xem trước, để xem đầy đủ mời bạn chọn Tải xuống

THÔNG TIN TÀI LIỆU

Thông tin cơ bản

Tiêu đề Diagnosis of the Powertrain Systems for Autonomous Electric Vehicles
Tác giả Tunan Shen
Người hướng dẫn Prof. Dr.-Ing. Michael Bargende, Prof. Dr.-Ing. Hans-Christian Reuss, Prof. Dr.-Ing. Jochen Wiedemann
Trường học University of Stuttgart
Chuyên ngành Automotive Engineering
Thể loại Dissertation
Năm xuất bản 2021
Thành phố Stuttgart
Định dạng
Số trang 144
Dung lượng 3,66 MB

Các công cụ chuyển đổi và chỉnh sửa cho tài liệu này

Nội dung

In the thesis, the faultanalysis of powertrain systems in today’s battery electric vehicles is presentedfirstly, followed by an introduction of the background context and the state-of-th

Trang 1

Diagnosis of the

Powertrain Systems for Autonomous

Trang 2

Wissenschaftliche Reihe

Fahrzeugtechnik Universität Stuttgart

Reihe herausgegeben von

Michael Bargende, Stuttgart, Deutschland

Hans-Christian Reuss, Stuttgart, Deutschland

Jochen Wiedemann, Stuttgart, Deutschland

Trang 3

Das Institut für Fahrzeugtechnik Stuttgart (IFS) an der Universität Stuttgarterforscht, entwickelt, appliziert und erprobt, in enger Zusammenarbeit mitder Industrie, Elemente bzw Technologien aus dem Bereich modernerFahrzeugkonzepte Das Institut gliedert sich in die drei Bereiche Kraftfahrwesen,Fahrzeugantriebe und Kraftfahrzeug-Mechatronik Aufgabe dieser Bereiche istdie Ausarbeitung des Themengebietes im Prüfstandsbetrieb, in Theorie und Simu-lation Schwerpunkte des Kraftfahrwesens sind hierbei die Aerodynamik, Akustik(NVH), Fahrdynamik und Fahrermodellierung, Leichtbau, Sicherheit, Kraftüber-tragung sowie Energie und Thermomanagement – auch in Verbindung mit hybri-den und batterieelektrischen Fahrzeugkonzepten Der Bereich Fahrzeugantriebewidmet sich den Themen Brennverfahrensentwicklung einschließlich Regelungs-und Steuerungskonzeptionen bei zugleich minimierten Emissionen, komplexeAbgasnachbehandlung, Aufladesysteme und -strategien, Hybridsysteme undBetriebsstrategien sowie mechanisch-akustischen Fragestellungen Themen derKraftfahrzeug-Mechatronik sind die Antriebsstrangregelung/Hybride, Elektromo-bilität, Bordnetz und Energiemanagement, Funktions- und Softwareentwick-lung sowie Test und Diagnose Die Erfüllung dieser Aufgaben wird prüf-standsseitig neben vielem anderen unterstützt durch 19 Motorenprüfstände,zwei Rollenprüfstände, einen 1:1-Fahrsimulator, einen Antriebsstrangprüfs-tand, einen Thermowindkanal sowie einen 1:1-Aeroakustikwindkanal Die wis-senschaftliche Reihe „Fahrzeugtechnik Universität Stuttgart“ präsentiert über die

am Institut entstandenen Promotionen die hervorragenden Arbeitsergebnisse derForschungstätigkeiten am IFS

Reihe herausgegeben von

Prof Dr.-Ing Michael Bargende

Stuttgart, Deutschland

Weitere Bände in der Reihehttps://link.springer.com/bookseries/13535

Trang 4

Tunan Shen

Diagnosis of the

Powertrain Systems for Autonomous

Electric Vehicles

Trang 5

Tunan Shen

Institute of Automotive Engineering (IFS),

Chair of Automotive Mechatronics

University of Stuttgart

Stuttgart, Germany

Zugl.: Dissertation Universität Stuttgart, 2021

D93

ISSN 2567-0042 ISSN 2567-0352 (electronic)

Wissenschaftliche Reihe Fahrzeugtechnik Universität Stuttgart

ISBN 978-3-658-36991-0 ISBN 978-3-658-36992-7 (eBook)

https://doi.org/10.1007/978-3-658-36992-7

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer

Fachmedien Wiesbaden GmbH, part of Springer Nature 2022

This work is subject to copyright All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprint- ing, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed.

The use of general descriptive names, registered names, trademarks, service marks, etc in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use.

The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Responsible Editor: Stefanie Eggert

This Springer Vieweg imprint is published by the registered company Springer Fachmedien Wiesbaden GmbH part of Springer Nature.

The registered company address is: Abraham-Lincoln-Str 46, 65189 Wiesbaden, Germany

Trang 6

I believe that the electric vehicle revolution is coming and automated drivingwill come soon Then, most people may not own a car in future any more,but hire self-driving, electric ride-shares to get around For those autonomouselectric vehicles, which owned by ride-share companies, the reliability andavailability are more important than today’s private owned cars I want to be

a part of the change, so I decided to pursue my PhD in the area of diagnosis

of powertrain system for autonomous electric vehicles since 2017, after threeyears working at Robert Bosch GmbH

My deep gratitude goes to my supervisor Prof Dr.-Ing Hans-Christian Reussforhis endless support, enthusiasm, knowledge and friendship

I am extremely grateful to my supervisors at Robert Bosch GmbH Dr AhmetKilic and Dr Christian Thulfaut for the valuable support and motivation Theirinsights and advice give me a great help

Special thanks to Dr Norbert Müller and Dr Achim Henkel for giving me thechance to carry out this work at Corporate Sector Research and Advance Engi-neering at Robert Bosch GmbH in Renningen My appreciation also extends

to Adam Babik, Dr Andreas Vogt, Dr Andreas Schönknecht, Daniel Acs, Dr.Deng Shi, Erik Hoevenaars, Dr Paul Mehringer, Rajaram Suresh and YupingChen for your valuable comments on this work and all my colleagues for thewonderful cooperation as well

Furthermore, I want to thank my wife Shiang and my parents You believed

in me when I was doubt You kept me motivated through my darkest thoughts.Finally, I would like to thank my daughter Cindy Thank you for coming into

my life Thank you for making me smile like crazy Thank you for making mehappy

Trang 7

Preface V Figures XI Tables XV Abbreviations XVII Symbols XIX Abstract XXI Kurzfassung XXIII

1 Introduction 1

1.1 Motivation 1

1.2 Objectives 2

1.3 Contributions of the Thesis 3

1.4 Organization of the Thesis 3

2 Fault Analysis 5

2.1 Battery 5

2.2 Inverter 6

2.3 Electric Machine 8

2.4 Gearbox 10

2.5 Scope of the Thesis 10

3 Background and State of the Art 13

3.1 Fault Diagnostic Methods 13

3.1.1 Signal-based Fault Detection Methods 13

3.1.2 Model-based Fault Detection Methods 14

3.1.3 Data-based Fault Detection Methods 15

3.2 Signal Processing Techniques 16

3.2.1 Time Domain Features 16

3.2.2 Statistical Features 16

3.2.3 Frequency Domain Features 17

3.2.4 Envelope Analysis 17

Trang 8

VIII Inhaltsverzeichnis

3.2.5 Time Frequency Domain Features 18

3.3 Machine Learning Algorithms 21

3.3.1 Decision Tree 21

3.3.2 Anomaly Detection 22

3.3.3 Artificial Neural Network 23

4 Diagnosis of Electrical Faults in Electric Machines 27

4.1 Related Works and Current Challenges 27

4.2 Contributions of the Thesis 29

4.3 Analytical Modeling of Faults 30

4.3.1 Analytical Modeling of a Healthy PMSM 30

4.3.2 Analytical Modeling of PMSM with TSC 32

4.3.3 Analytical Modeling of PMSM with PSC 32

4.3.4 Analytical Modeling of PMSM with UWR 33

4.3.5 Analytical Modeling of PMSM with Sensor Faults 33

4.4 Analysis of the Behavior of a PMSM in Different Conditions 34

4.5 Feature Engineering 38

4.5.1 Data Normalization 38

4.5.2 Feature Extraction 39

4.6 Diagnostic Concept 41

4.7 Physical Model based Diagnostic Model 42

4.8 Self Condition Monitoring (SCM) Diagnostic Model 48

4.9 Fleet Data-based Fault Diagnostic Model 52

4.10 Multi-stage Diagnostic Concept 56

4.11 Conclusion 65

5 Diagnosis of Mechanical Faults in Electric Machines 67

5.1 Fault Mechanisms of Bearing 67

5.2 Related Works 69

5.3 Current Challenges and Contribution of the Thesis 72

5.3.1 Current Challenges 72

5.3.2 Contributions of the Thesis 72

5.4 Data Set Description 73

5.5 Feature Engineering 75

5.5.1 De-noising 75

Trang 9

Inhaltsverzeichnis IX

5.5.2 Feature Extraction 76

5.5.3 Evaluation of Features 90

5.6 Validation with Other eAxles 92

5.7 Diagnostic Concept 97

5.8 Conclusion 101

6 Conclusion and Outlook 103

Bibliography 107

Trang 10

2.1 Failure mechanism of a battery fire 6

2.2 Distribution of fragile components 7

2.3 Distribution of failed components in electric machines 9

3.1 Scheme of a signal-based fault detection method 14

3.2 Scheme of a model-based fault detection method 15

3.3 Scheme of a data-based fault detection method 15

3.4 Envelope of a vibration signal 19

3.5 An example of STFT (a) Quadratic chirp signal; (b) STFT of quadratic chirp signal 20

3.6 A schematic structure of decision tree 21

3.7 (a) Distance-based anomaly detection; (b) Density-based anomaly detection 23

3.8 Diagram of an artificial neuron 24

3.9 Structure of neural network 25

4.1 Equivalent circuit of healthy and faulty stators of PMSM 31

4.2 The difference of a current sensor in healthy and faulty conditions 34

4.3 Operating area of PMSM 35

4.4 Current and torque of machine in different conditions at operating point Tre f = 100 Nm, nmech= 5000 rpm 37

4.5 Comparison of the three phase currents of healthy and faulty con-ditions at operating point Tre f = 40 Nm, nmech= 4000 rpm 38

4.6 Comparison of symptoms in different conditions 44

4.7 The tree structure of decision tree model 45

4.8 Normalized confusion matrices of the physical model on (a) trai-ning and (b) test data 46

4.9 Wrong predictions of the physical model 47

4.10 Distribution of distance 51

4.11 Confusion matrices of the SCM model on (a) training and validation data and (b) test data 51

4.12 Wrong predictions of the SCM model 52

Trang 11

XII Figures

4.13 Confusion matrices of NN model on (a) training, (b) validation and

(c) test data 55

4.14 Comparison of the probabilities of right and wrong predictions 56

4.15 Comparison the probabilities of right and wrong predictions 57

4.16 Confusion matrices of (a) physical model, (b) SCM model and (c)NN model on test data 58

4.17 Confusion matrices of the physical model at (a) frequently driven point, (b) low speed point, (c) high speed point 60

4.18 Confusion matrices of combined models (a) 1st stage, (b) 2nd stage, (c) 3rd stage 61

4.19 Multi-class confusion matrices of combined model at (a) 1st stage, (b) 2nd stage, and (c) 3rd stage 64

5.1 Rolling element bearing (a) Structure of a rolling element bearing; (b) Pictures of defective bearings 68

5.2 Ramp tests over lifetime 74

5.3 Rmsat different measurement time and speeds 74

5.4 The noisy data 77

5.5 The filtered data 77

5.6 Rmsof bearing 2-3 in XJTU-SY data set 79

5.7 Rmsof an eAxle 79

5.8 Process from raw data to health indicator 80

5.9 Normalized Rms in different clusters 81

5.10 Rms as a health indicator 81

5.11 Cage characteristic feature Pwrf t f of bearing 2-3 83

5.12 Cage characteristic feature Envf t f of bearing 2-3 83

5.13 Spectrogram of a ramp test 85

5.14 The first three harmonics of FTF 85

5.15 Cage characteristic feature Pwrf t f of an eAxle 86

5.16 Cage characteristic feature Envf t f of an eAxle 86

5.17 Spectrogram of ramp test No 1 88

5.18 Spectrogram of ramp test No 11 88

5.19 Spectrogram of ramp test No 22 89

5.20 Medianspecover the time 89

5.21 Comparison of features 91

5.22 The ranking of sensitivity 93

5.23 The ranking of monotonicity 93

Trang 12

Figures XIII

5.24 Feature P2P of four eAxles 95

5.25 Feature Envf t f of four eAxles 95

5.26 Feature Medianspecof four eAxles 96

5.27 Anomaly score of eAxle1 99

5.28 Prediction of the Medianspeccurves .100

5.29 Comparison of the synthetic and true Medianspeccurves 101

5.30 Prediction of the synthetic Medianspeccurve .102

Trang 13

3.1 Time domain features 16

3.2 Statistical features 17

3.3 Characteristic frequencies of bearing faults 18

3.4 Typical activation functions 25

4.1 Parameters of PMSM 36

4.2 Extracted time domain and statistical features 39

4.3 Comparison of time domain and statistical features in different conditions 40

4.4 Comparison of harmonic features in different conditions 41

4.5 The number of samples 49

4.6 Data split 53

4.7 Hyperparameters of NN model 53

4.8 Evaluation of single models 58

4.9 Evaluation of combined models 62

4.10 Evaluation of models with unknown faults 63

5.1 A short overview of open accessed bearing data sets 69

5.2 A list of features used for bearing RUL prediction 71

5.3 Comparison of two data sets 75

5.4 A list of extracted features from the eAxle data set 92

Trang 14

ADAC Allgemeine Deutscher Automobil Club

AI Artificial Intelligence

ANN Artificial Neural Network

ASCMO Advanced Simulation for Calibration, Modeling and

Optimization

BEV Battery Electric Vehicle

BPFI Ball Pass Frequency Inner race

BPFO Ball Pass Frequency Outer race

CART Classification And Regression Trees

CWRU Case Western Reserve University

FFT Fast Fourier Transformation

IFS Institute for automotive engineering

IGBT Insulated-Gate Bipolar Transistor

IMS the Center of Intelligent Maintenance Systems

Trang 15

XVIII Abbreviations

MEMS Micro-Electro-Mechanical Systems

MFPT Machinery Failure Prevention Technology Challenge

PMSM Permanent-Magnet Synchronous Motor

PSC Phase-to-phase Short Circuit

ReLu Rectified Linear Unit

SAE Society of Automotive Engineers

STFT Short Time Fourier Transformation

TSC Turn-to-turn Short Circuit

XJTU-SY Xi’an JiaotongUniversity and the Changxing Sumyoung

Technology Co

Trang 16

dp,q Euclidean distance between point p and q

-∥dx∥ Euclidean distance between point x and zero

point

eo f f set Offset error A

fb f f Ball fault frequency Hz

fbp f i Ball pass frequency inner race Hz

fbp f o Ball pass frequency outer race Hz

ff t f Fundamental train frequency Hz

Id,re f Amplitude of reference direct current A

Ii,norm Amplitude of normalized phase current A

Id,re f Amplitude of reference quadrature current A

Trang 17

-XX Symbols

-Greek Letters

-ηa Ratio of healthy turns at phase a in case of TSC

-ηb Ratio of healthy turns at phase b in case of TSC

-η Ratio of short circuited turns in case of TSC

Trang 18

In autonomous electric vehicles, the reliability of electrical powertrain systemmight be crucial A critical failure in the powertrain system would lead tobreakdown of the vehicle It is possible that the vehicle stops in the middle of atraffic lane without adequate protection Therefore, the powertrain system ofsuch an autonomous electric vehicle should be designed with fail-operationalability One solution to meet this requirement is system redundancy However,the additional cost and weight of a second powertrain system are not favorable

In contrast, an economically competitive solution is to prevent the vehiclefrom dangerous operating conditions by using predictive diagnosis of criticalfaults This work aims to enhance the availability of powertrain systems forautonomous electric vehicles by improving the diagnostic capability for criticalfaults Once a fault can be detected at an early stage, critical faults can beavoided by implementing proper fault reaction strategies such as operating indegradation mode until the degraded part is repaired In the thesis, the faultanalysis of powertrain systems in today’s battery electric vehicles is presentedfirstly, followed by an introduction of the background context and the state-of-the-art technologies Then, the electrical and mechanical faults of electricmachine are focused in this thesis

Winding faults are the most critical electrical faults in electric machines Inthis thesis, three winding faults, namely turn-to-turn short circuit (TSC), phase-to-phase short circuit (PSC), and unbalanced winding resistance (UWR) aremodeled analytically After analyzing the behaviors of the machine under va-rious faulty conditions, the three phase currents of electric machine are used

to detect the faults However, an erroneous current signal caused by devicetolerances, temperature drift, aging, and noise can lead to a misdiagnosis ofmachine faults Therefore, two typical sensor faults, i.e gain fault and offsetfault, are also modeled Several time and frequency domain features are extrac-ted from the three phase currents to distinguish different faults On this basis, amulti-stage diagnostic concept is proposed, which consists of three diagnosticmodels for different considerations The first model is a physical model; the

Trang 19

XXII Abstract

second model is a data-based anomaly detection model; the third model is

a neural network model based on fleet data To improve the total diagnosticcapability, a decision making function is designed to derive a comprehensivedecision from the predictions of various operating points and different models.This diagnostic concept is validated with simulation data for over ten thousandvarious faulty conditions covering diverse fault types, severity and operatingpoints The proposed diagnostic concept shows an outstanding diagnostic capa-bility for both winding and sensor faults under wide operating range compared

to diagnostic systems in state of the art Moreover, it has high robustness againstunknown faults

The most critical mechanical faults are bearing faults Although the fault chanisms of bearing faults in laboratory have been well investigated, thereare still many challenges for bearing fault diagnosis in actual products Thecurrent challenges can be summarized as following: limited data, limited label,and knowledge transfer from laboratory to real products To gain more datafrom industry products, the vibration data from the endurance tests of eAxles,which is a new product used in the powertrain system, are used to develop thediagnostic concept in the thesis The vibration data of predefined ramp testsduring the whole lifetime is used to detect the bearing faults Various featuresfrom time and frequency domain are extracted from the raw data and comparedwith those extracted from a set of bearing data acquired from a bearing testbench in the laboratory The difference in each feature between the two data setsare analyzed Their effectiveness and transfer ability are evaluated Then, thebest features are validated with data from non-trained defective eAxles In theend, a multi-level diagnostic concept is proposed The first level is based on theenvelope analysis of the vibration signal It is able to detect severe bearing faultand protect the product from further damage At the second level, the healthstate of the eAxle is monitored using an unsupervised learning algorithm, sothat the degradation of eAxle can be observed The prediction of health state isgiven by the third level It is demonstrated that a good diagnostic capability can

me-be achieved with the developed diagnostic concept

Combining the proposed two diagnostic concepts for electrical and mechanicalfaults in the electric machine, the majority of faults in the electric machine can

be detected predictively Once an incipient fault is detected, the autonomous

Trang 20

Abstract XXIII

electric vehicle is able to react to this fault in time and avoid critical faults thatlead to the breakdown of the powertrain system

Trang 21

Der zunehmende Personen- und Warenverkehr sowie die wachsende sierung erfordern zukünftig neue Mobilitätskonzepte Laut einer Studie vonRoland Berger wird deshalb der Markt für urbane Shuttles (Urban Automa-ted Shuttle, UAS) und Robo-Taxen (Urban Automated Taxi, UAT) bis 2030deutlich steigen Die Studie prognostiziert, dass der Anteil der Kilometer, diemit einem privat besessenen PKW gefahren werden, bis 2030 um ca 28 %sinkt, während der Anteil der mit einem UAT gefahrenen Kilometer um 27 %steigen wird [1] Voll automatisiert fahrende Fahrzeuge – die gemäß heutigerEinschätzung bereits rein aus marktwirtschaftlichen Gründen Elektrofahrzeugesein werden – nehmen demnach eine Schlüsselrolle in zukünftigen Mobili-tätszenarien ein Solche Fahrzeuge werden im Laufe ihrer Lebensdauer vielhäufiger im Einsatz sein als heutige, privat besessene PKW und erfordern somiteine wesentlich höhere Verfügbarkeit Darüber hinaus müssen Fahrzeuge mitSAE-Automatisierungslevel 4 und 5 (SAE, Society of Automotive Engineers)ohne den Fahrer als Rückfallebene auskommen und sich im Fehlerfall selbst

Urbani-in eUrbani-inen sicheren Zustand brUrbani-ingen Um diese Anforderung zu erfüllen, werdensicherheitskritische Systeme zur Erhöhung der Zuverlässigkeit typischerweiseredundant ausgelegt wie z.B bei Sensoren oder Computer-, Brems-, Lenk- undStromversorgungssystemen

(Teil-)Ausfälle des Antriebssystems können zur Reduktion oder sogar zum pletten Verlust der Antriebskraft führen Das Fahrzeug könnte dann, je nachVerkehrssituation, an einer für die Insassen gefährlichen Position – beispiels-weise auf der mittleren Fahrspur einer Autobahn – liegen bleiben Aus diesemGrund müsste der Antriebsstrang eines vollautomatisierten Elektrofahrzeugszur Vermeidung eines derartigen Szenarios fehlertolerant ausgelegt werden Umdiese Anforderung zu erfüllen, kann das Antriebssystem redundant gestaltetwerden Allerdings sind die hierfür erforderlichen zusätzlichen Kosten unddas zusätzliche Gewicht nachteilig Eine alternative Lösung besteht darin, dasAntriebssystem vor schädigenden Betriebsbedingungen, die zu (Teil-)Ausfällen

Trang 22

In Kapitel 2 werden die relevanten Fehler im Antriebssystem eines elektrischen Fahrzeugs gemäß heutigem Stand der Technik analysiert DieFehlerursachen und Ausfallmechanismen jedes Subsystems (inklusive Bat-teriesystem, Inverter, elektrische Traktionsmaschine und Getriebe) und dieDiagnosemöglichkeiten werden untersucht Da die kritischen Fehler in einemfrühen Zustand vorhergesagt werden müssen, wird die Umsetzbarkeit einerentsprechenden prädiktiven Diagnose für jeden Fehler getrennt bewertet Dabeiliegt der Fokus dieser Arbeit auf Fehlerzuständen in den elektromechanischenDomänen der genutzten elektrischen Traktionsmaschinen.

batterie-In Kapitel 3 werden zunächst die theoretischen Grundlagen cher Fehlerdiagnosemethoden erläutert Dazu erfolgt in Abschnitt 3.1 eineBetrachtung der drei relevanten Kategorien von Diagnoseverfahren: signalba-siert, modellbasiert und datenbasiert Anschließend werden in Abschnitt 3.2verschiedene Methoden zur Verarbeitung der Sensorsignale im Hinblick aufeine Verwertung als Kenngröße zur Diagnose, also für das sogenannte FeatureEngineering, beleuchtet Dies umfasst die statistische Auswertung genausowie die Auswertung der Signale im Zeit- und Frequenzbereichs In Abschnitt3.3 werden abschließend die Prinzipien der später verwendeten, maschinellenLernalgorithmen kurz erläutert

unterschiedli-Das entwickelte Diagnosekonzept für die elektrischen Fehler in einer schen Maschine wird schließlich in Kapitel 4 vorgestellt Zunächst werden

elektri-in Abschnitt 4.1 relevante Veröffentlichungen zur Fehlererkennung elektri-in eelektri-inerelektrischen Maschine zusammengefasst Die aktuellen Herausforderungen derFehlererkennung lassen sich wie folgt einordnen:

Trang 23

Kurzfassung XXVII

• Früherkennung

Die meisten Diagnosekonzepte sind gemäß Stand der Technik nicht inder Lage, die Fehler ausreichend frühzeitig zu erkennen, um einen de-gradierten Weiterbetrieb zu ermöglichen Die elektrische Maschine kanndaher nur abgeschaltet werden und es ist nicht möglich, das Fahrzeugzuverlässig an einen sicheren Ort zu bringen

• Fehleridentifikation

Die meisten Ansätze konzentrieren sich nur auf die Erkennung eines lers an sich, ohne gleichzeitig die Art des Fehlers zu erfassen Sie könnenbeispielsweise nicht zwischen Maschinen- und Sensorfehlern unterschei-den Daher könnte die elektrische Maschine aufgrund eines durch einenSensorfehler verursachten Fehlalarms abgeschaltet werden, obwohl einWeiterbetrieb möglich wäre

Feh-• Robustheit

Die meisten Ansätze berücksichtigen nur eine sehr begrenzte Anzahl vonArbeitspunkten der elektrischen Maschine Sie sind nicht für den breitenBetriebsbereich von Automobilanwendungen geeignet

In der vorliegenden Arbeit wird deshalb ein mehrstufiges Diagnosekonzeptvorgestellt Dieses ermöglicht einerseits die frühzeitige Erkennung sowie Un-terscheidung von drei relevanten Wicklungs- und zwei typischen Sensorfehlern

in der elektrischen Maschine Andererseits ist es gleichzeitig für einen weitenArbeitsbereich geeignet und robust gegenüber Signalrauschen Zudem kanndiese Methode durch die Kombination von modellbasierten (physikalischen)und datenbasierten Ansätzen ihre Diagnosefähigkeit über die Lebensdauer einesFahrzeugs kontinuierlich verbessern

Um das Konzept zu validieren, werden zunächst die drei Wicklungsfehler –Windungsschluss (TSC), Phasenkurzschluss (PSC) und unsymmetrischer Wick-lungswiderstand (UWR) – analytisch modelliert Nach der Analyse des Ver-haltens der Maschine unter verschiedenen fehlerhaften Betriebsbedingungenwerden die anliegenden Phasenströme der elektrischen Maschine verwendet, umdie Fehler zu erkennen Ein fehlerhaftes Stromsignal könnte jedoch auch durchGerätetoleranzen, Temperaturdrift, Alterung und Rauschen verursacht werden

Trang 24

XXVIII Kurzfassung

und somit zu einer ungewollten Fehldiagnose führen Daher werden auch zweider häufigsten Sensorfehler (Verstärkungs- und Offset-Fehler) modelliert, umdiese von den oben genannten Fehlern in der elektrischen Maschine unter-scheiden zu können Aus den sensorisch erfassbaren Phasenströmen werdenschlussendlich mehrere Merkmale im Zeit- und Frequenzbereich extrahiert undverwendet, um die verschiedenen Fehler eindeutig unterscheiden zu können.Das mehrstufiges Diagnosekonzept besteht aus drei verschiedenen Diagnose-modellen Das erste Modell ist ein physikalisches Modell Dieses Modell kannvor Serienstart bzw Produktion des Fahrzeugs auf Basis der bekannten physika-lischen Zusammenhänge entwickelt werden Sechs Kenngrößen aus dem Zeit-und Frequenzbereich werden hierbei betrachtet und das Modell erreicht einezufriedenstellende Diagnosegüte Allerdings ist diese reduziert, wenn sich dasFahrzeug in nicht-kalibrierten Betriebspunkten befindet oder ein unbekannterFehler auftritt

Das zweite Modell ist ein datenbasiertes Modell zur Überwachung des eigenenGesundheitszustands der betrachteten Komponenten (Self Conditioning Model;SCM) Dabei wird zunächst deren Systemverhalten im gesunden Zustand imjeweiligen Fahrzeug, in das das Antriebssystem verbaut wurde, erlernt, um dannAbweichungen von diesem Sollzustand detektieren zu können Dabei wird derGrad der Anomalie des aktuellen Zustands durch das SCM bewertet Prinzipbe-dingt können auch unbekannte Fehler präzise erkannt werden Um eine hoheDiagnosegüte zu erreichen, benötigt dieser Modellansatz jedoch eine gewis-

se Zeit, um zu Beginn des Fahrzeuglebens in verschiedenen Betriebspunktengenügend Daten zu sammeln

Das dritte Modell ist ein auf Flottendaten, also Fahrzeug übergreifenden Daten,basierendes, neuronales Netzwerkmodell Das Modell kann aktiviert werden,wenn ausreichend Daten, insbesondere Daten zu fehlerhaften Bedingungen,über die Gesamtheit der Fahrzeuge gesammelt wurden Dies bedarf einer deut-lich längeren Datenerfassungszeit im Vergleich zum SCM, kann jedoch insbe-sondere zum Ende der Fahrzeuglebensdauer, wenn aufgrund von Verschleiß dieFehlerhäufigkeit zunimmt, helfen, einen möglichen Fehler früher zu erkennenund damit die Verfügbarkeit des Antriebssystems zu erhöhen

Jedes einzelne Modell hat seine spezifischen Vor- und Nachteile Um die samtdiagnosegüte zu verbessern, wurde ergänzend eine Entscheidungsfunk-

Trang 25

Ge-Kurzfassung XXIX

tion entwickelt, die die Ergebnisse aus verschiedenen Betriebspunkten undDiagnose-Modellen kombiniert und gewichtet Die Entscheidungsfunktion ba-siert auf dem Prinzip des sogenannten Soft-Votings, bei dem jedes einzelneModell nicht nur die eigentliche Fehlerprädiktion, sondern dazu auch das zuge-hörige Vertrauensmaß (Confidence) mitliefern muss Damit wird dann in derZusammenschau der Teildiagnosen der unterschiedlichen Modelle die finaleDiagnoseentscheidung getroffen

Dieses Gesamtdiagnosekonzept wird anhand von numerischen ten validiert, in denen mehrere tausend fehlerhafte Zustände inkl verschiedenerFehlertypen, Schweregrade und Betriebspunkte enthalten sind Es zeigt dabeieine sehr hohe Diagnosegüte sowohl für Wicklungs- als auch für Sensorfehler

Simulationsda-in eSimulationsda-inem weiten Betriebsbereich und zeichnet sich darüber hSimulationsda-inaus auch durcheine hohe Robustheit gegenüber unbekannten Fehlern aus

In Kapitel 5 wird ein Diagnosekonzept für mechanische Fehler einer trischen Maschine vorgestellt Die relevantesten, mechanischen Fehler sindLagerfehler Zunächst werden daher verschiedene Fehlermechanismen vonLagern sowie relevante Veröffentlichungen zur Lagerfehlerdiagnose kurz vor-gestellt Obwohl die Mechanismen von Lagerfehlern im Labor gut untersuchtsind, gibt es immer noch viele Herausforderungen für ihre Diagnose am realenProdukt Die aktuellen Herausforderungen lassen sich wie folgt zusammenfas-sen: begrenzte Daten, begrenzte Zuordnung der Daten zu Fehlern (begrenzteLabel) und Übertragbarkeit von Labor- auf reale Einsatzbedingungen

elek-• Begrenzte Daten

Der Run-to-Failure-Test eines realen Produkts ist in der Regel sehr wändig Es dauert oft mehrere Monate oder Jahre, bis das Produkt seineLebensdauergrenze erreicht Darüber hinaus kann die Datenqualität ineinem realen Produkt durch viele unerwartete Faktoren beeinflusst wer-den Beispielsweise können Folgefehler, die durch den ursprünglichenFehler verursacht wurden, zu Testabbrüchen führen, so dass der eigentli-che, durch den ursprünglichen Fehler bedingte Degradationsprozess nichtbeobachtet werden kann

auf-• Onboard Überwachung des Fehlers

Bei der Lagerfehlererkennung ist es schwierig, genau zu bestimmen, wann

Trang 26

XXX Kurzfassung

die Signale einen ersten Hinweis auf einen Fehler geben und wie lange dieAnfangsphase dauert Dies gilt insbesondere für Fehler, die sich im Laufeder Zeit auf natürliche Weise durch Verschleiß entwickeln Während eineslaufenden Run-to-Failure-Tests ist der Zustand der Lager nicht beurteilbar

Er kann erst nachträglich durch Öffnen des Lagers und anschließenderBefundung ermittelt werden Dabei ist neben der Problematik der zeit-lichen Einordnung auch die Unterscheidung von Erst- und Folgefehlernschwierig Dies erschwert die Zuordnung von Signalauffälligkeiten zuFehlern deutlich

• Übertragbarkeit von Labor- auf reale Einsatzbedingungen

Die komplexen Einsatzbedingungen eines realen Produkts, einschließlichvariierender Betriebspunkte und Umgebungsbedingungen, sowie Sensor-einflüsse durch andere Systeme und Komponenten des Fahrzeugs oderauch Limitierungen bei der Positionierung eines Sensors, treten bei La-gertests unter Laborbedingungen typischerweise nicht auf Gleichzeitigkann der Laboraufbau selbst einen unerwünschten Störeffekt haben, insbe-sondere bei Schwingungsmessungen Im Labor als sinnvoll identifizierteKenngrößen und deren Charakteristika können sich deshalb unter Ein-satzbedingungen als nutzlos erweisen Die Übertragung des erworbenenWissens vom Labor auf ein reales Produkt in seiner nativen Umgebung isteine der größten Herausforderungen

Um zumindest Daten von Lagern in realen Produkten zu erhalten, werden indieser Arbeit Schwingungsmessungen aus Dauerlauftests kompletter Antrieb-sachsen verwendet Bei diesen werden vordefinierte Drehzahlrampen in einemTestzeitraum von ca 150 Stunden so oft wiederholt, bis die Achse ihre Lebens-dauer erreicht hat Die Dauer einer einzelnen Rampe beträgt eine Minute Dabeiwird die Achse zunächst aus dem Stillstand auf ihre Höchstdrehzahl beschleu-nigt und dann wieder bis zum Stillstand abgebremst Das Antriebsdrehmomentbleibt dabei auf einem konstanten, positiven Wert

Aus den Messungen werden verschiedene Merkmale aus dem Zeit- und quenzbereich extrahiert Diese Merkmale werden mit denselben Merkmalen ver-glichen, die aus Messdaten eines Lagerprüfstands im Labor extrahiert werden.Die einzelnen Merkmale beider Datensätze werden hinsichtlich ihrer Unter-

Trang 27

Fre-Kurzfassung XXXI

schiede, Diagnoseeignung und Übertragbarkeit analysiert und bewertet Hierbeizeigt sich, dass einige Merkmale, die sich in den Labortests als vielverspre-chend gezeigt haben, wie die Kurtosis oder die Amplituden lagerspezifischerFrequenzen, nicht für die Datensätze der gesamten Antriebsachse geeignet sind.Diese weisen deutlich mehr störendes Rauschen auf Im Vergleich dazu zeigensich Merkmale aus der Hüllkurvenanalyse Envf t f robust gegenüber Rauschen.Das neue Merkmal Medianspec, das die Amplitude des Hintergrundrauschensbeschreibt, wird in der Arbeit entwickelt Im Vergleich zu anderen betrachtetenMerkmalen, zeichnet sich dieses Merkmal durch eine sehr hohe Sensitivität undMonotonie aus, die es zu einem geeigneten Diagnosekriterium machen.Dies bestätigt sich auch in einer abschließenden Validierung, bei der die dreiaus den Vortests vielversprechendsten Merkmale, Medianspec, Envf t f, undP2P als Indikatoren ausgewählt und anhand von drei weiteren, geschädigtenAntriebsachsen validiert werden Hierbei zeigt nur das Merkmal Medianspec

einen konsistenten Degradationstrend bei allen Achsen, so dass darauf basierendeine Vorhersage des Gesundheitszustands der Lager in Antriebsachsen möglichscheint

Ähnlich zum Diagnosekonzept für die Fehler der elektrischen Maschine wird

in Abschnitt 5.7 abschließend wiederum ein mehrstufiges Diagnosekonzeptvorgestellt In der ersten Ebene dieses Diagnosekonzepts wird die zuvor be-schriebene Hüllkurvenanalyse des Schwingungssignals genutzt Sie ist in derLage, schwere Lagerfehler zu erkennen Auf der zweiten Ebene wird der Ge-sundheitszustand wiederum mithilfe eines unüberwachten Lernalgorithmus (SCM) erfasst Hierdurch kann die Degradation von Achsen bestimmt werden.Die eigentliche Vorhersage des Gesundheitszustands erfolgt durch die dritteEbene unter Verwendung eines Exponentialansatzes zur Berechnung einesGesundheitsindikators

Durch die Kombination der beiden vorgestellten Diagnosekonzepte für trische und mechanische Fehler in einem elektrischen Antriebssystem könnendie relevantesten Fehler prädiktiv diagnostiziert werden Sobald eine Tole-ranzschwelle für die Kritikalität eines Fehlers überschritten wird, kann dasFahrzeug durch entsprechende Betriebsstrategien rechtzeitig in einen sicherenFahr - und Betriebszustand gebracht werden Kritische Fehler, die zum Ausfalldes Antriebssystems führen können, werden dadurch vermieden Daher kann

Trang 28

elek-XXXII Kurzfassung

auf eine redundante Ausführung insbesondere der elektrischen Maschine imAntriebssystem verzichtet werden

Trang 29

1 Introduction

1.1 Motivation

Recently, electric vehicles and automated driving have attracted strong interests

in the field of automotive industry Many researches focus on autonomouselectric vehicles and aim to enhance the performance as well as the reliability ofthese vehicles, which will be used in highly automated driving applications [2,3] These autonomous electric vehicles will run much more over lifetime thantoday’s privately owned vehicles and require much higher availability Moreover,according to the Society of Automotive Engineers (SAE), the automated drivingsystems with automation level 4 and 5 are responsible for executing a fallbackdriving task, which brings the vehicle to a safe state in case of failure [4]

To fulfill this requirement, several critical systems have been designed withredundancy to enhance the reliability of systems, such as perception sensors,computing, braking, steering, and power supply systems [5, 6] However, ifthe powertrain system of a highly automated driving vehicle fails during atrip, it is possible that the vehicle stops in the middle of a traffic lane withoutsufficient protection From the author’s perspective, the vehicle should be able

to drive to a safe location automatically and ensure the safety of passengers

as well as the surroundings, even in the case of a failure in the powertrain.Therefore, the powertrain system of such an autonomous electric vehicle should

be designed with fail-operational ability One solution to fulfill this requirement

is to use a redundant system However, the additional cost and weight ofthe second powertrain system are not favorable An economical solution is toachieve incipient fault detection to prevent the vehicle from dangerous operatingconditions

In today’s battery electric vehicles, different interconnected electrical tems and components, various operating modes, functions and a large number

subsys-of sensors have increased the complexity subsys-of the powertrain system It is an mous challenge to detect and isolate the faults in a complex system Althoughthe On-board diagnostics (OBD) system provides the vehicle owner or repair

enor-© The Author(s), under exclusive license to

Springer Fachmedien Wiesbaden GmbH, part of Springer Nature 2022

T Shen, Diagnosis of the Powertrain Systems for Autonomous Electric Vehicles,

Wissenschaftliche Reihe Fahrzeugtechnik Universität Stuttgart,

https://doi.org/10.1007/978-3-658-36992-7_1

Trang 30

2 1 Introduction

technician the access to the status of a variety of vehicle sub-systems, especiallyelectrical systems, the root causes of many faults in the electric vehicles arestill inconspicuous Moreover, a component easily loses its function when aninternal fault is detected This kind of system is called fail-safe system, whichcannot fulfill the requirement of a fail-operational system According to [7], up

to 50 % of software code in a vehicle can be devoted to diagnostic functions, butrobust diagnostic systems, which are capable of monitoring the state of health

of the components, are especially desirable in electric vehicles Nevertheless,

an early detection of faults is not yet available due to the limited performance

of today’s diagnostic system

Last but not the least, for a privately owned vehicle, drivers or repair techniciancould feel, hear, see, or smell a special form of vibration, noises, colors, smells,etc Based on corresponding past experience, so in practice, they use the qualita-tive information to percept an incipient fault in the vehicle This kind of methodsare known as heuristic knowledge based methods [8] In an autonomous electricvehicle, the driver may be not always available, so the heuristic knowledgefrom the driver is limited Therefore, an advanced diagnostic system to detectincipient faults in the powertrain system for autonomous electric vehicles arenecessary

1.2 Objectives

The main goal of this thesis is to develop an intelligent diagnostic system

to detect incipient faults in the powertrain systems for autonomous electricvehicles Based on this target, the following objectives are set:

1 Analyzing the critical faults, which lead to the malfunction of powertrainsystems

2 Investigating the challenges induced by the gap between today’s diagnosticsystem and the desired one

3 Developing the diagnostic concept capable of detecting incipient faultsfocusing on a specific component

Trang 31

1.4 Organization of the Thesis 3

1.3 Contributions of the Thesis

The contributions of the thesis can be summarized as follows:

1 Comprehensive analysis of the critical faults in today’s electrical powertrainsystem

2 Evaluation of the feasibility of a predictive diagnostic system for maincomponents in the powertrain systems

3 Development of a predictive diagnostic concept for the electrical faults inelectric machines

4 Development of a predictive diagnostic concept for the mechanical faults inelectric machines

1.4 Organization of the Thesis

The rest of the thesis is organized as follows In Chapter 2, the most criticalfaults in the powertrain system of a battery electric vehicle are analyzed Theroot causes and failure mechanisms of each subsystem are investigated Afterevaluating the feasibility of constructing a predictive diagnostic system, thefaults in electric machine are focused In Chapter 3, the background of relevantdiagnosis, signal processing techniques and machine learning algorithms areintroduced briefly Then, the diagnostic concept for the electrical faults in theelectric machine is proposed in Chapter 4 The concept is able to identifyvarious electrical faults at different operating points The performance of thediagnostic method can be increased over the lifetime because of its learningability In Chapter 5, the mechanical faults of electric machines are focused Thevibration data from a industry product, eAxle, is used to conduct fault detection.With proper signal processing, effective features for the fault detection areextracted from the raw data Then, a multi-level diagnostic concept is developed

to monitor the health state of the eAxle Finally, conclusions and future workare included in Chapter 6

Trang 32

2 Fault Analysis

The powertrain system of a battery electric vehicle (BEV) consists of a batterysystem, inverter, electric machine, gearbox and other mechanical parts Thepowertrain will be shut down if any subsystem or component fails In thischapter, the failure mechanisms of each subsystem are summarized Further-more, various diagnostic methods of each fault in the state of the art are brieflyinvestigated As the critical faults are desired to be predicted at an early stage,the feasibility of corresponding predictive diagnosis is evaluated, and finally,the common faults in the electric machine are selected as the focus of the thesis

2.1 Battery

A battery system consists of numerous interconnected battery cells, differentvoltage, current, and temperature sensors as well as battery management system(BMS) The battery safety reports [9] and [10] summarized 28 and 41 electricvehicle incidents due to battery fire until 2019 According to these reports,the most common fault (over 90 %) of today’s battery system is short circuit.Mechanical abuse e.g damage [11], electrical abuse e.g water immersion [12]and cell internal short circuit [13] are the main causes of short circuit Figure 2.1illustrates the failure mechanism of a battery fire

Lots of researchers focused on the detection of short circuit [14, 15, 16, 17]based on existing cell voltage, current measurements and equivalent batterymodel

Besides, some approaches focused on the failure mode and diagnosis of sensorfaults in the battery [18, 19, 20] The reason for that is, a sensor fault can lead

to an erroneous estimation of battery status and the electric vehicles have to beshut down despite the fact that the battery cells still work in normal condition

© The Author(s), under exclusive license to

Springer Fachmedien Wiesbaden GmbH, part of Springer Nature 2022

T Shen, Diagnosis of the Powertrain Systems for Autonomous Electric Vehicles,

Wissenschaftliche Reihe Fahrzeugtechnik Universität Stuttgart,

https://doi.org/10.1007/978-3-658-36992-7_2

Trang 33

6 2 Fault Analysis

Figure 2.1: Failure mechanism of a battery fire [21]

Hence, the sensor faults also need to be detected and distinguished from cellfailure to enhance the reliability of battery system

2.2 Inverter

Inverter plays an important role in the automotive electric drive system Authors

of [23] summarized the results of various reports and concluded that more than

30 % of all the faults in variable-speed motor drives were caused by the failure

of power electronics Figure 2.2 shows the result from an industry-based survey,

in which 31 % of the responders believed that semiconductor power devices(power electronics) are the most fragile components in their applications [22]

Trang 34

2.2 Inverter 7

Figure 2.2: Distribution of fragile components [22]

Several publications have covered the topics on fault diagnosis of inverter Thefault diagnosis of an insulated-gate bipolar transistor (IGBT) is the most popularone Diagnostic methods of the two main faults of IGBT in various invertertopologies, short circuit faults [24, 25] and open circuit faults, [24, 26, 27] havebeen well investigated In addition, several researchers focused on the agingeffect of IGBT [24, 28] Moreover, different concepts to monitor the healthycondition of IGBT were developed under laboratory conditions [29, 30, 31].Besides, capacitor is also a reliability-critical component in inverter Due totheir aging effect, the capacitance of capacitors reduces gradually, while theequivalent series resistance increases Consequently, the temperature of capaci-tor increases, resulting in a decline in the performance of the inverter Typicalsolution in the industry to avoid such effect is to overdesign the capacitors,which increases the material cost Alternatively, the condition monitoring tech-nologies of capacitors have been investigated to prevent the critical fault in [31].However, the additional costs for the monitoring system, e.g sensors, plugshave to be taken into account

Some other researches looked at sensor faults Analogous to sensor faults in thebattery system, an erroneous sensor may cause a false alarm and stop the vehicleunexpectedly [32] and [33] explained two diagnostic concepts to detect and

Trang 35

8 2 Fault Analysis

identify current sensor faults by using average normalized current and stationaryreference frame In [34], a model-based diagnostic method for current sensor,voltage sensor, and position sensor in the drive system was introduced Withproper adaption in the control system, the influence of sensor faults can becompensated [32]

2.3 Electric Machine

An electric machine can fail due to diversified causes The typical faults andtheir occurrence distribution according to a recent reliability study of inductionmachines (IM) is presented by Figure 2.3 It is clear that bearing (51 %) is themost critical component for most industrial applications, while stator windingfaults are on the second place with 16 % of total breakdown in study [35] Alt-hough this study focused on IM, this distribution of faults can be used as a goodreference for the commonly used machine in the electric vehicle: permanentmagnet synchronous machine (PMSM) The authors of [7] analyzed the rootcauses of different failure modes and effects in IM and PMSM Especially forbearing, 12 failure modes were summarized in [36], such as excessive load,and lubricant failure In practice, a defective bearing may be caused by multi-ple causes One fault may initiate other faults during the degradation process

A “self-healing” effect of bearing was observed in [37] In this case, the itial defect was cracks, which caused increasing vibration After some time

in-of operation, the damaged surface was smoothed through rolling contact Itwas reflected as an “increase-decrease” degradation trend When the damageachieved a sufficiently conspicuous extent, the amplitude of vibration started

to increase again These complex fault mechanisms bring a great challenge

to bearing fault diagnosis Recently, fault diagnosis of bearing have attractedmore and more attention from academia and industry Various approaches havebeen proposed with focuses of extracting health indicators [38, 39], monitoringdegradation processes [37, 40], and prediction of remaining useful life [37, 41].Moreover, winding faults have been investigated by several works According

to [42], the root causes of winding faults are thermal stress, electrical stressand mechanical stress An increase in temperature significantly accelerates the

Trang 36

2.3 Electric Machine 9

Figure 2.3: Distribution of failed components in electric machines [35]

aging of the insulation The insulation lifetime decreases by 50 % when thetemperature increase by 10 °C above the maximum allowable temperature Theelectrical stress caused by the dielectric material, the phenomena of trackingand corona, and the high transient voltages For the electric machines used inthe powertrain system, the inverters with high switching voltage and switchingfrequency can cause high transient voltage and curtail the lifetime of insulationmaterial In addition, the coil movement and strikes from the rotor can alsocause mechanical stress and lead to winding faults in the end The most commontechniques to detect winding faults are surge test and partial discharge test, butboth of them are offline tests Recently, various online monitoring methods havebeen developed using different physical quantities e.g temperature monitoring,leakage currents and zero sequence voltage In [42], several online monitoringmethods were summarized

Trang 37

10 2 Fault Analysis

2.4 Gearbox

Gearbox is an important mechanical component in the electric drive systems.Failure occurrence in the gearbox results in a performance degradation, whichaffects the efficiency of the powertrain and increases the energy consumption

In critical cases, the gearbox can block the wheel shaft, which would lead

to severe consequences Many methods have been proposed for gearbox dition monitoring, such as vibration-, current-, acoustic emission-, oil-, andtemperature-based approaches [43, 44, 45] Vibration-based approaches havebeen proven to be effective for gearbox fault diagnosis by using appropria-

con-te signal processing con-techniques, e.g Fourier analysis, envelope analysis, andwavelet transformation (WT) However, vibration-based approaches are intru-sive because they require the installation of additional sensors in the gearbox.Current-based methods are non-intrusive But it might be a challenge to extractuseful information from current signals for gearbox fault diagnosis [46]

2.5 Scope of the Thesis

In the previous sections, the most critical faults in the powertrain system ofautonomous electric vehicles are summarized, and the state of the art faultdiagnosis are investigated briefly Since the target of this study is to enhance theperformance of current diagnostic methods and achieve predictive diagnosis ofcritical faults, the faults are individually evaluated and justified for predictivediagnosis According to [47], the following criteria are considered:

• Probability of the faults

• Economical feasibility of a predictive diagnostic system compared to othermethods, e.g redundancy and preventive maintenance

• Technical feasibility of a predictive diagnostic system incl diagnosticcoverage, reaction time, and practicality

• Organizational feasibility, e.g competence of existing knowledge levels.Finally, electric machine is selected as the candidate for the predictive diagno-stic concept and the most common faults, namely bearing and stator winding

Trang 38

2.5 Scope of the Thesis 11

faults are focused in this thesis The structure of electric machines is wellknown, and its fault behaviors can be well predicted with the help of practicalresources Compared to the faults of power electronics, the faults in the electricmachine develop much more slowly When a fault is detected, proper reactions,e.g warning, degradation and maintenance can be implemented within a certaintime On the other hand, battery system and inverter are ruled out Concerningthe battery system, two particular features are notable to limit the prospect ofapplying predictive diagnosis First, a low cost redundancy design can be achie-ved by connecting battery cells in parallel as shown in [48] Second, advancedknowledge in battery electrochemistry is needed for proposing a predictive dia-gnostic concept Neither is inverter a good candidate for the predictive diagnosis,because the most fragile parts in the inverter, power electronics, always fail inseveral seconds as soon as a fault occurs In other words, allowable reactiontime is probably too short to perform predictive diagnosis Furthermore, voltagemeasurement is required to be sufficiently precise to achieve an acceptableaccuracy of the diagnosis Nevertheless, the fault detection methods of gearboxfaults are analogous to bearing faults in the electric machine, solely the electricmachine faults are focused in the thesis

Trang 39

3 Background and State of the Art

In this chapter, the background of fault diagnostic methods are firstly ced, followed by introduction of signal processing techniques used for featureengineering In the end, the principles of machine learning algorithms employed

introdu-in the present study are briefly explaintrodu-ined

3.1 Fault Diagnostic Methods

In general, the diagnostic methods in today’s vehicles can be divided into twoparts: the analytical knowledge and heuristic knowledge based methods Theanalytical knowledge based methods use the measured signals, variables, andstates in the vehicle to produce quantifiable, analytical information Since theintroduction of OBD in 1969, many safety related faults can be detected automa-tically The OBD system gives vehicle owner or repair technician access to thestatus of various vehicle sub-systems, especially electrical systems However,there are still some faults in the vehicle requiring human expert knowledge.Driver or repair technicians may feel, hear, see, or smell a special form ofvibration, noises, colors, smells, etc based on their experience They use thequalitative information to detect a fault in the vehicle This kind of methodsare called heuristic knowledge based methods [49] In an autonomous electricvehicle, the driver may not always be available, so the heuristic knowledge ofthe driver is limited To achieve the same performance of diagnosis, extendedanalytical knowledge methods are required In the following subsections, threekinds of analytical knowledge based diagnostic methods based on [49] areexplained

3.1.1 Signal-based Fault Detection Methods

A signal-based fault detection method uses the measured variables to monitorthe current condition of a process Figure 3.1 presents the scheme of a typical

© The Author(s), under exclusive license to

Springer Fachmedien Wiesbaden GmbH, part of Springer Nature 2022

T Shen, Diagnosis of the Powertrain Systems for Autonomous Electric Vehicles,

Wissenschaftliche Reihe Fahrzeugtechnik Universität Stuttgart,

https://doi.org/10.1007/978-3-658-36992-7_3

Trang 40

14 3 Background and State of the Art

Figure 3.1: Scheme of a signal-based fault detection method

signal-based fault detection method Some observed variables of the processare used as the inputs of the signal model that contains special mathematicalformulations and calculates relevant features from the measured variables Bylooking into the change of such features, the fault of the process can be detected.Limit checking and plausibility checking are the two basic signal-based faultdetection methods, which are widely used in the automotive branch because ofits simplicity and effectiveness In practice, all safety related signals have to becompared with other signals from independent sources to ensure the plausibility

of the signals In addition, Fourier analysis, correlation analysis and waveletanalysis are also used in signal models.Signal-based fault detection methods areusually effective for detecting critical faults that can change the behavior of theprocess dramatically However, the usage is limited under transient conditionsand for incipient faults

3.1.2 Model-based Fault Detection Methods

Model-based detection methods aim to describe the dependency between inputand output signals by mathematical equations Figure 3.2 presents the scheme

of a typical model-based fault detection method As shown in Figure 3.2, theinputs of the process are also given to a process model, which simulates thisprocess virtually and predict the outputs simultaneously The outputs of theprocess model will be compared to the measured values The discrepancy isanalyzed to extract suitable features subsequently These features can be used

to detect fault The modeling of the process model can be divided into threecategories The theoretical model contains the functional description betweeninputs and outputs with physical model and its parameters This kind of models

is called white box models On the other hand, the experimental models, whichare also called black box models, assume that the structure of the model is onlybased on measured data The grey box models are those between these two

Ngày đăng: 18/03/2023, 07:29

TÀI LIỆU CÙNG NGƯỜI DÙNG

TÀI LIỆU LIÊN QUAN

🧩 Sản phẩm bạn có thể quan tâm

w