Non-Destructive Testing and Condition Monitoring Techniques in Wind Energy

By Fausto Pedro Garcia Marquez and Mayorkinos Papaelias

Non-Destructive Testing and Condition Monitoring Techniques in Wind Energy looks at the complex and critical components of energy assets and the importance of inspection and maintenance to ensure their high availability and uninterrupted operation. Presenting the main concepts, state-of-the-art advances and case studies, this book approaches the topic by considering it as an integral part of the overall operation of any wind energy project. Linking the essential NDT subject with its sub disciplines, the book uses computational techniques, dynamic analysis, probabilistic methods, and mathematical optimization techniques to support analysis of prognostic problems with defined constraints and requirements.

This book is the first of its kind and will provide useful insights to industrial engineers and scientists, academics and students in the possibilities that NDT and condition monitoring technologies can offer.

• Presents advances in Non-Destructive Techniques and Condition Monitoring Systems applied in the energy industry
• Provides case studies in Fault Detection and Diagnosis and Prognosis for critical variability
• Offers technical maintenance actions for the observation and analyses of inspection, monitoring, testing, diagnosis, prognosis and active maintenance actions in wind

• Language: English
• Publisher: Academic Press
• Release date: Jun 24, 2023
• ISBN: 9780323951005

Book preview

Non-Destructive Testing and Condition Monitoring Techniques in Wind Energy

Edited by

Fausto Pedro Garcia Marquez

University of Castilla-La Mancha, Ciudad Real, Spain

Mayorkinos Papaelias

University of Birmingham, Birmingham, United Kingdom

Valter Luiz Jantara Junior

School of Metallurgy and Materials, University of Birmingham, Birmingham, United Kingdom

Table of Contents

Cover image

Title page

Copyright

List of contributors

Biographies

Foreword

Preface

Introduction to non-destructive testing and condition monitoring techniques in wind energy

Background and purpose of the book

Target audience

1. SCADA-based fault detection in wind turbines: data-driven techniques and applications

1. SCADA-based condition monitoring

2. Early fault detection with normal behavior models and SCADA data

3. Case study: early fault detection in gear bearings with multi-target normal behavior models and SCADA data

2. Fault detection in small wind turbines using condition monitoring techniques and machine learning algorithms (a predictive approach)

1. Introduction

2. Related works

3. Condition monitoring techniques

4. System design

5. Machine learning algorithms

6. Implementation

7. Results and discussion

8. Conclusion

3. Prediction and classification of different wind turbine alarms using K-nearest neighbors

1. Introduction

2. Approach

3. Case study and results

4. Conclusions

4. Artificial neural networks applied for wind turbines maintenance management in unmanned aerial vehicle acoustic inspection case

1. State of the art

2. Wind turbines acoustic inspection case

3. Artificial neural networks application

4. Results

5. Conclusions

5. Quantitative interlink wear estimation method for the mooring chain

1. Introduction

2. Proposed interlink wear estimation method

3. Application to a spar-type floating structure moored with three catenary lines

4. Wear estimation considering the rolling motion between adjacent chain links

5. Conclusions

Appendix A: Newly designed wear test machine possible to use mooring chain surface under high pressure

6. Enhanced sparse representation-based intelligent recognition framework for fault diagnosis of wind turbine drive trains

1. Introduction

2. Related works

3. Enhanced sparse representation-based intelligent recognition framework

4. Experiment validations

5. Conclusions

7. An optimal combined production and maintenance policies for a wind farm with environmental and operational considerations

1. Introduction

2. Literature review

3. Problem setting and notations

4. Description of production and maintenance policies

5. Mathematical formulation of the production strategy

6. Mathematical formulation of maintenance strategies

7. Optimization algorithms

8. Numerical examples

9. Conclusion

8. The valuation of geothermal power projects in Indonesia using real options valuation

1. Introduction

2. Literature review

3. Methodology

4. Result and discussion

5. Conclusion

9. A robust multiple open-switch fault diagnosis approach for converter in wind energy system

1. Introduction

2. A brief literature review and motivation

3. System description and fault analysis

4. Fault diagnostic method

5. Simulation result and analysis

6. Conclusion

7. Funding statement

10. Condition monitoring in wind turbines: a review

Abbreviations

1. Introduction

2. Requirements for condition monitoring in WECS

3. Features extracted for condition monitoring in WECS

4. Technologies for condition monitoring in WECS

5. Research challenges

6. Conclusion

11. Artificial intelligence techniques and cloud computing for wind turbine pitch bearing fault detection

1. Introduction

2. Temporal convolutional augmented Bayesian search (TCABS)

3. Discussion for TCABS

4. Fault diagnosis

5. Simulation

6. Experiments

7. Damage validation

8. Conclusion

12. Alarms and false-alarm analysis by support vector machine in wind turbines

1. Introduction

2. Approach

3. Case study and results

4. Conclusions

13. Background, advancement, and applications of in situ structural health monitoring based on different modes of failure detection in composites: a review

Abbreviations

1. Introduction

2. Nanomaterials and in situ SHM

3. Applications of advanced in-situ SHM in the detection of specific modes of failures in composites

4. Computational modeling and in-situ SHM

5. Conclusion

Data availability statement

Conflicts of interest

14. Numerical simulations of offshore wind farms considering accidental scenarios

Abbreviations

1. Introduction

2. Scenarios of accidents and accidental loads for offshore wind farms

3. Numerical simulations of offshore wind turbines with internal accidents

4. Numerical simulations of wind turbines with external accidents

5. Concluding remarks and future outlook

15. Multibody dynamic analysis of onshore horizontal-axis wind turbine

1. Introduction

2. Overview of the chapter

3. Mathematical model

4. Numerical simulation

5. Conclusion

6. Appendix: coordinate transformation

16. Foundation monitoring system of offshore wind turbines

1. Introduction

2. Numerical simulation of scouring effect on dynamic characteristics of offshore wind turbine

3. Numerical simulation of scouring effect on dynamic responses of offshore wind turbines

4. Real-time monitoring system for scouring around monopile foundations of offshore wind turbines

5. Real-time uneven settlement and stress monitoring systems for jacket-foundation offshore wind turbines

6. Conclusion

Index

Copyright

Academic Press is an imprint of Elsevier

125 London Wall, London EC2Y 5AS, United Kingdom

525 B Street, Suite 1650, San Diego, CA 92101, United States

50 Hampshire Street, 5th Floor, Cambridge, MA 02139, United States

The Boulevard, Langford Lane, Kidlington, Oxford OX5 1GB, United Kingdom

Copyright © 2023 Elsevier Inc. All rights reserved.

No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopying, recording, or any information storage and retrieval system, without permission in writing from the publisher. Details on how to seek permission, further information about the Publisher’s permissions policies and our arrangements with organizations such as the Copyright Clearance Center and the Copyright Licensing Agency, can be found at our website: www.elsevier.com/permissions.

This book and the individual contributions contained in it are protected under copyright by the Publisher (other than as may be noted herein).

Notices

Knowledge and best practice in this field are constantly changing. As new research and experience broaden our understanding, changes in research methods, professional practices, or medical treatment may become necessary.

Practitioners and researchers must always rely on their own experience and knowledge in evaluating and using any information, methods, compounds, or experiments described herein. In using such information or methods they should be mindful of their own safety and the safety of others, including parties for whom they have a professional responsibility.

To the fullest extent of the law, neither the Publisher nor the authors, contributors, or editors, assume any liability for any injury and/or damage to persons or property as a matter of products liability, negligence or otherwise, or from any use or operation of any methods, products, instructions, or ideas contained in the material herein.

ISBN: 978-0-323-99666-2

For information on all Academic Press publications visit our website at https://www.elsevier.com/books-and-journals

Publisher: Megan Ball

Acquisitions Editor: Edward Payne

Editorial Project Manager: Ali Afzal-Khan

Production Project Manager: Fizza Fathima

Cover Designer: Greg Harris

Typeset by TNQ Technologies

List of contributors

Pedro José Bernalte Sánchez,     Ingenium Research Group, Universidad Castilla-La Mancha, Ciudad Real, Spain

Reza Amali Bilqist,     Industrial Engineering Department, Faculty of Engineering, Universitas Indonesia, Depok, Indonesia

Maryem Bouzoubaa,     Laboratoire d’Innovation Durable et de Recherche Appliquée, International University of Agadir, Bab Al madina, Quartier Tillila, Agadir, Morocco

Arunasis Chakraborty,     Department of Civil Engineering, Indian Institute of Technology Guwahati, Guwahati, Assam, India

Fulei Chu,     State Key Laboratory of Tribology, Department of Mechanical Engineering, Tsinghua University, Beijing, China

Muhammad Dachyar,     Industrial Engineering Department, Faculty of Engineering, Universitas Indonesia, Depok, Indonesia

G. Edwin Prem Kumar,     Department of Information Technology, Sri Krishna College of Engineering and Technology, Coimbatore, Tamil Nadu, India

Farizal,     Industrial Engineering Department, Faculty of Engineering, Universitas Indonesia, Depok, Indonesia

Fausto Pedro García Márquez

Ingenium Research Group, Universidad Castilla-La Mancha, Ciudad Real, Spain

ETSI Industrial, University of Castilla-La Mancha, Ciudad Real, Spain

Koji Gotoh,     Department of Marine Systems Engineering, Kyushu University, Fukuoka, Japan

Zied Hajej,     University of Lorraine, LGIPM, Metz, France

Zhiyu Jiang,     Department of Engineering Sciences, University of Agder, Grimstad, Norway

Mohamed Ali Kammoun,     University of Lorraine, LGIPM, Metz, France

Yun Kong

School of Mechanical Engineering, Beijing Institute of Technology, Beijing, China

State Key Laboratory of Tribology, Department of Mechanical Engineering, Tsinghua University, Beijing, China

Khalid Lafdi,     Department of Mechanical and Construction Engineering, Northumbria University, Newcastle Upon Tyne, United Kingdom

Jinping Liang,     School of Astronautics, Northwestern Polytechnical University, Xi'an, PR China

M. Lydia,     Department of Mechatronics Engineering, Sri Krishna College of Engineering and Technology, Coimbatore, Tamil Nadu, India

Angela Meyer,     Bern University of Applied Sciences, Bern, Switzerland

Arka Mitra,     Department of Civil Engineering, Indian Institute of Technology Guwahati, Guwahati, Assam, India

Ana María Peco Chacón,     Ingenium Research Group, Universidad Castilla-La Mancha, Ciudad Real, Spain

Yumna Qureshi,     Institute of Space Technology, Islamabad, Pakistan

Nidhal Rezg,     University of Lorraine, LGIPM, Metz, France

Saptarshi Sarkar,     Chalmers University of Technology, Gothenburg, Sweden

Iku Sato,     Ocean Renewable Energy Division, TODA CORPORATION, Tokyo, Japan

Isaac Segovia Ramírez,     Ingenium Research Group, Universidad Castilla-La Mancha, Ciudad Real, Spain

Takaaki Takeuchi,     Department of Navel Architecture and Ocean Engineering, Osaka University, Suita, Japan

Dongyue Tang

College of Civil Engineering, Tongji University, Shanghai, China

Shanghai Research Institute of Building Science Co., Ltd., Shanghai, China

Mostapha Tarfaoui

ENSTA Bretagne, IRDL-UMR CNRS 6027, Brest, France

Green Energy Park (IRESEN/UM6P), Benguerir, Morocco

Zhuang Tian,     School of Astronautics, Northwestern Polytechnical University, Xi'an, PR China

Tomoaki Utsunomiya,     Department of Marine Systems Engineering, Kyushu University, Fukuoka, Japan

Amrit Shankar Verma,     Department of Mechanical Engineering, The University of Maine, Orono, ME, United States

S. Yoganand,     School of Computer Science and Engineering, VIT University, Vellore, Tamil Nadu, India

Chao Zhang,     Department of Electrical and Electronic Engineering, The University of Manchester, Manchester, United Kingdom

Long Zhang,     Department of Electrical and Electronic Engineering, The University of Manchester, Manchester, United Kingdom

Ming Zhao,     College of Civil Engineering, Tongji University, Shanghai, China

Daming Zhou,     School of Astronautics, Northwestern Polytechnical University, Xi'an, PR China

Biographies

Prof. Fausto Pedro Garcia Márquez

Fausto Pedro García Márquez has been a Full Professor at the University of Castilla-La Mancha (UCLM), Spain, since 2013, is an Honorary Senior Research Fellow at Birmingham University, UK, and a lecturer at the Postgraduate European Institute. He was a Senior Manager at Accenture from 2013 to 2014 and is currently the Director of www.ingeniumgroup.eu. He obtained his European Ph.D. with maximum distinction. Since 2021, he has been a Senior Member at IEEE and an Honored Honorary Member of the Research Council of Indian Institute of Finance. He became the Committee Chair of the International Society for Management Science and Engineering Management (ISMSEM) in 2020. Among his awards are the Nominate Prize (2022), Grand Prize (2021), Runner Prize (2020) and Advancement Prize (2018), Runner (2015), Advancement (2013), and Silver (2012) from the International Society of Management Science and Engineering Management (ICMSEM), and the First International Business Ideas Competition 2017 Award (2017). He has published more than 200 papers and is the author or editor of more than 45 books, over 100 international chapters, and 6 patents. He has been a principal investigator in 4 European projects, 8 national projects, and more than 150 projects for universities and companies. His main interests are artificial intelligence, maintenance, management, renewable energy, transport, advanced analytics, and data science.

Dr. Valter Luiz Jantara Junior

Dr. Valter is a postdoctoral research fellow at the School of Metallurgy and Materials at the University of Birmingham, UK. He obtained his Ph.D. in 2019 having studied damage mechanics and condition monitoring of wind turbine gearbox materials. His main research interests are damage modeling, finite element analysis, failure analysis, surface engineering, condition monitoring, and advanced material characterization. He has been distinguished with the Len Gelman Award for the best paper by a person in the early stages of their career, in the Proceedings of the First World Congress on Condition Monitoring, 2017. He is a member of the Institute of Materials, Minerals and Mining (IOM3). He is experienced in the following areas: electron (including FEG, EBSD, EDS) and optical microscopy; tensile and fatigue testing; tribological testing; plasma nitriding, including DC and active screen; vibration analysis; acoustic emission monitoring; and modeling of materials.

Prof. Mayorkinos Papaelias

Professor Mayorkinos is has done his Ph.D. in Metallurgy (University of Birmingham). Mayorkinos leads the research activity in Non-Destructive Testing and Structural Health Condition Monitoring at the Birmingham Railway Centre for Research and Education and conducts research in structural health condition monitoring of wind turbine towers, and advanced condition monitoring of wind turbine gearboxes and rotating machinery. He served as a technical consultant to TWI, ENGITEC, Innovative Technology and Science Ltd., and Instituto de Soldadura e Qualidade. He is the editor of two books on fault detection and condition monitoring and has contributed chapters to books in fault detection and rail inspection. Mayorkinos is the chairman of the Education Committee of the International Society for Condition Monitoring of the British Institute of Non-Destructive Testing.

Foreword

The energy sector is undergoing an important transition, to meet the world's energy needs, to reduce greenhouse gas emissions, to support climate change mitigation, and to protect our planet for future generations. A major element of the energy transition is the integration of large-scale renewable energy sources, including wind energy, in the energy mix of countries.

For millennia, humanity has harnessed the wind for mobility with sailboats, while the wind has been used for centuries to perform mechanical work using large surfaces that capture the kinetic energy of the passing air flow. More recently, wind power has achieved technological maturity and, along with solar energy and bioenergy, wind energy is playing a significant role in the energy transition.

In support of the global research community and the commercial wind energy sector, the Wind Energy Engineering series publishes research and application-oriented book titles in the overarching subjects related to wind energy engineering, with a focus on scientific and technical content that supports all stages of research and application.

Building on the research and editorial expertise of Profs. Fausto Pedro Garcia Marquez and Mayorkinos Papaelias, along with the participation of an early career researcher, Dr. Valter Luiz Jantara Junior, the editors have brought together over 30 authors to produce this timely book which is specific to an area of growing interest in the wind energy sector.

As in every large-scale energy production system, reliability of wind power plants is a critical aspect for operations, while protection of the operating assets ensures the long-term viability of the wind power plant. Nondestructive testing and condition monitoring techniques contribute to achieve reliability and sustainability in various industrial sectors. Non-Destructive Testing and Condition Monitoring Techniques in Wind Energy adapts and compiles these concepts in a single volume, specifically for wind energy applications.

Covering leading edge issues, such as advances in mathematics, modeling, computational techniques, and dynamic analysis, along with case studies, this book is of interest for all researchers and industrial practitioners interested in the reliability of wind power operations and the long-term viability of wind energy assets.

Prof. Yves Gagnon PEng, DSc

Université de Moncton, Canada

Series Editor, Wind Energy Engineering Series

Preface

Fausto Pedro Garcia Márquez ¹ , Mayorkinos Papaelias ² , and Valter Luiz Jantara Junior ² ,      ¹ Ingenium Research Group, Castilla-La Mancha University, Ciudad Real, Spain,      ² School of Metallurgy and Materials, University of Birmingham, Birmingham, United Kingdom

The literature available on Non-Destructive Testing (NDT) and Condition Monitoring for Renewable Energy sources is very fragmented. Most of the relevant literature sources on the subject can be found in the form of conference and journal papers. To our knowledge, there is no single source of literature that focuses solely NDT and Condition Monitoring applications and techniques in renewable energy. Therefore, this book is quite timely and will face very little competition. Nonetheless, it is expected that once it is published a series of books will follow, but these are more likely to be more focused on individual NDT techniques or renewable energy sources. Therefore, the proposed book will remain of interest and a central source of reference to a significantly wide audience for several years after it has been published.

In 2019 was published the book entitled Non-Destructive Testing and Condition Monitoring Techniques for Renewable Energy Industrial by Mayorkinos Papaelias (Author), Fausto Pedro Garcia Marquez (Author), and Alexander Karyotakis (Author), where two of the authors, Mayorkinos and Fausto, are the same authors to the present book. The main contribution of this book with regards to the mentioned one is that this one is focused only on wind energy and not to any renewable energy industry.

This book presents the main concepts, the state of the art, advances, and case studies in Non-Destructive Techniques and Condition Monitoring Systems applied to energy industry. Fault detection, diagnosis, and prognosis are being a critical variable in the industry in order to reach the competitiveness. Therefore, a correct management of the corrective, predictive, and preventive politics in any industry is required. This book shows content complementary to other subdisciplines such as economics, finance, marketing, decision and risk analysis, engineering, etc.

This book comprises real case studies in multiple disciplines. It considers the main topics as prognostic and subdisciplines. It is essential to link these topics with financial, schedule, resources, downtimes, etc., in order to increase productivity, profitability, maintainability, reliability, safety, availability, and reduce costs and downtimes.

Advances in mathematics, models, computational techniques, dynamic analysis, etc., are employed in prognostic, where this book presents the most important.

Computational techniques, dynamic analysis, probabilistic methods, and mathematical optimization techniques are expertly blended to support analysis of prognostic problems with defined constraints and requirements.

Introduction to non-destructive testing and condition monitoring techniques in wind energy

Fausto Pedro Garcia Márquez ¹

Mayorkinos Papaelias ²

Valter Luiz Jantara Junior ²

¹ Ingenium Research Group, Castilla-La Mancha University, Ciudad Real, Spain      ² School of Metallurgy and Materials, University of Birmingham, Birmingham, United Kingdom

Background and purpose of the book

The economic significance of wind energy sources has increased profoundly over the last 2 decades. Wind energy assets, mainly in the form of wind turbines, have become an important part of the global economy with their net worth value exceeding several hundreds of billions of Euros. Investments in renewable energy are expected to continue to grow until at least 2030, before beginning to stabilize.

Wind energy assets are complex systems with several critical components. Therefore, it is of paramount importance that they are inspected and maintained adequately in order to ensure their high availability and uninterrupted operation. Unexpected failures and faults can result in the underperformance of a particular renewable energy project which may compromise its financial success and endanger future investment. Therefore, the wind energy industry requires reliable maintenance management to ensure the operations of the engines, components, structures, etc., to guarantee the energy supply to industry and society. Any failure, i.e., termination of the ability of an item to perform a required function, generates downtimes, costs, risks for the workers, etc. The high competitiveness in the current industry does not lead these failures to this industry.

The advances in information and communication systems, together with the technologies, lead the industry to incorporate new condition monitoring systems and nondestructive techniques. They require also advanced analytics in order to format, save, and analyze the signals and information, from qualitative and quantitative points of view.

In order to reduce the failures occurrence probability, a correct maintenance task is required. British Standard BS EN-13306:2017 defines maintenance as managerial actions during the life cycle of an item intended to retain it in, or restore it to, a state in which it can perform the required function. Technical maintenance actions include observation and analyses of the item state (e.g., inspection, monitoring, testing, diagnosis, prognosis, etc.) and active maintenance actions (e.g., repair, refurbishment). The correct maintenance support to a maintenance organization to carry out the correct tasks is called maintenance supportability.

Nondestructive testing (NDT) and condition monitoring should be considered as an integral part of the overall operation of any wind energy project ensuring its reliable and safe operation and enabling maintenance actions to be carried out effectively and efficiently resulting in significant financial savings for both manufacturers and operators. This book is the first of its kind and will provide a useful insight to industrial engineers and scientists, academics, and students into the possibilities that NDT and condition monitoring technologies can offer. It will also provide a useful reference to researchers working in the field.

Target audience

This book is intended for professionals in the field of engineering, information science, mathematics, economists, and researchers who wish to develop new skills in Testing NDTs, or who employ the NDT discipline as part of their work. The authors of this volume describe their original work in the area or provide material for cases and studies successfully applying the NDT discipline in real-life cases and theoretical approaches.

The book is also of interest for operators of wind energy systems, where they will find new approaches to apply to their problems to be solved, and similar cases studies to have them as reference.

1: SCADA-based fault detection in wind turbines

data-driven techniques and applications

Angela Meyer     Bern University of Applied Sciences, Bern, Switzerland

Abstract

The installed wind power capacity is growing strongly in many countries worldwide while the profit margins for wind energy are decreasing in the markets. Many wind farm operation managers are seeking to further reduce the operation and maintenance cost of their turbines. Remote condition monitoring and data science techniques enable round-the-clock diagnoses of the wind turbines' subsystems and triggering of alerts when incipient fault conditions are being detected. This is a crucial prerequisite for fast and cost-efficient responses to unforeseen maintenance needs in onshore and offshore wind turbines. This chapter provides an overview of data-driven fault detection methods based on data from the supervisory control and data acquisition system with a focus on normal behavior modeling. It also demonstrates their application for the early detection of developing faults in wind turbine drivetrains and presents a new effective approach to SCADA-based fault detection for multiple simultaneously monitored target variables with a case study for an onshore wind farm.

Keywords

Conditionmonitoring; Diagnostics; Normal behaviour models; SCADA-based fault detection; Wind turbines

1. SCADA-based condition monitoring

Megawatt-scale wind power plants are typically equipped with supervisory control and data acquisition (SCADA) systems. These are operation control and data acquisition systems with a central processing unit that can send commands to actuators, a data storage facility, sensors, and a communication network. In addition to controlling the wind turbine's operation, SCADA systems often collect and give access to data from hundreds of sensor channels from different wind turbine (WT) subsystems. The sensor instrumentation, the monitored WT components and physical variables, data models and aggregation can differ across manufacturers. Typically, the SCADA data contain temperatures measured at components along the drive train. Moreover, the SCADA system also gives access to variables describing the environmental conditions under which the WT operates, in particular the wind speed, wind direction, and the outdoor air temperature. It also provides information about the current active power generation and operating state variables such as the rotor speed and the generator speed and current. An overview of SCADA variables that are being recorded in most utility-scale wind turbines is provided in Table 1.1. Modern wind turbines are equipped with multiple complementary sensor systems monitoring their operating and health states in order to deliver data about the WT's operation and control states and the environmental conditions [1,2]. In addition to the SCADA sensor systems, the WT drive train and tower are often equipped with accelerometers to monitor their vibrational characteristics. Further sensor systems may also be present, such as oil quality sensors monitoring the gearbox oil composition and temperature.

SCADA data have been proposed and applied for fault detection tasks in wind turbine components for more than a decade (e.g. Refs. [3–6]) with the objective to improve the reliability of WT subsystems, to reduce their failure rates and the resulting downtimes [7–10]. SCADA systems are usually logging WT condition data as mean values at regular intervals, such as every 10 min. In particular, they also log the temperatures of critical components, such as bearings, and operating fluids. These temperatures are highly relevant variables in SCADA-based condition monitoring because failure processes often entail the generation of excessive heat. For instance, electrical discharges from generator windings or increased friction conditions in the gearbox due to oil loss events typically cause the associated components to heat up beyond a temperature level that would be normal at that operating state. Consequently, excessive heat detected by a wind turbine's temperature sensors can be used as an indicator of developing component failures. For this reason, SCADA-based fault detection approaches in the wind turbine drive train often involve the monitoring of component temperatures.

Table 1.1

2. Early fault detection with normal behavior models and SCADA data

The potential of vibration-based fault detection methods for WT drive train components has been demonstrated in various studies [11–16]. Vibration monitoring requires the installation, operation, and maintenance of acceleration sensing and acquisition systems near the monitored components. A large number of wind turbines currently in operation are not yet equipped with such systems despite the wealth of asset health information the sensors can give access to. This often is due to the capital expenditure needed to acquire the systems. SCADA-based fault detection methods have been proposed as an inexpensive alternative to the vibration monitoring of WT drive trains because they do not require dedicated sensors (e.g. Refs. [12,17–20]).

A range of approaches to SCADA-based fault detection in WT drive trains has been investigated. They include normal behavior modeling, trend monitoring, cluster analysis, the analysis of SCADA system alarm logs, and damage modeling [17]. This chapter focuses on normal behavior modeling because it can be considered one of the most relevant SCADA-based fault detection techniques in practice, as demonstrated by, e.g., Ref. [21]. Normal behavior models characterize the machine state during normal operation in the absence of faults. They have been in use for monitoring the health status of turbine components in the gearbox [22,23] and the generator [24]. They have also been successfully employed for monitoring active power generation [25–27]. Normal behavior models are empirical relationships between input and target variables whereby the target variables correspond to the variables to be monitored. For instance, a model of the generated power (target variable) can be established based on the wind speed and rotor speed as possible input variables. Normal behavior models are estimated from historical SCADA data using statistical or machine learning regression algorithms. The models do not incorporate any explicit knowledge about physical processes in the monitored wind turbine, such as aging. Rather, they map the input variables to the target variables of interest based on historical SCADA data in order to enable an estimation of the target variables' expected values and a comparison to their actual values. If the expected values and the actual values deviate significantly, the current operating state is an outlier in a statistical sense. This situation can point to a fault condition and should trigger more in-depth analyses.

A major advantage of normal behavior modeling with SCADA data is the empirical character of the models which naturally accounts for turbine- and site-specific phenomena, such as wake effects. Normal behavior models constitute highly customized descriptions of a turbine's operation behavior in the absence of fault processes. Those descriptions are uniquely adapted to the monitored turbine, its site and surroundings, its current configuration, and control software. This customization enables accurate and near real-time detection of deviations from normal operation conditions, based on which further diagnostics and eventually maintenance actions can be performed.

A further advantage of normal behavior models for remote fault detection tasks lies in the fact that those models can in principle be established and applied for any turbine assembly and component and for any fault type whose development is reflected in the turbine's sensor signals. In contrast, the multitude of turbine types and compositions and the variety of fault types that can arise in a given component bring about that supervised WT fault detection models can usually not be trained from historical examples: training fault detection models with supervised learning algorithms requires large enough training data sets which contain a sufficiently large number of fault observations for the same component and fault type. Such training datasets are typically not available. Therefore, normal behavior modeling remains the most relevant approach to SCADA-based fault detection in practice. However, it also faces challenges. The most important one is probably that it is an outlier detection approach, which means the model building does not involve any (or involves only little) information about fault conditions. This means that, while normal behavior models excel at identifying outliers that can indicate fault conditions, they cannot inherently discriminate relevant outliers from less relevant ones, unlike supervised learning approaches. This fact can complicate the interpretation of detected outlier events and demand for further diagnostics.

For more than a decade, normal behavior models with single target variables have been constituting the state-of-the-art SCADA-based fault detection in research and in practice. Wind farm monitoring centers often run a large number of normal behavior models with single target variables simultaneously. An early example of such monitoring software has been demonstrated by Ref. [28]. In practice, the normal behavior modeling with individual target variables can quickly result in thousands of single-target normal behavior models per monitored WT. All of those models need to be incorporated into the software and processes of the monitoring centers. Moreover, the models need to be stored, maintained, and updated when needed. The normal operation behavior can change from time to time, for instance, after a software update or hardware replacements have been performed. When this happens, the models need to be updated to the new conditions. Associated threshold values used in fault diagnostics and alerting need to be stored, maintained, and also updated. Thus, the single-target approach to fault detection requires a significant amount of resources when the number of single-target normal behavior models becomes large.

To reduce the resource demands of state-of-the-art SCADA-based fault detection methods, we recently proposed and demonstrated the application of normal behavior models with multiple target variables [29,30]. Multi-target models [31–37] are statistical or machine learning models that estimate multiple target variables simultaneously. It has been demonstrated in other research areas that multi-target models hold the potential to increase prediction accuracy compared to single-target models and that they are less susceptible to overfitting [32–34]. We have investigated multi-target normal behavior models for fault detection in WT in the case study below. In the case study, we demonstrate that they can achieve at least the same fault detection delay and in some cases even shorter delays than single-target normal behavior models. Moreover, the multi-target models can also accomplish the same degree of prediction stability. Our following case study (cf. Ref. [30]) demonstrates these advantages of multi-target normal behavior models for the SCADA-based fault detection in WT drive trains.

3. Case study: early fault detection in gear bearings with multi-target normal behavior models and SCADA data

As the number of SCADA channels has been growing strongly, thousands of independent single-target models are in place today for monitoring a single turbine. Multitarget learning was recently proposed to limit the number of models [29]. This case study applied multi-target neural networks to the task of early fault detection in drive-train components. The accuracy and delay of detecting gear-bearing faults were compared to state-of-the-art single-target approaches. We found that multitarget multilayer perceptrons (MLPs) detected faults at least as early and in many cases earlier than single-target MLPs. The multi-target MLPs could detect faults up to several days earlier than the single-target models. This can deliver a significant advantage in the planning and performance of maintenance work. At the same time, the multi-target MLPs achieved the same level of prediction stability.

3.1. Introduction

The global wind power capacity is growing strongly with a total installed volume of 743 GW in 2020 and an increase of 93 GW from 2019 [38]. The newly installed wind turbines are getting larger and increasingly more complex. At the same time, the operating cost of wind farms still makes up a major fraction, approximately 30%, of their lifetime cost [39]. Major faults can result in days and even weeks of downtime [9,40]. Therefore, they can substantially reduce the owner's return on investment and pose a considerable economic risk. As a result, many operators want to closely monitor the health state of their turbines in order to be alerted as early as possible of any developing technical problems and to prevent any major damage and downtime. To this end, an automated condition monitoring of wind turbine subsystems provides an essential prerequisite for informed operational decision-making and fast responses in case of unforeseen maintenance needs [1,2]. Data-driven automated monitoring methods have been proposed, amongst others, based on sensor data logged in the turbines' supervisory control and data acquisition (SCADA) systems [17,18,41,42]. Temperature can be an important indicator of different types of developing machine problems such as mechanical faults which can give rise to excessive friction generating heat. Therefore, a major focus of the proposed SCADA-based condition monitoring approaches is the temperature-based detection of developing faults in the wind turbine subsystems based on models of the turbine's normal operation behavior in the absence of operational faults [3,5,27,43–51]. The present study focuses on the gear-bearing temperature as an indicator of developing gearbox faults. The goal of this study is to assess the potential of multi-target regression models for automated SCADA-based fault detection. Specifically, this work has investigated and compared the delays in detecting gear bearing faults using single-target versus multi-target models of the turbines' normal operation. Moreover, the stability of the alarm signal after the first detection of a developing fault is assessed.

3.2. Data and methods

Condition monitoring data from the SCADA system of three commercial onshore turbines was analyzed in this work. The turbines are variable-speed three-bladed horizontal axis systems with pitch regulation from an onshore wind farm. Their rated power was 3.3 MW, and they operated with a three-stage planetary gearbox. The turbines' rotors were 112 m in diameter with the hub located at 84 m height above ground. The turbines' cut-in, rated, and cut-out wind speeds were specified at 3, 13, and 25 m/s, respectively.

In this study, 14 months of 10-min mean SCADA signals served to train and test the models specified below. The data were anonymized to maintain the privacy of the wind farm operator. We report the results for one of the wind turbines. It was randomly selected and the results were not affected by the choice of the turbine. We focus on monitoring the gear bearing condition based on the temperature of the gear bearing. The temperature is an important SCADA-based indicator of incipient fault processes in gearbox components [17,22,43,49,51]. In the present study, the condition of the gear bearing has been monitored based on two normal operation models of the bearing temperature. Wind speed v wind , wind direction α wind , and air temperature T air constitute the models' input variables which were provided as 10-min averages of measurements from nacelle-mounted anemometers and thermometers. The input variables were selected due to their relevance in explaining and predicting the behavior of the target variables.

A multi-target fully connected feedforward neural network was designed to predict the gear bearing temperature T gear along with the hydraulic oil temperature T oil and the transformer winding temperature T tr from the input variables at high accuracy, T gear , T oil , T tr ∼ v wind + α wind + T air . The single-target model estimates the gear bearing temperature only, T gear ∼ v wind + α wind + T air . In addition, the two fully connected feedforward neural networks (multi-layer perceptrons, MLP) were trained and tested to assess the normal operating behavior of the gear bearing based on the provided SCADA data. The model architectures were developed to obtain high predictive accuracy on the training set without overfitting the training data. In this process, the number of neurons and weights to be trained was increased only if this resulted in higher predictive accuracy. The resulting model architectures are detailed in Table 1.2.

Table 1.2

In this case study, our goal is to systematically assess the ability of multitarget neural networks to detect developing faults that result in rising component temperatures. We demonstrate this approach by the example of the gear-bearing temperature. However, it is equally applicable to faults in other subsystems and other components that result in elevated SCADA-logged temperatures.

A major challenge in data-driven fault detection and isolation is the scarcity of actually observed fault instances. We addressed this point by combining gear-bearing temperature measurements with a multitude of synthetic temperature trends in order to mimic the bearing temperature rise induced by a developing fault. To this end, a synthetic temperature trend was overlaid on the normalized gear-bearing temperature signal. One of 10 different linear temperature trends was added to the normalized bearing temperature. Temperature trends with integer slopes in the range of 1–10 were used to simulate slowly and fast-evolving fault processes. The temperature trend onset time was randomly sampled from a 2-week time window in months 12 and 13 of the 14-month observation period. Fifty different onset times have been randomly sampled from the 2-week window for each of the 10 temperature slopes. This ensured that the results did not depend on the choice of the trend onset time.

3.3. Results

Two common alarm criteria [18] were applied using the residuals of the gear-bearing temperature. The residuals were computed as the