Structural Health Monitoring: from Structures to Systems-of-Systems★

9th IFAC Symposium on Fault Detection, Supervision and 9th on Safety of Symposium Technical Processes 9th IFAC IFAC Symposium on Fault Fault Detection, Detection, Supervision Supervision and and 9th IFAC Symposium on Fault Detection, Supervision and Safety of Technical Processes Available online at www.sciencedirect.com September 2-4, 2015. Arts et Métiers ParisTech, Paris, France Safety of Technical Processes Safety of Technical Processes September 2-4, 2015. Arts et Métiers ParisTech, Paris, France September September 2-4, 2-4, 2015. 2015. Arts Arts et et Métiers Métiers ParisTech, ParisTech, Paris, Paris, France France

ScienceDirect

IFAC-PapersOnLine 48-21 (2015) 001–017

Structural Health Monitoring: from Structural Health Monitoring:  Structural Health Monitoring: from from Structures to Systems-of-Systems Structures to Systems-of-Systems Structures to Systems-of-Systems  Keith Worden, Elizabeth J. Cross, Nikolaos Dervilis, Keith Worden, Elizabeth J.Ifigeneia Cross, Nikolaos Dervilis, Keith Worden, Elizabeth Cross, Dervilis, Antoniadou KeithEvangelos Worden, Papatheou, Elizabeth J. J.Ifigeneia Cross, Nikolaos Nikolaos Dervilis, Evangelos Papatheou, Antoniadou Evangelos Papatheou, Ifigeneia Antoniadou Evangelos Papatheou, Ifigeneia Antoniadou Dynamics Research Group, Department of Mechanical Engineering, Dynamics Research Group, Department Mechanical Engineering, Dynamics Research Group, Department of Mechanical Engineering, University of Sheffield, Street,of S1 3JD, UK. Dynamics Research Group,Mappin Department ofSheffield Mechanical Engineering, University of Sheffield, Mappin Street, Sheffield S1 3JD, UK. University of Sheffield, Mappin Street, Sheffield S1 3JD, (e-mail: [email protected]). University of (e-mail: Sheffield,[email protected]). Mappin Street, Sheffield S1 3JD, UK. UK. (e-mail: [email protected]). (e-mail: [email protected]). Abstract: Almost all engineering disciplines face the problem of damage detection and Abstract: Almost alltheengineering disciplines face the problem of damage detection and Abstract: Almost disciplines face the of detection and assessment. Although terminology might change between disciplines, the problems faced, Abstract: Almost all alltheengineering engineering disciplines face between the problem problem of damage damage detectionfaced, and assessment. Although terminology might change disciplines, the problems assessment. Although the terminology might change between disciplines, the problems faced, and the methods of solution, show great commonality. The objective of this paper is to discuss assessment. Although the terminology might change between disciplines, the problems faced, and methods of solution, show great commonality. The of this paper is to discuss and the methods of show commonality. The objective of paper discuss how the Structural Health Monitoring (SHM) is carried out inobjective the context of aerospace and civil and the methodsHealth of solution, solution, show great great commonality. The objective of this this paper is is to toand discuss how Structural Monitoring (SHM) is carried out in the context of aerospace how Structural Healthand Monitoring (SHM) is is carried out in in the the contextisof ofconducted aerospacewithin and civil civil structural monitoring to draw parallels with how damage detection the how Structural Health Monitoring (SHM) carried out context aerospace and civil structural to draw parallels with how damage detection is conducted within the structural monitoring and to draw parallels with how damage detection is conducted within the electrical, monitoring control andand process engineering communities. A four-stage methodology for SHM, structural monitoring and to draw parallels with how damage detection is conducted within the electrical, control and process engineering communities. A of four-stage methodology for SHM, electrical, control process engineering A four-stage methodology SHM, based on machine learning is discussed. Thecommunities. different levels diagnostic informationfor available electrical, control and and process engineering communities. A of four-stage methodology for SHM, based on machine learning is discussed. The different levels diagnostic information available based discussed. different levels of available dependon onmachine the typelearning of data is available for The learning and this is also discussed.information Detection alone can based onon machine learning is discussed. The different levels of diagnostic diagnostic information available depend the type data available learning this is also discussed. can depend on the type of data available for learning and this is also discussed. Detection alone can be carried out usingof normal conditionfor data only and achieved using Detection supervisedalone learning depend on the type of data available for learning and this is also discussed. Detection alone can be carried out using normal condition only and this is achieved using supervised learning learning be carried out normal condition data only this is using where detection thresholds are critical.data If damage state data are available, supervised be carried out using using normalare condition data only and and thisdata is achieved achieved using supervised supervised learning learning where detection thresholds critical. If damage state are available, where detection thresholds are critical. If damage state data are available, supervised learning can be applied and more diagnostic information becomes available. The various concepts are where detection thresholds are critical. If damage state data are available, supervised learning can be applied and more diagnostic information becomes available. The various concepts are can be applied and more diagnostic information becomes available. The various concepts are discussed in terms of a number of case studies drawn from wind turbine monitoring programmes; can be applied andof more diagnostic information becomes available. The various programmes; concepts are discussed terms a number of case studies drawn from wind turbine monitoring discussed in terms of of studies drawn from wind monitoring programmes; this allowsin the illustration of both SHM methods and machine condition monitoring methods. discussed inthe terms of a a number number of case case studies drawnand from wind turbine turbine monitoring programmes; this allows illustration of both SHM methods machine condition monitoring this the illustration of both methods and machine condition monitoring methods. The allows possibility of moving towards aSHM population-based approach to SHM applicable to methods. systemsthis allows the illustration of both SHM methods and machine condition monitoring methods. The possibility of moving towards aa population-based approach to SHM applicable to systemsThe possibility of moving towards population-based approach to SHM applicable to systemsof-systems is outlined via one of the case studies. The possibility of moving towards a population-based approach to SHM applicable to systemsof-systems is outlined via one one of the the case studies. studies. of-systems is outlined via of case of-systems is outlined via one of the case studies. © 2015, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved. Keywords: Data-based damage identification; Structural Health Monitoring (SHM); Condition Keywords: damage Health (SHM); Condition Keywords: Data-based Data-based damage identification; identification; Structural Health Monitoring Monitoring (SHM);(SPC); Condition Monitoring (CM); Non-Destructive EvaluationStructural (NDE); Statistical Process Control Keywords: Data-based damage identification; Structural Health Monitoring (SHM); Condition Monitoring (CM); Non-Destructive Evaluation (NDE); Statistical Process Control (SPC); Monitoring (CM); Non-Destructive Evaluation (NDE); Statistical Process Control (SPC); Fault Detection and Isolation (FDI); monitoring of offshore wind turbines. Monitoring (CM); Non-Destructive Evaluation (NDE); Statistical turbines. Process Control (SPC); Fault Fault Detection Detection and and Isolation Isolation (FDI); (FDI); monitoring monitoring of of offshore offshore wind wind turbines. turbines. Fault Detection and Isolation (FDI); monitoring of offshore wind 1. INTRODUCTION if and how the structure has changed. In contrast, the 1. INTRODUCTION if and how viewpoint the structure changed. In contrast, the 1. INTRODUCTION if and the has changed. In the data-based useshas techniques the pattern 1. INTRODUCTION if and how how viewpoint the structure structure has changed.from In contrast, contrast, the data-based uses techniques from the pattern data-based viewpoint uses techniques from the pattern and machine learning communities in order Almost all engineering disciplines are concerned with some recognition data-based viewpoint uses techniques from the pattern recognition and about machine learning communities in order order Almost all engineering disciplines are concerned with some recognition and machine in make inferences the learning structuralcommunities state from measured Almost all disciplines are with form of damage detection or identification. The motivation and about machine learning communities in order Almost all engineering engineering disciplines are concerned concerned with some some recognition make inferences the structural state from measured form of damage detection or identification. The motivation make inferences about the structural state from measured and processed data from the structure without reference form of damage detection or identification. The motivation is clear, the implementation of an effective monitoring inferencesdata about thethe structural state from measured form of damage detection or identification. The monitoring motivation make and processed from structure without reference is clear, the implementation of an effective and processed data from the structure without reference any physics-based model. Both viewpoints have their is clear, the implementation of an effective monitoring strategy can allow both cost and safety benefits (Farrar to and processed data from the structure without reference is clear, the implementation of an effective monitoring to any physics-based model. Both viewpoints have their strategy can allow both cost and safety benefits (Farrar to any physics-based model. Both viewpoints have adherents, and the pros and cons of the two approaches strategy can allow both cost and safety benefits (Farrar and Worden (2007)). Within the aerospace and civil ento any physics-based model. Both viewpoints have their their strategy can (2007)). allow both cost the and aerospace safety benefits (Farrar adherents, and the pros and cons of the two approaches and Worden Within and civil enadherents, and the pros and cons of the two approaches is still the subject lively debate. and Worden (2007)). Within the aerospace and civil engineering communities, the relevant discipline concerning adherents, and the of pros and(occasionally cons of the heated) two approaches and Worden (2007)). Within the aerospace and civil enis still the subject of lively (occasionally heated) debate. gineering communities, the relevant discipline concerning is still subject lively (occasionally heated) debate. wishing to of dismiss model-based approach, the gineering communities, the discipline concerning damage awareness is usually called Structural Health Mon- Without is still the the subject of lively the (occasionally heated) debate. gineering communities, the relevant relevant disciplineHealth concerning Without wishing to dismiss the model-based approach, the damage awareness is usually called Structural MonWithout wishing to dismiss the model-based approach, the current paper will concentrate on the data-based alternadamage awareness is usually called Structural Health Monitoring, or SHM. Although the discipline is comparatively wishing toconcentrate dismiss the on model-based approach, the damage or awareness is usuallythe called Structural Health Mon- Without current paper will the data-based alternaitoring, SHM. Although discipline is comparatively current paper will concentrate on the data-based alternative as it is the authors’ preferred approach. itoring, or SHM. Although the discipline is comparatively young compared to other engineering fields, it has now current paper will concentrate on the data-based alternaitoring, or SHM. Although the discipline is comparatively tive as as it it is the the authors’ authors’ preferred preferred approach. approach. young compared to other fields, ithave has now young compared to engineering fields, has now matured to the point where engineering clear methodologies been tive as it is is theprinciple authors’for preferred approach. young compared to other other engineering fields, it ithave has been now tive An organising a data-based approach to SHM matured to the point where clear methodologies matured to the point where clear methodologies have been framed and monographs have been written (e.g. Farrar An organising principle for a data-based approach SHM matured to the point where clear methodologies have been An organising organising principle for aaofdata-based data-based approach to SHM can be formulated in terms a four-stage process to (Farrar framed and monographs have been written (e.g. Farrar An principle for approach to SHM framed and monographs have been written (e.g. Farrar and Worden (2012)). It can be argued that two dominant can be formulated in terms of a four-stage process (Farrar framed and monographs have been written (e.g. Farrar can be formulated in terms of a four-stage process (Farrar and be Worden (2007)): and Worden (2012)). It can be argued that two dominant can formulated in terms of a four-stage process (Farrar and Worden (2012)). argued two ‘competing’ philosophies forbe SHM havethat emerged over its and Worden and Worden philosophies (2012)). It It can can be argued that two dominant dominant and Worden (2007)): (2007)): ‘competing’ for SHM have emerged over its and (2007)): ‘competing’ philosophies for SHM have emerged over its (i) Worden Operational evaluation. period of development; each regarding the problem from ‘competing’ philosophies for SHM have emerged over its (i) Operational evaluation. period of development; each regarding the problem from (i) Operational evaluation. period of development; each regarding the problem from (ii) Data acquisition, normalisation and cleansing. a different viewpoint. The division in the subject isfrom be(i) Operational evaluation. period of development; each regarding the problem (ii) Data acquisition, normalisation and cleansing. aatween different viewpoint. The division in the subject is be(ii) Data acquisition, normalisation and cleansing. different viewpoint. The division in the subject is be(iii) Feature selection and information condensation. model-based SHM and data-based SHM. The first (ii) Data acquisition, normalisation and cleansing. a different viewpoint. The division in the subject is be(iii) Feature selection and information condensation. tween model-based SHM and data-based SHM. The first (iii) Feature selection and information condensation. tween model-based SHM and data-based SHM. The first (iv) Statistical model development for feature discriminaviewpoint is primarily based on establishing a physical (iii) Feature selection and information condensation. tween model-based SHM and data-based SHM. The first (iv) tion. Statistical model model development development for for feature feature discriminadiscriminaviewpoint is primarily on establishing aa physical Statistical viewpoint is primarily based on establishing physical law-based model of thebased structure of interest in its nor- (iv) (iv) Statistical model development for feature discriminaviewpoint is primarily based on establishing a physical tion. law-based model of the structure of interest in its nortion. law-based model of the structure of interest in its normal (undamaged) condition; whenofsubsequent data are tion. law-based model of the structure interest in its norThis paper is concerned with elaborating on the stages mal (undamaged) when data are mal (undamaged) condition; when subsequent data are accumulated, they condition; are checked for subsequent conformity with the paper is concerned elaborating on the stages mal (undamaged) condition; when subsequent data the are This This paper is with elaborating on stages (ii)-(iv) in some detail andwith arguing that general principles accumulated, they are checked for conformity with This paper is concerned concerned with elaborating on the the stages accumulated, they are checked for conformity the model, and any discrepancies can be used to with diagnose (ii)-(iv) in some detail and arguing that general principles accumulated, they are checked for conformity with the (ii)-(iv) in some detail and arguing that general principles can be drawn out, allowing principled strategies for datamodel, and any discrepancies can be used to diagnose (ii)-(iv) in some detail and arguing that general principles model, and be drawn out, allowing principled strategies for datamodel, and any any discrepancies discrepancies can can be be used used to to diagnose diagnose can  can be drawn out, allowing principled strategies for based damage identification across all engineering conThe support of the UK Engineering and Physical Sciences can be damage drawn out, allowing principled strategies for datadata based identification across all engineering conThe support of the UK Engineering and Physical Sciences  based damage identification across all engineering contexts. Stage (i) will be rather neglected in the current Research Council (EPSRC) through grant reference numbers The support of the UK Engineering and Physical Sciences  based damage(i)identification across all engineering conThe support of (EPSRC) the UK Engineering and reference Physical numbers Sciences texts. Stage will be rather neglected in the current Research Council through grant texts. Stage (i) will be rather neglected in the current EP/J016942/1 and EP/K003836/2, and that of the EU Framework Research Council (EPSRC) through grant reference numbers paper as it is(i)more concerned with answering questions texts. Stage will be rather neglected in the current Research Council (EPSRC) through grant reference numbers EP/J016942/1 and EP/K003836/2, and that of the EU Framework paper as it is more concerned with answering questions 7 Programme for theEP/K003836/2, ITN project SYSWIND, is the gratefully acknowlEP/J016942/1 and and EU paper as is concerned with answering questions about the process that occur the data EP/J016942/1 and and that that of of EU Framework Framework paper as it itSHM is more more concerned withupstream answeringof questions 7 Programme Programme for for theEP/K003836/2, ITN project project SYSWIND, SYSWIND, is the gratefully acknowlabout the SHM process that occur upstream of the data edged. 7 the ITN is gratefully acknowlabout the SHM process that occur upstream of 7 Programme for the ITN project SYSWIND, is gratefully acknowlabout the SHM process that occur upstream of the the data data edged. edged.

edged. Copyright 2015 IFAC 1 Hosting by Elsevier Ltd. All rights reserved. 2405-8963 © 2015, IFAC (International Federation of Automatic Control) Copyright 2015 IFAC 1 Copyright © 2015 IFAC 1 Peer review© of International Federation of Automatic Copyright ©under 2015 responsibility IFAC 1 Control. 10.1016/j.ifacol.2015.09.497

SAFEPROCESS 2015 2 September 2-4, 2015. Paris, France

Keith Worden et al. / IFAC-PapersOnLine 48-21 (2015) 001–017

acquisition. This is not to say that operational evaluation is any less important than the other stages, in fact, it is arguably the most important; however, the focus of this paper will be much more on the data processing subsequent to acquisition. Before leaving operational evaluation for now, it is worth pointing out that a useful framework has been formulated as a sequence of questions to be asked (Farrar and Worden (2007)):

• Fault Detection and Isolation (FDI).

SHM is mostly relevant to structures such as aircraft and buildings and usually implies a permanently-installed sensor network that monitors the behaviour of the structure. Sometimes it is assumed that SHM means that monitoring is carried out on-line; however, in general it can be carried out when the structure is either in or out of operation. Typical sensors are optical fibres, electrical resistance strain gauges or acoustic devices. CM is relevant to rotating and reciprocating machinery, such as used in manufacturing or power generation. CM also uses online techniques that are often vibration based and uses accelerometers as sensors. NDE is usually carried out offline after any damage has been located or inferred using on-line sensors. (There are exceptions to this rule, NDE is used as a monitoring tool for e.g. pressure vessels and rails where the a priori damage location is not fixed.) NDE is therefore primarily used for characterisation and as a severity check when there is a prior knowledge of the location of the damage. Typical techniques include ultrasound, thermography and shearography. SPC is rather different, being process-based rather than structure-based and uses a variety of sensors to monitor changes in the process of interest . SPC is predominantly associated with the Chemical Engineering industry; however, its basic techniques for change detection in time series underpin many damage detection schemes (Montgomery (2009)). Finally, FDI is usually associated with the electrical or control engineering disciplines and can accommodate electromechanical structures; in this context, control is important and loss of control may be regarded as ‘damage’ to the system of interest as much as actual physical degradation. This latter point is something which FDI shares with SPC. A more detailed account of the five contexts is given in a little while.

(i) What are the life-safety and/or economic justification for performing SHM? (ii) How is damage defined for the system being investigated and, for multiple damage possibilities, which cases are of the most concern? (iii) What are the conditions, both operational and environmental, under which the system to be monitored functions? (iv) What are the limitations on acquiring data in the operational environment? One can immediately see that these are important issues, concerned with the cost-benefit analysis of performing any monitoring in the first place. However, as stated, this paper will now proceed to discussing aspects of SHM, and other damage identification frameworks concerned with signal or data processing. In particular, it will be argued that techniques from pattern recognition and machine learning hold the key to success for data-based approaches. A great deal of the material in the following two sections can be found in a slightly different form in Worden and DulieuBarton (2004). The authors have included this material here (but have taken the opportunity to update it where possible) in order that the paper be as self-contained as possible, as they consider this a desirable property for a keynote presentation. The layout of the paper is as follows: the next section will discuss how SHM can be formulated in terms of a hierarchical structure and then argue that simple modifications of that structure provide parallel frameworks for damage identification in other contexts. Section Three will then give an overview of how machine learning provides the tools to address the damage identification problem at its various levels. The three sections that follow will each provide a case study; the first is concerned with condition monitoring of a wind turbine gearbox, the second with damage detection in a large-scale fatigue test of a wind turbine blade and the third with how SHM might be accomplished for a population of structures (in that case, an offshore wind farm). The paper is then finalised with brief conclusions.

There is the potential to add HUMS (Health and Usage Monitoring Systems) to the above list as a technology that has proved very useful in the aerospace industry (Chronkite (1993)). In fact, HUMS technologies are arguably the only damage awareness approaches to have made the jump from a research interest to applied technology in industry. The most notable fact is that HUMS have been accepted into the rotorcraft industry as a matter of legislation (in some countries at least). It is argued here that HUMS is a distinct discipline from those above as it is largely concerned with a very specific aspect of damage awareness; it is purely concerned with transforming online measurements of loads and/or stresses into a running tally of fatigue accumulation for the structure of interest. At first glance it may seem that the five areas of SHM, CM, NDE, SPC and FDI are very different because of their application range. In fact there are many similarities that should be explored together if a unified approach to damage identification is to be created. All of the techniques used in SHM, CM, NDE, SPC and FDI initially use sensors that provide signals that are dependent on the behaviour of the structure, system or process. If damage evaluation is to evolve beyond a simple threshold approach (i.e. the signal has exceeded a certain limit and therefore the structure must be taken off-line, the damage located and repaired) then techniques for processing the signal to derive features that enable quantification must

2. DAMAGE IDENTIFICATION HIERARCHIES Damage evaluation is relevant to almost all the engineering disciplines; however, the issues and methods tend to be discussed in context-dependent ways, sometimes with slightly different terminologies. There are arguably five key multidisciplinary areas for which monitoring and assessing damage are principal concerns: • • • •

Structural Health Monitoring (SHM), Condition Monitoring (CM), Non-Destructive Evaluation (NDE), Statistical Process Control (SPC), 2

SAFEPROCESS 2015 September 2-4, 2015. Paris, France

Keith Worden et al. / IFAC-PapersOnLine 48-21 (2015) 001–017

3

CLASSIFICATION: the method gives information about the type of damage. ASSESSMENT: the method gives an estimate of the extent of the damage. PREDICTION: the method offers information about the safety of the structure e.g. estimates a residual life.

be established. In doing this, the degree to which a signal processing regime is intelligent must be defined. Rytter (1993) has done this, to a certain extent and his hierarchical approach to damage identification will be discussed next. A full treatment of intelligent damage identification would address the many issues which can arise from the sensor systems and networks commonly employed for monitoring; however, this will not be covered here, the curious reader can consult Farrar and Worden (2007) as an introduction.

Classification is important, if not vital, for effective identification at Level Five and possibly at Level Four. Level Five is distinguished from the others in that it cannot be accomplished without an understanding of the physics of the damage, i.e. characterisation. Level One is also distinguished in the following sense - it can be accomplished with no prior knowledge of how the system will behave when damaged. This will be discussed in the following section.

Detection of damage is a facet of the broader problem of damage awareness or damage identification. The objective of a monitoring system must be to accumulate sufficient information about the damage for appropriate remedial action to be taken to restore the structure or system to high-quality operation or at least to ensure safety. Also, efficiency demands that only the necessary information should be returned by the monitor. With this in mind, it is helpful to think of the identification problem as a hierarchical structure. As mentioned above, this train of thought began with Rytter (1993); the original specification was defined in the context of SHM only, and cited four levels:

One of the aims of this lecture is to discuss a unified framework for intelligent damage detection, applicable over a broad range of application areas. At least it should be applicable within the contexts discussed above, namely: SHM, CM, NDE, SPC and FDI. The hierarchical framework presented above is general enough to encompass these disciplines, and can also accommodate context-specific issues as identified at the operational evaluation stage. To recap, one of the requirements of pre-planning is to:

DETECTION: the method gives a qualitative indication that damage might be present in the structure. LOCALISATION: the method gives information about the probable position of the damage. ASSESSMENT: the method gives an estimate of the extent of the damage. PREDICTION: the method offers information about the safety of the structure e.g. estimates a residual life.

identify context-specific features and decide the appropriate level for monitoring. Essentially, what is required here are context-specific versions of Rytters hierarchy which was initially defined only with SHM in mind. The following subsections discuss the damage identification levels for the other four contexts.

The vertical structure is clear, each level requires that all lower-level information is available. In general, the hierarchy can be adapted to different damage detection contexts from its SHM origins. This will be discussed later. Interestingly, the context of HUMS mentioned above also fits Rytter’s model, but can allow progression directly, via the assumption of damage accumulation, to PREDICTION, as it attempts to monitor fatigue-life exhaustion.

2.1 Condition Monitoring In machine condition monitoring, the fundamental requirement is to detect the damage, and in many situations this will be all that is required. If higher-level information is required, it is with a slightly different emphasis to SHM. Often, in monitoring a machine, the location of the problem is not an issue. Machines are largely heterogeneous structures and each of the critical components will usually be subject to different failure mechanisms. The damage types are different for each component e.g. bearings, gearboxes, rotors. This fact means that effective damage classification usually also amounts to effective location; often each important component will be individually given a sensor. Another aspect of machine monitoring, is that the many of the components will be low-cost items. This means that it is economic to replace them at the first sign of damage. This in turns means less emphasis on estimating the severity of the damage or determining the remaining safe-life, although residual-life studies exist. The relevant parts of Rytters structure for CM would be:

Note that the damage identification scheme should if at all possible, be implemented on-line i.e. during operation of the structure; in this case prediction must also be understood as an estimate of the residual safe-life of the structure obtained during operation. For an aircraft in flight, for example, this is critical. If the diagnostic system signals serious damage but fails to indicate that there is time to land, the aircraft may be lost needlessly and at great expense when the crew bail out (the primary concern is of course that the crew do bail out, issues of life safety far outweigh economic considerations). Few have argued against the structure above summarising the main issues in damage identification (at least for SHM), except that one major exception can be remedied by the introduction of a new level. At the risk of repetition, the new structure is:

DETECTION: the method gives a qualitative indication that damage might be present in the structure. CLASSIFICATION: the method gives information about the type of damage. ASSESSMENT: the method gives an estimate of the extent of the damage. PREDICTION: the method offers information about the safety of the structure e.g. estimates a residual life.

DETECTION: the method gives a qualitative indication that damage might be present in the structure. LOCALISATION: the method gives information about the probable position of the damage.

3

SAFEPROCESS 2015 4 September 2-4, 2015. Paris, France

Keith Worden et al. / IFAC-PapersOnLine 48-21 (2015) 001–017

2.2 Non-Destructive Evaluation

Having added CONTROL as one of the desirables for SPC damage identification, consistency implies that a level REPAIR be added to the level specifications for the other contexts. Rather than do this, the functions CONTROL and REPAIR will be subsumed into the overall processing strategy as described in the next section. This reduces the structure above for SPC to:

Again, the emphasis is different to SHM. Most methods of NDE involve a priori specification of the area of interest, i.e. eddy current methods, radiography, thermography etc. This means that location is not an issue. There are exceptions; ultrasonic methods of damage detection are usually classed as NDE methods, and in the case of Ascan, C-scan etc. that assume a prior location, this is appropriate. However, methods based on, for example, ultrasonic Lamb wave scattering also have the potential to locate damage, even over reasonable distances for example, there are applications to pipe inspection. For many of the NDE methods, the extent of the damage is evident, e.g. area of a delamination in C-scan. It is arguable that the prediction level is not critical for NDE as almost all inspection methods involve taking the structure or system out of service. In fact, almost all applications will be off-line. The important parts of the identification hierarchy for NDE are thus:

DETECTION: the method gives a qualitative indication that damage might be present in the system. ISOLATION: the method gives information about the source of the damage. 2.4 Fault Detection and Isolation

As a discipline, this is rather similar to SPC in the sense that CONTROL is a major aspect. FDI is one of the most-often used terms to cover damage identification in the electrical and control engineering communities (Gertler (1998); Chen and Patton (1999)). The discipline might be said to consider coupled electro-mechanical sysDETECTION: the method gives a qualitative indica- tems in much the same way that SPC can be concerned with coupled thermomechanical-chemical systems. A comtion that damage might be present in the structure. CLASSIFICATION: the method gives information about plication in both cases is provided by the fact that the systems of interest may be operating in open or closed-loop. the type of damage. ASSESSMENT: the method gives an estimate of the In any case, one might argue that the Rytter hierarchy for SPC also applies to FDI. extent of the damage. 2.3 Statistical Process Control

3. DATA PROCESSING FOR DAMAGE IDENTIFICATION

In process control, the fundamental objective is, as always, the detection of damage. In this case, the term location is somewhat inadequate as the damage may take several forms, ranging from a true engineering problem like a blocked valve in the plant to a substandard batch of reactants. A better term, and one that is often used, is isolation. Severity and residual-life prediction are not issues either. If a blocked valve is present and detected, it is likely that the system is already producing an unacceptable reduction in quality, the process is stopped and the valve is cleared. Alternatively, if the system is controllable, and the damage is the result of an out-ofcontrol process, it will be desirable to bring the system back inside proper control limits. The emphasis in process control is to maintain consistent operation in the face of variations in the dynamics of the process. This focus places SPC in contrast to the other contexts described so far, in which the system or structure must sooner or later be subject to remedial action i.e. repair or replacement, if and when damage is detected. The resulting identification hierarchy is a little different for SPC:

Once the operational evaluation stage has passed and the sensor network has been designed, the health monitoring system can begin to deliver data. The choice and implementation of algorithms to process the data and carry out the identification is arguably the most crucial ingredient of an intelligent damage detection strategy. As discussed earlier, before even choosing the algorithm, it is necessary to choose between two complementary approaches to the problem: • damage identification is an inverse problem; • damage identification is a pattern recognition problem.

DETECTION: the method gives a qualitative indication that damage might be present in the system. ISOLATION: the method gives information about the source of the damage. CONTROL: the method gives a control input to restore normal condition.

The first approach usually adopts a physics-based model of the structure and tries to relate changes in measured data from the structure to changes in the model, sometimes locally linearised models are used to simplify the analysis. The algorithms used are mainly based on linear algebra or optimisation theory. The second approach is based on the idea that measured data from the system of interest are assigned a damage class by a pattern recognition algorithm. This is the approach that is chosen here for detailed discussion. Somewhat confusingly, the data-based approach also makes extensive use of models; however, these are almost exclusively statistical models rather than the physics-based ones of the inverse-problem approach.

SPC initially considered a single measurand at a time. The development of data fusion algorithms and Multivariate SPC (MSPC) allowed multiple parameters to be processed simultaneously (Montgomery (2009)). This gives substantially more accurate isolation at the expense of more costly statistical processing.

The levels in Rytter’s hierarchy are addressed in different ways according to whether one pursues a model-based or data-based approach. For example, consider the level LOCALISATION. In a model-based approach to SHM, the equations of motion of the system of interest will often be constructed on the basis of Newton’s second law and 4

SAFEPROCESS 2015 September 2-4, 2015. Paris, France

Keith Worden et al. / IFAC-PapersOnLine 48-21 (2015) 001–017

for the classifier. The availability of damage state data is a serious issue for data-based SHM; however, as will be discussed later, the DETECTION level need only require data from the normal condition of the structure. This paper is concerned only with data-based damage identification from this point onwards, the curious reader can consult Friswell (2007) for more discussion of (physical) model-based approaches.

will usually constitute a set of second-order differential equations. The variables in the equations of motion will be the displacements at some set of reference positions on the structure; the parameters of the equations will summarise the physical properties of the structure in the vicinity of the reference points. For N reference points, or degrees-offreedom, one will have, [m]{¨ y } + [c]{y} ˙ + [k]{y} = {x}

5

(1)

The data processing element of a monitoring system comprises all actions on the data upstream from the point of acquisition by the sensors. The ultimate product of the analysis is a decision as to the health of the system. The analysis has been neatly summed up by Lowe Lowe (2000), as the D2D (Data to Decision) process, a process based on ideas of data fusion. The field of sensor or data fusion emerged mainly as a result of various defence organisations attempting to formalise procedures for integrating information from disparate sources. The first definition of data fusion came from the North American Joint Directors of Laboratories (JDL), and is,

where [m], [c] and [k] are respectively the mass, damping and stiffness matrices of the structure, {x} and {y} are the vectors of excitation forces and responses respectively and overdots denote differentiation with respect to time. (Throughout this paper, straight brackets will denote matrices and curved brackets, vectors.) A typical inverse-problem approach to SHM would be model updating Friswell (2007); in this approach, one would establish an initial model of the form (1), either by system identification or by constructing a finite element model. It would be assumed then that the model represented the normal condition of the system of interest. During the monitoring phase, data measured from the structure would be tested in order to see that it remains consistent with the normal condition model, potentially by re-estimating system parameters or adjusting them in order to bring the model into correspondence with new data. DETECTION is accomplished when a significant (according to some meaningful statistical measure) parameter change is needed in order to maintain fidelity of the model - the implication being that the system has changed. Because parameters are associated with specific DOFs of the model, the parameters that change determine the LOCALISATION of the damage, and because the parameters have physical meanings, the extent of the changes establishes the ASSESSMENT level. Another point to note in terms of the model-based approach is that much more general models may be needed for the damage ID contexts apart from SHM. It is likely that a model-based approach to SPC would require differential equations expressing the chemical processes or reactions taking place; similarly the FDI context would likely require equations for electrical components. At a quite fundamental level, nonlinear effects may need to be accommodated.

A multilevel, multifaceted process dealing with the automatic detection, association, correlation, estimation and combination of data from single and multiple sources. The objective was to determine battlefield situation and assess threat on the basis of data from various sources; the ideas generated were quickly recognised to have broader application, initially in the fields of meteorology and traffic management, but later in the medical field and in NDE. The military terms used in the first studies of fusion soon gave way to a more general terminology. One of the first general fusion models for data fusion was the Waterfall model developed in the UK Defence Evaluation Research Agency (DERA) (Bedworth (1994)). It is no longer widely used; however, it will suffice to illustrate the main stages of the D2D process (Figure 1).

Sensing

In a data-based strategy for SHM, the approach is quite different. One measures or extracts quantities from the sensor measurements - features - that show sensitivity to the damage e.g. natural frequencies in the context of vibration-based SHM. A machine learning analysis is then applied to the feature data in order to construct statistical models of the different classes one wishes to distinguish and a classifier to determine which class new data might belong to. The classes would always include one for the normal condition of the system, but should also contain classes for any damage state that one wishes to distinguish. For example, in a data-based approach, LOCALISATION might be achieved by dividing the structure into substructures and assigning a class label for data corresponding to damage in given substructures. The issue here is that one would need data from the structure of interest in many different damage states and this may be impossible, or very expensive, to obtain. Similarly, the ASSESSMENT level would require data for different damage extents, each data set represented by a class label

Signal Processing Feature Extraction Pattern Processing Situation Assessment Decision Making

Fig. 1. Waterfall model of data fusion. 5

SAFEPROCESS 2015 6 September 2-4, 2015. Paris, France

Keith Worden et al. / IFAC-PapersOnLine 48-21 (2015) 001–017

at this stage that the information discarded in the dimension reduction is not relevant for diagnosing the damage. Feature extraction should only discard components of the data that do not distinguish the different system states and thus concentrate the information about damage. Some practitioners use the term symptoms for what are defined here as features. This is a little imprecise, in common English usage, a symptom is evidence of the presence of a problem. Most of the time, the monitored quantities will show no evidence of damage because damage is not present. Only when damage occurs do the features display symptoms.

Apart from the situation assessment stage, which is an artefact of the old military terminology, all of the important processing stages are in the model. Beyond the sensor level which is considered to generate the raw data, the first stage is signal processing. This should more properly be called pre-processing. The purpose is to prepare the data for feature extraction (more of that later). The pre-processing stage can encompass two tasks. The first of these is commonly data cleansing. Examples of cleansing processes are: filtering to remove noise, spike removal by median filtering, removal of outliers (care is needed here as the presence of outliers is one indication that the data is not from normal condition), and treatment of missing data values. The second (optional) preprocessing stage is a preliminary attempt to reduce the dimensions of the data vectors and further de-noise the signal. For example, given a random time-series with many points, it is often useful to convert the data to a spectrum by Fourier transformation. If the signal is divided into contiguous blocks before transformation and the resulting spectra are averaged, the number of points in the spectrum can be much lower than in the original time-history and noise is averaged away. Another advantage of treating the time signal this way is that the data vector obtained is independent of time (assuming all relevant dynamics is stationary). Pre-processing is usually carried out on the basis of engineering judgement and experience; the aim would be to reduce the dimension of the data set from possibly many thousands to perhaps a hundred - but to preserve (or hopefully amplify) any signatures of damage.

The next stage is pattern processing. This is the application of an algorithm which can assign a class label, and thus determine the damage state, on the basis of the given feature vector. An example would be a neural network that has been trained to return the damage type and severity when presented with say, condensed spectral information from a gearbox. There is an important distinction to be made here between two main types of learning: supervised and unsupervised learning. In supervised learning, the training data consists of a set of feature vectors together with their known class labels. Supervised learning algorithms learn to associate labels with features from the training data and can then be applied to new data. In contrast, unsupervised algorithms are usually applied in a two-class problem where feature data are only known for one class. Unsupervised classifiers work by learning the characteristics of the known class and then attempting to judge if new data are consistent with the known class or not. Unsupervised learning can also be used in order to explore the properties of data sets via clustering and visualisation algorithms; in fact, well-known dimension reduction algorithms like PCA and factor analysis can be formulated in unsupervised learning terms (Roweis and Ghahramani (1999)).

The second stage is feature extraction. The term feature comes from the pattern recognition literature and is short for distinguishing feature. As the fundamental problem of pattern recognition is to assign a class label to a vector of measurements, this task is made simple if the data contains dominant features that distinguish them from data from other classes. In general, the components of the signal that distinguish the various damage classes will be masked by features that characterise the normal operating condition of the structure, particularly when the damage is not yet severe. The aim of feature extraction is to magnify the characteristics of the various damage classes and suppress the normal background. For example, suppose the raw data from the sensors is a time-series of accelerations from the outside of a gearbox casing. Further suppose that the time data has been pre-processed and converted into an averaged spectra. Feature extraction in this situation could be by keeping only the spectral lines at the meshing frequency and its harmonics as these lines are known to be sensitive to damage (strictly speaking, this is feature selection). So feature extraction can be carried out on the basis of engineering judgement also. Alternatively, statistical or information theoretic algorithms can be used to reduce the dimension like Principal Component Analysis (PCA) Worden et al. (2011). The resulting low-dimensional data set is the feature vector or pattern vector the pattern recognition algorithm will use to assign a class. The aim of this stage would be to generate a feature vector of dimension less than ten. A low-dimensional feature vector is a critical element in any pattern recognition problem as the number of data examples needed for training grows explosively with the dimension of the problem. Care must be taken

Three main types of algorithm can be distinguished depending on the desired diagnosis. Novelty detection. In this case, the algorithm is required to simply indicate if the data comes from normal operating condition or not. This is a two-class problem which has the advantage that unsupervised learning can be used and only data from the normal condition of the structure or system is needed for training. There are many methods for novelty detection, some of the main ones are: outlier analysis, kernel density methods, autoassociative neural networks, Kohonen networks, growing radial basis function networks, methods based on SPC control charts and one class support vector machines. Some of these methods will be illustrated on the case studies later in the paper, many more are discussed in the comprehensive reviews Markou and Singh (2003a,b); Pimentel et al. (2014). While the DETECTION level in damage identification is distinguished by the fact that unsupervised learning can be used, this does not come without a price. The issue is that damage detection is not the same problem as novelty (or change) detection. The features from a structure in its normal condition may show changes as a result of benign variations in its environment or in its mode of operation (sometimes called confounding influences in this context). For example, if the natural frequencies 6

SAFEPROCESS 2015 September 2-4, 2015. Paris, France

Keith Worden et al. / IFAC-PapersOnLine 48-21 (2015) 001–017

7

fused together at the pre-processing or feature extraction stages. In general fusion models also allow information to be combined at the pattern level or the decision level. The objective at all times is to reach a decision with higher confidence than can be reached using any of the information sources alone. A better fusion paradigm than the waterfall model is the Omnibus model (Bedworth and O’Brien (1999)). This is illustrated in Figure 2.

of a bridge are used as features for monitoring, they may change for harmless reasons like variations in the ambient temperature or pattern of traffic loading. This issue means that damage detection algorithms usually need to supplement the actual novelty detection stage with the prior application of a method for projecting out the effects of benign changes, or false alarms will occur regularly (Sohn (2007); Worden et al. (2007b)). Classification. In this case, the output of the algorithm is a discrete class label. In order to apply such an algorithm, the damage states must be quantised, i.e. for location, the structure should be divided into labelled substructures. In this case, the algorithm could only locate to within a sub-structure, so resolution of what is essentially a continuous parameter may not be good unless many labels are used. However, this type of algorithm is useful in the sense that the algorithms can be trained to give the probability of class membership, this gives an in-built confidence factor in the diagnosis. In the case where the desired diagnosis is from a discrete set, e.g. for diagnosing damage type, this class of algorithms are singled out. Examples of algorithms include: neural network classifiers trained with the 1 of M rule, support vector machines, linear and quadratic discriminant analysis, kernel discriminant analysis and nearest neighbour classifiers. A number of powerful Bayesian algorithms have also appeared in the fairly recent past, including Bayesian MLPs and Relevance Vector Machines (Worden et al. (2011)). Regression. In this case the output of the algorithm is one or more continuous variables. For location purposes, the diagnosis might be the Cartesian coordinates of the fault, for severity assessment it could be the length of a fatigue crack. The regression problem is often nonlinear and is particularly suited to neural networks. As in the classification case, it is often possible to recover a confidence interval for a neural network prediction. In particular, Bayesian algorithms like Bayesian neural networks and Gaussian processes provide confidence intervals for predictions natural (Worden et al. (2011));

Decision Making

Soft Decision Fusion

Hard Decision Fusion

Context Processing Decide Pattern Processing Feature Extraction

Control Orientate

Act

Resource Tasking

Observe Signal Processing Sensor Data Fusion

Sensing

Sensor Management

Fig. 2. Omnibus model of data fusion. This has several advantages over the waterfall model. First of all it includes the possibility of action. In the context of structural health monitoring, this could be REPAIR; in the context of SPC or FDI this could be CONTROL. In the event of a sensor failure, this could entail the reallocation of sensors or the switching in of redundant sensors. The model has a loop structure that makes clear the fact that action need not interrupt the monitoring process, but may enhance it. The next sections of the paper will illustrate some of the ideas discussed above, but not all. In most cases, the main concern in the following case studies is DETECTION; however, a range of signal processing and data analysis techniques will be covered.

In all cases, the pattern processing is subject to an important limitation. There is a trade-off between the resolution of the diagnosis and the noise rejection capabilities of the algorithm (this can be considered to be an axiom of SHM (Worden et al. (2007a)). Put simply, if the data is always noise-free, there will be very little fluctuation in the measurement from normal operating condition; in this case, small damages will cause detectable deviations. If there is much noise on the training data, it will be difficult to distinguish fluctuations due to noise and deviations due to damage unless the damage is severe. One of the tasks of feature extraction is to eliminate as far as possible, fluctuations on the normal condition data. This optimisation for performance is a requisite feature of intelligent damage detection.

4. CASE STUDY I: CONDITION MONITORING OF A WIND TURBINE GEARBOX The objective of this case study is to illustrate a databased approach to a condition monitoring problem; that of detecting damage in a wind turbine gearbox. The damage identification strategy is only really adddressed as far as DETECTION, and an unsupervised learning approach is adopted. One approach to novelty detection is via techniques from the statistical discipline of outlier detection; the simplest approach assumes that the selected features defining the normal condition follow a Gaussian distribution (Worden et al. (2000)). In Worden et al. (2000) the Mahalanobis Squared-Distance (MSD) measure is used as a discordancy (or novelty) measure and the threshold value for damage is estimated via a Monte Carlo method. The MSD measure has a very important limitation, which is that the outlier detection must be conducted exclusively; the training set cannot have any

The final stage in the D2D chain for the Waterfall model is the decision. This is a matter of considering the outputs of the pattern recognition algorithm and deciding whether action needs to be taken, and what that action should be. The Waterfall model is a fusion model in the sense that it allows for a sensor network with different types of sensors, all of which generate relevant information that can be 7

SAFEPROCESS 2015 8 September 2-4, 2015. Paris, France

Keith Worden et al. / IFAC-PapersOnLine 48-21 (2015) 001–017

examples of a damage state in it, or they will bias the method and reduce its diagnostic power. Here, a different outlier detection method will also be shown, which does not suffer from the masking effect of multiple outliers in the training set; it can separate within one data set, the normal condition from the outliers. The method is based on the idea of spatial adaptation, a term from the statistical literature, describing methods that can be used for regression and classification and operate without a predefined fixed basis, but rather create this basis during their application, according to the data being analysed.

where λU is termed the universal threshold. For a normal, random variable sample, that consists of n data points and whose standard deviation is estimated by σ and the mean is zero, the expected absolute maximum is,  λU σ = 2 log nσ (4)

The thresholding algorithm consists of a number of sequential iterations that stop when the number of inliers becomes constant (or equivalently, the number of new points identified as outliers, falls to zero). If ui is a sample from the dataset being analysed, then each iteration has the following steps:

The new (i.e. new to the SHM context (Antoniadou (2014))) method discussed here was developed originally in Goring and Nikora (2002) where it was used for acousticDoppler velocimetry data analysis. Spike detection in this case was achieved by exploiting the three following concepts (Goring and Nikora (2002); Mori et al. (2007)): a) that differentiation enhances the high frequency portion of the signal, b) that the expected maximum of a random series is given by a universal threshold, and c) the idea of a “good” data cluster in a dense cloud in phase space or within Poincar´e maps. The basic ideas behind the approach are now summarised.

• Calculate ∆ui and ∆2 ui for the first and second derivatives from, ∆ui = (ui+1 − ui−1 )/2 (5)

∆2 ui = (∆ui+1 − ∆ui−1 )/2 (6) • Calculate the standard deviations of all three variables σu , σ∆u and σ∆2 u , and then the expected maxima using the universal criterion. • Calculate the rotation angle of the principal axis of the ∆2 ui versus ui using the cross correlation,  u i ∆2 u i −1 (7) θ = tan (  2 ) ui • Each set of variables {ui , ∆ui , ∆2 ui }, determines a point in spherical coordinates. For each pair of these variables, an ellipse can be calculated. For ∆ui versus ui the major axis is λU σu and the minor axis is λU σ∆u ; for ∆2 ui versus ∆ui the major axis is λU σ∆u and the minor axis is λU σ∆2 u ; and for ∆2 ui versus ui the major and minor axes a and b respectively, can be shown by geometry to be the solution of, (λU σu )2 = a2 cos2 θ + b2 sin2 θ (8)

4.1 Background Theory Outlier Analysis Only the briefest summary is provided here, the reader may consult Worden et al. (2000) for more details. In general, the methods compute discordancy measures for data and compares them with a threshold. In the case of multivariate data, the simplest possible discordancy measure is the MSD, defined by, Dζ2 = ({x}ζ − {x})T [Σ]−1 ({x}ζ − {x}) (2)

where {x}ζ is the feature vector corresponding to the candidate outlier, {x} is the sample mean of the normal condition features and [Σ] is the normal condition feature sample covariance matrix. The MSD is a scalar discordancy measure which can simply be compared to a threshold value in order to establish if a point represents an outlier. The thresholds here are computed using the Monte Carlo approach described in Worden et al. (2000); they are based on extreme value statistics and are a function of the dimension of the feature vectors and the size of the training set.

(λU σ∆2 u )2 = a2 sin2 θ + b2 cos2 θ (9) • For each projection in space the points that lie outside of the ellipse are identified and replaced with a smoothed estimate in order to perform the next iteration. At each iteration, replacement of the outliers reduces the standard deviations calculated in the second step and thus the size of the ellipsoid reduces until further outlier replacement has no effect.

Phase-Space Thresholding Method This method essentially constructs an ellipsoid in a three-dimensional phase space, without using the central statistics of the data; points lying outside the ellipsoid are designated as spikes or outliers. The three-dimensional phase space map or Poincar´e map is a simultaneous plot of a variable with its derivatives. As in standard outlier analysis a threshold region is determined which signifies ‘normal’ condition; the elliptical region fixed by the boundary threshold separates inliers from (multiple) outliers. The threshold used in this case is defined by a universal criterion. The universal threshold arises from a theoretical result in the landmark paper by Donoho and Johnstone (1994). For n independent, identically distributed, standard, normal random variables ξi the expected absolute maximum is,  E(|ξi |max ) = 2 log n = λU (3)

4.2 Experimental Datasets The experimental gearbox vibration data analysed in this study comes from an NEG Micon NM 1000/60 wind turbine in Germany. The measurements of the experimental data have been taken by members of the company EC Grupa, a Polish engineering company that maintains the wind turbine system from which the datasets were obtained. The gearbox consists of three gear stages: one planetary gear stage and two spur gear stages. The measurements come from a single accelerometer chosen by EC Grupa to be carrying more information concerning the damage in the gearbox. The sampling frequency of the measurements was 25000 Hz. Acceleration signals from the gearbox were obtained on three different dates: 31/10/2009, 11/2/2010 and 4/4/2010. The first dataset was indicated as the one to be 8

SAFEPROCESS 2015 September 2-4, 2015. Paris, France

Keith Worden et al. / IFAC-PapersOnLine 48-21 (2015) 001–017

used as a reference; this is generally sensible because one expects damage to increase. If it transpired that the system was undamaged at the first test, then one would see clear signatures of damage in the later data. If it transpired that damage was already present during the first test, one can still use it as a reference and look for increased signatures of damage later. The second dataset was considered to be one describing an early tooth damage of the gearbox and the third one was the dataset of the vibration signal with progressed tooth damage in the gearbox.

9

4

2

x 10

IMF 1

1 0 −1 −2 0

2000

4000

6000

8000

10000

12000

2000

4000

6000

8000

10000

12000

2000

4000

6000

8000

10000

12000

2000

4000

6000

8000

10000

12000

Sample point

4

Generally, it is known that it is in the harmonics of the meshing frequency of the damaged gear pair that damage features occur. The frequency components of the signal are the following:

1

x 10

IMF 2

0.5

• 15.07, 30.14 Hz: relative meshing frequency and second harmonic of the planetary gear, • 89.535, 179.07 Hz: relative meshing frequency and second harmonic of the 1st parallel gear stage, • 352, 705, 1056, 1410, 2115, 2820 Hz: relative meshing frequency and harmonics of the 2nd parallel gear stage.

0 −0.5 −1 0

Sample Point

4

1

x 10

0.5

IMF 3

In order to perform time-frequency analysis, the signals were decomposed using the empirical mode decomposition (EMD) into a set of signal components (oscillatory functions) in the time-domain called intrinsic mode functions (IMFs). The EMD is now well known in the signal processing field (Huang et al. (1998)); the procedure for extracting the IMFs from the signal analysed is known as the sifting process and is an empirical algorithm described in detail in Huang et al. (1998). Each extracted IMF (decomposed signal component in the time-domain) represents a frequency band in the frequency domain with the first IMF representing the highest frequency component of the signal and subsequent IMFs representing corresponding lower frequency regions of the signal. Figure 3 shows the EMD results (the first four of the 13 IMFs extracted) for data from the final monitored date 4/4/2010. More examples of data can be found in Antoniadou (2014) and Antoniadou and Worden (2014).

0 −0.5 −1 0

Sample Point

4000

IMF 4

2000 0 −2000 −4000 0

Sample Point

Fig. 3. First four IMFs of the wind turbine gearbox data from the date 4/4/2010.

In this example, the second IMF produced by the decomposition proved to be the one most sensitive to damage since it described in the frequency domain a harmonic of the meshing frequency of the damaged gear stage, this directly relates the visible impulses in the IMF to such damages as gear tooth faults Randall (1993). Envelope analysis can then be performed in order to enhance the impulsive nature of the damage features. Envelope analysis is most commonly performed using the Hilbert Transform (HT); however, an alternative approach is used here which exploits the Teager Kaiser energy operator and an energy separation algorithm. For certain signals and under appropriate circumstances, the new approach is known to improve results (Maragos et al. (1993); Antoniadou et al. (2015)). Finally, the square of the envelope of the second IMF is chosen in order to perform outlier analysis as this corresponds to the instantaneous power in the signal. Figure 4 shows the envelope results corresponding to the second IMF in Figure 3.

7

4

x 10

Power

3 2 1 0 0

360

720

1080

Gear Rotation [deg]

1440

Fig. 4. Power (TKEO/ Desa-1) of the 2nd IMF for the data from the date 4/4/2010. 4.3 Results and Discussion Concerning the outlier analysis results and the way the features were selected from the power diagrams, a 10dimensional feature was defined as a 10-point time window. One series of 200 features was defined as a reference 9

SAFEPROCESS 2015 10 September 2-4, 2015. Paris, France

Keith Worden et al. / IFAC-PapersOnLine 48-21 (2015) 001–017

and used for the training data. The training data were chosen carefully in order that no peaks (of the power measure) would be included in them. Whenever a fault appears, the outlier statistics (MSD) plot should show a peak that is distinct from the normal condition data. The results of the outlier analysis using the MSD, presented here in Figure 5 show that the method detects the impulsive peaks in the power. The results are presented in a log scale in the figure. There are a significant number of outliers in areas where the damaged gear rotation degrees does not coincide with the expected period of damage.

particular, and multiple outlier detection in general. The method is clearly very well suited to condition monitoring applications when the faults observed are gear tooth faults or bearing faults, as these are known to produce impulses in the vibration signals. 6

6

x 10

5

Power

4

3

2 3

10

0 0

10

360

720

1080

1440

720

1080

1440

720

1080

1440

Gear Rotation [deg]

6

10

1

10

x 10

9 8 0

10

7

Power

Mahalanobis squared distance

1

2

−1

10

0

360

620

Gear rotation [deg]

1080

1440

6 5 4 3

4

2 1

3

10

0 0

360

Gear Rotation [deg]

2

10

7

4 1

x 10

3.5

10

3 0

10

2.5

Power

Mahalanobis squared distance

10

−1

10

0

360

720

Gear rotation [deg]

1080

1440

2 1.5 1

4

10

Mahalanobis squared distance

0.5 3

10

0 0

360

Gear Rotation [deg]

2

10

Fig. 6. Outlier detection using the phase-space thereshold method for the power of the 2nd IMF for: (top) the 31/10/2009 dataset (training data 280-480); (middle) the 11/2/2010 dataset (training data 200-400); (bottom) the 4/4/2009 dataset (training data 350-550).

1

10

0

10

−1

10

0

360

720

Gear rotation [deg]

1080

1440

Fig. 5. Outlier statistics for the power of the 2nd IMF for: (top) the 31/10/2009 dataset (training data 280-480); (middle) the 11/2/2010 dataset (training data 200400); (bottom) the 4/4/2009 dataset (training data 350-550).

5. CASE STUDY II: DAMAGE DETECTION IN A WIND TURBINE BLADE Experimental measurements from vibration analysis were extracted from a 9m CX-100 wind turbine blade (see Fig.7) by researchers at Los Alamos National Laboratory (LANL) throughout a full-scale fatigue test conducted at the National Renewable Energy Laboratorys (NREL) and National Wind Technology Center (NWTC). The blade was harmonically excited at its first natural frequency using a Universal Resonant EXcitation (UREX) system (Dervilis et al. (2014)).

In order to eliminate the additional alarms detected and continue in an adaptive signal processing framework, the spatially adaptive thresholding method will be used and its results on the same dataset will be presented at this point. Figure 6 shows the final outlier detection results using the phase space method. The diagrams are very satisfactory; all of the power peaks have been detected, so in this particular case the method indeed manages to detect the points where there is an increase of power (presumed to be related to tooth damage) while at the same time labelling the points that do not have such a dramatic power increase as inliers. The results are indeed and improvement over the MSD and the application of the phase-space thresholding technique appears to be a very promising strategy for condition monitoring in

The blades are complicated structures with different materials combined and they are rotating continuously and simultaneously changing direction. In this study (described in considerably more detail in Dervilis et al. (2014)), methods such as Probabilistic Principal Component Analysis (PPCA) (essentially a Bayesian variant of PCA) were adopted in order to transform the raw measurements into a lower-dimensional representation. Once a particular 10

SAFEPROCESS 2015 September 2-4, 2015. Paris, France

Keith Worden et al. / IFAC-PapersOnLine 48-21 (2015) 001–017

11

Fig. 7. Wind turbine blade experiment. feature was obtained, a decision algorithm was trained to reveal the condition of the structure. The key point here is again the use of pattern recognition tools for the monitoring of turbine blades; in this case by using vibration data and high frequency response function measurements (FRFs) to provide damage-sensitive features (Dervilis et al. (2014)). Active sensing measurements were used to provide the data for the novelty detection methods. Within this active sensing system, two different sensor arrays were implemented called here the INNER and OUTER sensor arrays; they consisted of six and seven sensors respectively. An independent actuator was assigned to each of the arrays. The excitation frequency band chosen for the experiment was between 5 kHz and 40 kHz with a sampling rate of 96 kHz; this gave a resolution of 7200 spectral lines for the FRFs.

Fig. 8. Auto-associative neural network. It is clear how this works. If learning has been successful, then for all data in the training set ν({x}) ≈ 0 if {x} represents the normal condition. If {x} corresponds to damage, ν({x}) will be significantly non-zero. (Note that there is no guarantee that ν will increase monotonically with the level of damage, this is why novelty detection only gives a Level One diagnostic.) Note that universal approximation property of the neural network (Bishop (1995)) means that the AANN novelty detector can learn the properties of any normal condition distribution, it does not have to be Gaussian or even unimodal.

The dimension of the feature set if the raw FRFs were used was far too high, this was therefore reduced using PPCA. It was decided to retain the first five principal components and thus give a five dimensional feature (Dervilis et al. (2014)).

In the results presented here, the novelty indices for points from the training set are presented in a black colour, while those from testing data are blue. Figs. 9, 10, 11, 12 show the results of novelty detection based on FRFs from the two sensor arrays (INNER and OUTER) Each figure shows the results from two novelty detectors, respectively trained on low and high frequencies families of spectral lines from the FRFs for the same sensor (Dervilis et al. (2014)).

After selecting the training, validation and testing data for each of the 13 sensors labelled INNER and OUTER, a trained five-layer Auto-Associative Neural Network (ANN) was used via unsupervised learning to model the normal condition of the data for novelty detection. (Normal condition points are those coloured black in the figures presenting novelty detection results). The AANN was implemented as a standard Multi-Layer Perceptron (MLP) network (Bishop (1995)), which is asked to reproduce at the output layer, those patterns that are presented at the input. This would be a trivial exercise except that the network structure is given a ‘bottleneck’ i.e. the patterns are passed through hidden layers which have fewer nodes than the input layer (Figure 8). This constraint forces the network to learn the significant features of the patterns; the activations of the smallest, central layer, correspond to a compressed representation of the input. Training proceeds by presenting the network with many versions of the pattern corresponding to normal condition (sometimes corrupted by added noise) and requiring a copy at the output.

The lower ranges of the frequency responses are not as useful features as those from the higher ranges of the FRFs, as in most of sensors no strong indication of novelty is seen in the upper plots within the figures. In contrast, the high frequency ranges of the FRFs were proved to be a significant asset in detecting damage. The advantage of using high frequency vibrations to monitor the blade for damage is that the wavelength of the modes is smaller, so they are more sensitive to smaller cracks (Dervilis et al. (2014)). The results indicate that measured FRFs as features allow the novelty detection technique to be sensitive to small structural changes. Novelty detection applied to each sensor’s FRF measurements presented early outliers, about 20-25 days before a visible crack was visually observed on the surface of the blade. The initial damage was introduced internally and started from the main carbon spar (Dervilis et al. (2014)).

The novelty index ν({x}) corresponding to a pattern vector {x} is then defined as the Euclidean distance between the pattern and the result of presenting it to the network {ˆ x}, ν({x}) = ||{x} − {ˆ x}|| (10) 11

SAFEPROCESS 2015 12 September 2-4, 2015. Paris, France

Keith Worden et al. / IFAC-PapersOnLine 48-21 (2015) 001–017

0.5

2

Novelty Index

Novelty Index

0.4 0.3 0.2 0.1 0 0

100

200

300 Observations

400

Novelty Index

Novelty Index

0.5

100

200

300 Observations

400

150

200

250 300 350 Observations

400

450

500

400

450

500

4

2

50

100

150

200

250 300 Observations

350

Fig. 11. Low frequencies range novelty index (top) and high frequencies range novelty index for OUTER sensor 1 (bottom). 0.8

Novelty Index

0.6

0.4

0.2

100

200

300 Observations

400

0.6

0.4

0.2

500

0 0

1.5

50

100

150

200

250 300 350 Observations

400

450

500

50

100

150

200

250 300 350 Observations

400

450

500

4

1

Novelty Index

Novelty Index

100

6

0 0

500

0.8

0.5

0 0

50

8

Fig. 9. Low frequencies range novelty index (top) and high frequencies range novelty index for INNER sensor 2 (bottom).

Novelty Index

0.5

0 0

1

0 0

1

500

1.5

0 0

1.5

100

200

300 Observations

400

500

3

2

1

0 0

Fig. 10. Low frequencies range novelty index (top) and high frequencies range novelty index for INNER sensor 4 (bottom).

Fig. 12. Low frequencies range novelty index (top) and high frequencies range novelty index for OUTER sensor 3 (bottom).

6. CASE STUDY III: POPULATION-BASED SHM FOR AN OFFSHORE WIND FARM

to this problem is to adopt a population-based approach where one develops a methodology which allows inferences to be made between different structures. Such a methodology would mean that any damage state data from any structure in the population informs the detection and classification processes for all structures. Ultimately, one would wish for such a methodology to extend to heterogeneous populations, but for now, development work is perhaps best restricted to populations of nominally identical structures. In this respect, the offshore wind farm

6.1 Background The final case study will illustrate some first steps in how one might extend SHM from the monitoring of individual structures to populations of structures, or systems-ofsystems. As discussed earlier, one of the major issues in implementing SHM is the difficulty of obtaining damage state data for high-value structures. One possible solution 12

SAFEPROCESS 2015 September 2-4, 2015. Paris, France

Keith Worden et al. / IFAC-PapersOnLine 48-21 (2015) 001–017

discussed in the following provides an almost ideal basis for research on population-based SHM.

13

conventional farms (Dahlberg (2009)), and this is generally expected to affect their performance.

In terms of the monitoring of individual turbines, there have been several different proposed approaches, approaches for the monitoring of wind turbines, ranging from traditional non-destructive evaluation (NDE) and vibration-based approaches on the blades, to advanced signal processing and machine learning approaches on the gearboxes (as illustrated in previous sections). General reviews can be found in Hameed et al. (2009) and Garc´ıa M´ arquez et al. (2012). Most modern wind farms will contain some form of Supervisor Control and Data Acquisition (SCADA) system installed which can provide for the measurement and recording of several different variables, such as wind speed, bearing and oil temperatures, voltage, and power produced, among others. As the SCADA system records constantly and is primarily used to monitor and control plant, it forms an ideal basis for an online SHM approach. In addition, SCADA extracts are perhaps the most direct and potentially useful data obtained from wind turbines, except of course any direct measurements acquired on the turbines themselves (through accelerometers, laser vibrometry or any other sensor).

Fig. 13. Lillgrund wind farm and the distribution of the wind turbines (Dahlberg (2009)). The available data used in this study correspond to a full year of operation. All the SCADA extracts consist of ten minute averages, with the maximum, mean, minimum and standard deviation of the ten minute intervals being recorded and available. The actual sampling frequency is less than ten minutes, but it is not disclosed here.

The use of SCADA data for monitoring has been discussed in several studies, such as Zaher et al. (2009) and Kusiak and Li (2011), and in most cases it aims at developing a complete and automatic strategy for the monitoring of the whole turbine or wind farm, although sub-components (e.g. bearings, generator) may be individually assessed as well. Among the various approaches, power curve monitoring has been popular and successful. Wind turbines have been designed by manufacturers to have a direct relationship between wind speed and the power produced, and as they require a minimum speed to produce the nominal power, but limit the power generated from higher wind speeds, the power curve usually resembles a sigmoidal function which is readily modelled using regression methods. A critical analysis of methods for modelling the power curves can be found in Thapar et al. (2011); in general researchers have exploited the deviation from a reference curve to perform SHM on turbines i.e. a form of novelty detection is applied to the curves. The following discussion will show how power curves might be used in a population-based strategy founded on machine learning; previous studies using learning methods on power curves include Kariniotakis et al. (1996); Li et al. (2001); Gill et al. (2012); Kusiak et al. (2009); Marvuglia and Messineo (2012); Kusiak and Verma (2013). The algorithms used here are artificial neural networks and Gaussian processes.

6.3 Power Curve Monitoring Neural Networks In the current study, the most common ANN structure - the multi-layer perceptron (MLP) - is used. All the modelling here used the Netlab package in Matlab (Nabney (2001)). Since ANNs have been extensively and successfully used for nonlinear regression, they would seem ideal for learning the power curve of wind turbines. In the current case, the wind speed (from anemometers in each tower) is available, in ten minute averages, from the SCADA records. The SCADA data also provides a status for the operation of the turbines, usually in the form of an ‘error code’. For the creation of the healthy power curve here (the reference curve), data from the whole year were used, but only when they corresponded to time instances with a status code equal to ‘0’, which signifies ‘no error’ in the turbines. The one-year healthy data were separated into training, validation and testing sets as is best practice in machine learning. In the ANN case, the training set is used to establish the parameters (weights) of the model, while the validation set is used to optimise the structure for the network (e.g. number of neurons). In the end, the finally selected network is tested with fresh data in order to ensure that it can generalise to new data. Only the number of hidden neurons in the network is subject to optimisation, the number of neurons in the input and output layers are fixed by the dimensions of the input features and output targets; the number of hidden units here was allowed to range from 1 to 10 in the optimisation. The optimal number of training cycles used was also established from the validation step and was found to be 300 here. The measure of the goodness of the regression fit was provided by the normalised mean-square error (MSE) given in equation (11),

6.2 Description of the Wind Farm The Lillgrund wind farm is situated in the sea area between Denmark and Sweden and consists of 48 identical wind turbines of rated power 2.3 MW (Dahlberg (2009)); their distribution can be seen in Figure 13. (The original labelling of the turbines shown in the figure carried information in the letter and number codes, but for convenience here the turbines are simply numbered from 1 to 48). It is important to note that the spacing between the turbines in the specific wind farm is significantly closer than most 13

SAFEPROCESS 2015 14 September 2-4, 2015. Paris, France

M SE(ˆ y) =

Keith Worden et al. / IFAC-PapersOnLine 48-21 (2015) 001–017

n 100  (yi − yˆi )2 N σy2 i=1

(11)

45

4.5

40 Turbine training number

where the caret denotes an estimated quantity. In total, 48 different networks (one for each turbine) were finally selected to create reference power curves for the turbines. After that, each network was provided with wind speed data from the rest of the turbines and was asked to predict the power produced. In Figure 14 the MSE errors of each trained network, when tested with wind speed data for the turbine for which they were trained, and subsequently the remaining turbines, is shown. Each axis of the confusion matrix shown in Figure 14 corresponds to 1 up to 48 turbines, where on the y-axis is the number of the trained turbine and on the x-axis the number of the tested turbine. In general, an MSE error below 5.0 can be considered a good fit, and one below 1.0 is considered excellent (this is a subjective judgement based on previous experience with this MSE).

4

35

3.5

30

3

25 2.5 20 2

15

1.5

10

1

5 5

10

15

20 25 30 35 Turbine testing number

40

45

Fig. 14. Confusion matrix with MSE errors created from the neural networks - testing set.

From the results it is clear that almost all the trained networks are very robust (generalise well to other turbines) and the maximum MSE error is around 5.0, which mainly occurs in turbines 3 and 4, which are located in the outside row of the wind farm. It can also be seen that in the diagonal of the confusion matrix (which corresponds to the testing set of the trained turbines when tested with data from themselves), the MSE error is very low, with the maximum appearing in turbine 39 (MSE=1.4708), and the minimum in turbine 31 with MSE at 0.5408. When the trained networks for an individual turbine were fed data originating from the same network, but which did not correspond to ‘no error’ statuses, the MSE error was everywhere larger as can be seen in Figure 15, the lowest was 4.7991 which appeared in turbine 12 and it was still much larger than the value of 0.8262 for the healthy data. In turbine 4, for example, the MSE increased from 0.768 to 149.033 and the standard deviation of the regression error from 0.0593 to 0.3685. Subsequent examination of the data and the associated error codes revealed that the majority of the instances where the regression error becomes high (in turbine 4) happened when the turbine was not working, either from emergency stops or manual stops. The emergency stops are associated with faults, but as these results derive from actual working data, the types of faults are limited to what was present during the recording period. Essentially, Figure 14 shows a map of potential thresholds, which can be used for the monitoring (in a novelty detection scheme) of the turbines individually or as a population.

Non−healthy Healthy

2

MSE error − log scale

10

1

10

0

10

5

10

15

20 25 30 Turbine number

35

40

45

Fig. 15. MSE error of the neural network models when presented with data not corresponding to error code ‘0’. vector. A mean prediction and confidence intervals on this prediction can be calculated from the GP distribution. The initial step in applying GP regression is to specify mean and covariance functions, which express prior belief in the process being modelled. These functions are fixed by specifying a number of hyperparameters including the correlation length for input data and the anticipated noise variance. Here, a zero-mean function and a squaredexponential covariance function are applied (Rasmussen and Williams (2006)). The training algorithm for the GP, together with the means of optimising the hyperparameters (a maximum likelihood method) is explained in detail in Rasmussen and Williams (2006); the software used here was taken from an internet resource provided by the authors of that reference.

Gaussian Processes The power curve regression was conducted using another approach, using the Gaussian process (GP) algorithm. This is a research area of increasing interest not only for regression but also for classification purposes; readers are referred to Rasmussen and Williams (2006) for more details. The GP is a stochastic nonparametric Bayesian approach to regression and classification problems; it is computationally very efficient and nonlinear learning is relatively easy. GP regression takes into account all possible functions that fit to the training data vector and gives a whole predictive distribution for a given input

In Figure 16, a similar confusion matrix to that produced for the MLPs is shown. The results are very good, with the same level of robustness and similar levels of MSE error as for the MLP. The worst results are again in turbines 3 and 4. In terms of the comparison between ANNs and GPs, it appears that the results are very similar with the networks performing with a slightly lower MSE error. However, it should be noted that the GPs are trained with about a 14

SAFEPROCESS 2015 September 2-4, 2015. Paris, France

Keith Worden et al. / IFAC-PapersOnLine 48-21 (2015) 001–017

third of the data that the neural networks were provided (for reasons of computational efficiency), but the testing sets are everywhere the same.

15

2.5

45

4.5

40

4

35

3.5

30

Average MSE error

Turbine training number

2

3

25

2.5

1.5

1

0.5

20 2 15

0 0

1.5

10

10

1

5 5

10

15

20 25 30 35 Turbine testing number

40

40

50

Fig. 18. Average MSE error showing how well the power produced in each turbine is predicted by the networks trained in the rest of the turbines.

45

Fig. 16. Confusion matrix with MSE errors created from the Gaussian processes - testing set.

2.5

Figures 17 to 20 show the average MSE errors contained in the confusion matrices shown in Figures 14 and 16. Figures 17 and 19 show how well each trained (reference) power curve predicts the power produced in the rest of the turbines. Figures 18 and 20 show how well the power produced in each turbine is predicted by the curves trained from each of the rest of the turbines. All these figures show again, that the worst turbines are 3 and 4, which predict for, and are also predicted by, the rest of the turbines with a greater error than the rest. The low MSE errors show the potential of individual power curves as features for the monitoring of the whole farm.

Average MSE error

2

1.5

1

0.5

0 0

10

2.5

20 30 Turbine number

40

50

Fig. 19. Average MSE error showing how well each turbine predicts the produced power for the other turbines Gaussian processes.

2 Average MSE error

20 30 Turbine number

1.5 2.5 1 2

0 0

10

20 30 Turbine number

40

Average MSE error

0.5

50

Fig. 17. Average MSE error showing how well neural networks trained to predict the power produced in each turbine, predict the produced power in the rest of the turbines.

1.5

1

0.5

0 0

10

20 30 Turbine number

40

50

7. CONCLUSIONS Fig. 20. Average MSE error showing how well each turbine (power produced) is predicted by the others Gaussian processes.

Only the briefest conclusions appear to be warranted here. The objective of the paper has been to discuss the databased approach to damage identification and to show that 15

SAFEPROCESS 2015 16 September 2-4, 2015. Paris, France

Keith Worden et al. / IFAC-PapersOnLine 48-21 (2015) 001–017

there is commonality in how this can be applied in a number of different engineering disciplines. Although the basic similarities between damage identification methodologies are sometimes obscured by context-specific variations in terminology etc. the paper has attempted to bring them to light. It is argued here that the damage identification hierarchy proposed by Rytter in the SHM context is more broadly applicable as an organising principle and that, together with an appropriate data fusion strategy, databased methods provide an effective means of addressing the lower levels of diagnostics. The application of databased methods for damage ID to several case studies (all concerning real-life systems and structures associated with offshore wind turbines) has been presented; the studies show that data-based methods can provide sensitive diagnosis of damage for both individual structures and (potentially) for populations of structures or systems-ofsystems.

Chronkite, J. (1993). Practical application of health and usage monitoring (HUMS) to helicopter rotor, engine and drive system. In Proceedings of the 49th Forum of the American Helicopter Society. Dahlberg, J.A. (2009). Assessment of the lillgrund windfarm: Power performance, wake effects. Technical report, Vatenfall A. B. Dervilis, N., Choi, M., Taylor, S., Barthorpe, R., Park, G., Farrar, C., and Worden, K. (2014). On damage diagnosis for a wind turbine blade using pattern recognition. Journal of Sound and Vibration, 333(6), 1833–1850. Donoho, D. and Johnstone, I. (1994). Ideal spatial adaptation by wavelet shrinkage. Biometrika, 81, 425 – 455. Farrar, C.R. and Worden, K. (2007). An introduction to structural health monitoring. Philosophical Transactions of the Royal Society - Series A, 365, 303–315. Farrar, C.R. and Worden, K. (2012). Structural Health Monitoring: A Machine Learning Perspective. John Wiley and Sons. Friswell, M. (2007). Damage identification using inverse methods. Philosophical Transactions of the Royal Society - Series A, 365, 393–410. Garc´ıa M´arquez, F.P., Tobias, A.M., Pinar P´erez, J.M., and Papaelias, M. (2012). Condition monitoring of wind turbines: Techniques and methods. Renewable Energy, 46, 169–178. Gertler, J. (1998). Fault detection and diagnosis in Engineering Systems. Marcel Dekker. Gill, S., Stephen, B., and Galloway, S. (2012). Wind turbine condition assessment through power curve copula modeling. Sustainable Energy, IEEE Transactions on, 3(1), 94–101. Goring, G. and Nikora, V. (2002). Despiking acoustic Doppler velocimeter data. Journal of Hydraulic Engineering, 128, 117 – 126. Hameed, Z., Hong, Y., Cho, Y., Ahn, S., and Song, C. (2009). Condition monitoring and fault detection of wind turbines and related algorithms: A review. Renewable and Sustainable Energy Reviews, 13(1), 1– 39. Huang, N., Shen, Z., Long, S., Wu, M., Shih, H., Zheng, Q., Yen, N., Tung, C., and Liu, H. (1998). The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis. Proceedings of the Royal Society of London. Series A: Mathematical, Physical and Engineering Sciences, 454(1971), 903–995. Kariniotakis, G., Stavrakakis, G., and Nogaret, E. (1996). Wind power forecasting using advanced neural networks models. IEEE Transactions on Energy Conversion, 11(4), 762–767. Kusiak, A. and Li, W. (2011). The prediction and diagnosis of wind turbine faults. Renewable Energy, 36(1), 16–23. Kusiak, A. and Verma, A. (2013). Monitoring wind farms with performance curves. IEEE Transactions on Sustainable Energy, 4(1), 192–199. Cited By (since 1996)5. Kusiak, A., Zheng, H., and Song, Z. (2009). On-line monitoring of power curves. Renewable Energy, 34(6), 1487–1493. Li, S., Wunsch, D., O’Hair, E., and Giesselmann, M. (2001). Using neural networks to estimate wind tur-

ACKNOWLEDGEMENTS The authors would like to acknowledge EC Grupa and Prof. Wieslaw Staszewski and Dr. Tomasz Barszcz for providing the wind turbine gearbox data and their help in the condition monitoring research of this paper. The authors would also like to thank C.R. Farrar, K. Farinholt, S.G. Taylor from Los Alamos National Laboratories, USA, M. Choi from the Chonbuk National University, Korea, and G. Park from Chonnam National University, Korea, and LANL, USA, for their support and guidance on the study of the turbine blade experimental research. Finally, the authors would like to thank A.E. Maguire from Vattenfall Research and Development, for providing access to data and for help and support regarding the Lillgrund wind farm. REFERENCES Antoniadou, I. (2014). Accounting for nonstationarity in the condition monitoring of wind turbine gearboxes. Ph.D. thesis, Department of Mechanical Engineering, University of Sheffield, UK. Antoniadou, I., Manson, G., Staszewski, W., Barszcz, T., and Worden, K. (2015). A time-frequency analysis approach for condition monitoring of a wind turbine gearbox under varying load conditions. Mechanical Systems and Signal Processing, 64–65, 188–216. Antoniadou, I. and Worden, K. (2014). Use of a spatially adaptive thresholding method for the condition monitoring of a wind turbine gearbox. In Proceedings of 7th European Workshop on Structural Health Monitoring, Nantes, France. Bedworth, M. (1994). Probability moderation for multilevel information processing. Technical Report DRA/CIS(SE1)/651/8/M94.AS03BP032/1, Defence Research Agency, Malvern. Bedworth, M. and O’Brien, J. (1999). The omnibus model: a new model of data fusion. Technical Report Preprint, Defence Research Agency, Malvern. Bishop, C. (1995). Neural networks for pattern recognition. Clarendon Press Oxford. Chen, J. and Patton, R. (1999). Robust model-based fault diagnosis for dynamic systems. Kluwer Academic Publishers. 16

SAFEPROCESS 2015 September 2-4, 2015. Paris, France

Keith Worden et al. / IFAC-PapersOnLine 48-21 (2015) 001–017

bine power generation. IEEE Transactions on Energy Conversion, 16(3), 276–282. Lowe, D. (2000). Feature extraction, data visualisation, classification and fusion for damage assessment. In Oral Presentation at EPSRC SIDAnet Meeting, Derby, UK. Maragos, P., Kaiser, J., and Quatieri, T. (1993). Energy separation in signal modulations with application to speech analysis. IEEE Transactions on Signal Processing, 41(10), 3024 – 3051. Markou, M. and Singh, S. (2003a). Novelty detection: a review - part 1: statistical approaches. Signal Processing, 83, 2481–2497. Markou, M. and Singh, S. (2003b). Novelty detection: a review - part 2: neural network based approaches. Signal Processing, 83, 2499–2521. Marvuglia, A. and Messineo, A. (2012). Monitoring of wind farms’ power curves using machine learning techniques. Applied Energy, 98(0), 574 – 583. Montgomery, D. (2009). Introduction to Statistical Quality Control. John Wiley and Sons. Mori, N., Suzuki, T., and Kakuno, S. (2007). Noise of acoustic Doppler velocimeter data in bubbly flows. J. Eng. Mech., 133(1), 122–125. Nabney, I.T. (2001). Netlab: Algorithms for Pattern Recognition. Springer. Pimentel, M., Clifton, D., Clifton, L., and Tarassenko, L. (2014). A review of novelty detection. Signal Processing, 99, 215–249. Randall, R. (1993). Vibration-based condition monitoring: Industrial, aerospace and automotive applications. Wiley. Rasmussen, C.E. and Williams, C.K.I. (2006). Gaussian processes for machine learning. The MIT Press, Cambridge, MA, USA. Roweis, S. and Ghahramani, Z. (1999). A unifying review of linear gaussian models. Neural Computation, 11, 305– 345. Rytter, A. (1993). Vibration-based inspection of civil engineering structures. Ph.D. thesis, Department of Building Technology and Structural Engineering, University of Aalborg, Denmark. Sohn, H. (2007). Effects of environmental and operational variability on structural health monitoring. Philosophical Transactions of the Royal Society A, 365, 539–560. Thapar, V., Agnihotri, G., and Sethi, V. (2011). Critical analysis of methods for mathematical modelling of wind turbines. Renewable Energy, 36(11), 3166–3177. Worden, K., Manson, G., and Fieller, N. (2000). Damage detection using outlier analysis. Journal of Sound and Vibration, 229(3), 647–667. Worden, K., Staszewski, W., and Hensman, J. (2011). Natural computing for mechanical systems research: A tutorial overview. Mechanical Systems and Signal Processing, 25(1), 4–111. Worden, K. and Dulieu-Barton, J. (2004). Damage identification in systems and structures. International Journal of Structural Health Monitoring, 3, 85–98. Worden, K., Farrar, C.R., Manson, G., and Park, G. (2007a). The fundamental axioms of structural health monitoring. Proceedings of the Royal Society - Series A, 463, 1639–1664. Worden, K., Manson, G., and Surace, C. (2007b). Aspects of novelty detection. In L. Garibaldi, C. Surace, and ...

17

(eds.), Proceedings of DAMAS 2007, Torino, Italy, 3–16. Transtech. Zaher, A., McArthur, S., Infield, D., and Patel, Y. (2009). Online wind turbine fault detection through automated SCADA data analysis. Wind Energy, 12(6), 574–593.

17