Towards an Autonomous Damage Detection of Concrete Slab Bridges

(1)

University of Amsterdam

Masters Thesis

Towards an Autonomous Damage Detection of

Concrete Slab Bridges

Author: Jeremy Barisch-Rooney

Daily supervisor: dr. Árpád Rózsás Examiner: dr. Valeria Krzhizhanovskaya Second supervisor: dr. Arthur Slobbe Assessor: prof. dr. ir. A.G. Alfons Hoekstra

A thesis submitted in partial fulfilment of the requirements for the degree of Master of Science in Computational Science

in the

Computational Science Lab Informatics Institute

(2)

Declaration of Authorship

I, Jeremy Barisch-Rooney, declare that this thesis, entitled ‘Towards an Autonomous Damage Detection of Concrete Slab Bridges’ and the work presented in it are my own. I confirm that:

⌅ This work was done wholly or mainly while in candidature for a research degree at the University of Amsterdam.

⌅ Where any part of this thesis has previously been submitted for a degree or any other qualification at this University or any other institution, this has been clearly stated. ⌅ Where I have consulted the published work of others, this is always clearly attributed. ⌅ Where I have quoted from the work of others, the source is always given. With the exception

of such quotations, this thesis is entirely my own work. ⌅ I have acknowledged all main sources of help.

⌅ Where the thesis is based on work done by myself jointly with others, I have made clear exactly what was done by others and what I have contributed myself.

(3)

“I have just one wish for you – the good luck to be somewhere ... where you do not feel forced by a need to maintain your position in the organization, or financial support, or so on, to lose your integrity. May you have that freedom."

(4)

UNIVERSITY OF AMSTERDAM

Abstract

Faculty of Science Informatics Institute

Master of Science in Computational Science

Towards an Autonomous Damage Detection of Concrete Slab Bridges by Jeremy Barisch-Rooney

The probability of a deteriorating bridge to fail increases over time until it is no longer considered safe for use. Many bridges in Europe are reaching the end of their design life because of the post-war surge in bridge construction, yet maintenance can extend the life of a bridge. Sensor data can be used to detect damage1 _{in real-time without the delay or cost of a manual maintenance check.}

One of the biggest problems with damage detection of concrete slab bridges is the availability of sensor data from the damage state. Reasons for this are the installation costs of a sensing system and it is unclear how stakeholders of the bridge could benefit. However in order to estimate the potential benefit of such an installation, data must first be available for the purpose of analysis. A finite element (FE) model can be of help to generate data for analysis. In this thesis a FE model of a concrete slab bridge under operational conditions is built and analysed.

The FE model in this thesis is a 3D linear model of one concrete slab bridge in Amsterdam, the bridge 705. The mechanical behaviour of the model is validated against measurements taken during an experimental campaign. Time series of traﬃc that crosses the bridge is modeled based on data collected from Weigh-in-motion sensors. Time series of the eﬀect of temperature is modeled based on data from the NOAA (National Oceanic and Atmospheric Administration). Two forms of damage are modeled, the settlement of piers and areas of the bridge deck where cracking has occurred, which we refer to as “crack zones”. Concrete creep and shrinkage are modeled using a method found in the EuroCode standard. To aid other researchers in collecting data, all model components have been made available for re-use via a software package for Python named bridge-sim. A paper is being reviewed by the Journal of Open Source Software which is a peer-reviewed open-access journal that publishes articles about research software.

Results show that while the eﬀects of temperature, concrete creep and shrinkage are significant at some points on the bridge deck, they are relatively small close to the piers – our model indicates that pier settlement as low as 1.9 mm can be detected by comparing the diﬀerence between piers at 1_{Damage detection refers to determining whether damage is present or not in a structure. Damage localization,}

(5)

symmetrical points on the bridge. Investigation of crack zones in our model show that on a sunny day the greatest change in longitudinal strain due to a 0.5 m × 16.6 m crack zone is 23% (of the maximum strain in healthy condition) but primarily only in a 0.2 m2 _{area, and not likely to be} detected with a strain gauge installation. However for larger crack zones that may be the cause of an explosive blast the change in response is significant. A method for detecting crack zones by comparing diﬀerences in responses from sensors on symmetrical spans of the bridge is investigated. For this case we can detect a 2 m × 16.6 m crack zone in our model using vertical translation sensors placed 2 m apart.

(6)

Acknowledgements

I would like to extend an absolutely huge thank you to Árpád Rózsás and Valeria Krzhizhanovskaya. Clearly two very smart researchers, I am confident you can both achieve whatever you set your sights on.

Thank you to my mother, one of the most caring people I know, thank you for always looking out for me. And to my father for inspiring me, turning 70 and traveling around South America with a backpack smaller than what most children bring to school!

I also want to thank Arthur Slobbe and anyone else who gave me advice or had to put up with me during this time! Finally thank you to the many contributors and authors of open source software upon which this thesis is built!

(7)

Nomenclature

" Strain Stress ⌫ Poisson’s ratio Creep coeﬃcient E Young’s modulus G Shear modulus

MSL Parameter: the maximum length of any shell element TREF Temperature that the bridge concrete was set at Tbot Temperature at the bottom of the bridge deck Ttop Temperature at the top of the bridge deck

ULS Parameter: the amount of unit point-load simulations per traﬃc lane XXB Denotes stress or strain type: longitudinal on the bottom of the deck XXT Denotes stress or strain type: longitudinal on the top of the deck ZZB Denotes stress or strain type: transverse on the bottom of the deck X Longitudinal axis of bridge

Y Vertical axis of bridge Z Transverse axis of bridge ANN Artificial neural network

CTE Coeﬃcient of thermal expansion DOF Degree(s) of freedom

DSS Decision support software FE Finite element

NDE Non-disruptive evaluation

NDW National Data Warehouse for Traﬃc Information (Netherlands) NOAA National Oceanic and Atmospheric Administration (USA) OCSVM One-class support-vector machine

OpenSees Open source FE software for earthquake engineering simulation RNN Recurrent neural network

SHB Sydney Harbour Bridge SHM Structural health monitoring SVM Support-vector machine PyPI Python Package Index

(8)

4.8 Parameter Selection . . . 60 5 Analysis 63 5.1 Experiments . . . 63 5.1.1 Question 1 . . . 63 5.1.2 Question 2 . . . 63 5.1.3 Question 3 . . . 63 5.1.4 Question 4 . . . 64 5.1.5 Question 5 . . . 64 5.2 Results . . . 66 5.2.1 Question 1 . . . 66 5.2.2 Question 2 . . . 66 5.2.3 Question 3 . . . 67 5.2.4 Question 4 . . . 68 5.2.5 Question 5 . . . 68 5.3 Discussion . . . 73 5.3.1 Pier Settlement . . . 73

(10)

CONTENTS

5.3.2 Crack Zones . . . 75

6 Conclusions and Recommendations 77

6.1 Conclusions . . . 77 6.1.1 Re-usable System . . . 78 6.2 Recommendations . . . 79

References 82

Appendices 87

A Temperature Load Calculation 87

B Simulations on the Lisa Cluster 88

C Selection of Parameter ULS 89

D Convergence of Responses 90

E Materials Properties of Bridge 705 92

F Details of The Experimental Campaign 93

G Validation Against Static Load Tests 98

H Verification of Point-loads 115

I Verification of Pier Settlement 121

J Verification of Temperature Load 124

K Verification of Self-weight 126

(11)

List of Figures

1 Bridge 705 at Cornelis Lelylaan, Amsterdam . . . 2

2 Accelerometers on the Sydney Harbour Bridge . . . 11

3 A simple spring system . . . 16

4 Stress XXT in three loading scenarios . . . 21

5 Boundary conditions and pier labels for the generated FE model of bridge 705 . . . . 23

6 Photo of the supporting piers of bridge 705 . . . 24

7 Frequency domain of diﬀerent eﬀects a bridge is subjected to . . . 26

8 Example of a temperature profile through the bridge deck . . . 27

9 Temperature profile model for bridge 705 . . . 29

10 Temperature model for the year of 2019 . . . 31

11 Contour plots of the eﬀect of temperature on bridge 705 . . . 32

12 Distribution of vehicles used in traﬃc simulations . . . 33

13 Responses to concrete shrinkage over 100 years . . . 34

14 Responses to concrete creep over 100 years . . . 35

15 Young’s modulus EX for a crack zone . . . 37

16 Loading specification of Truck 1 . . . 39

17 R-squared of Y translation responses . . . 41

18 R-squared of strain XXB responses . . . 41

19 Influence lines of Y translation at sensors U25-U27 . . . 42

20 Influence lines of strain XXB at sensors O3, O4 and O9 . . . 43

21 Comparison of Y translation from static simulation and dynamic test . . . 44

22 Comparison of strain XXB from static simulation and dynamic test . . . 45

23 Strain XXB convergence at increasing distances from a point-load at A . . . 46

24 Y translation from a 100 kN point-load at A with AxisVM . . . 49

25 Y translation from a 100 kN point-load at A with OpenSees . . . 49

26 Stress XXB from a 100 kN point-load at A with AxisVM . . . 50

27 Stress XXB from a 100 kN point-load at A with OpenSees . . . 50

28 Y translation from settlement of pier 4 with AxisVM . . . 52

29 Y translation from settlement of pier 4 with OpenSees . . . 52

30 Stress XXB from settlement of pier 4 with AxisVM . . . 53

31 Stress XXB from settlement of pier 4 with OpenSees . . . 53

32 Strain XXB convergence at increasing distance from pier 4 . . . 54

33 Y translation from uniform temperature loading with AxisVM . . . 55

34 Y translation from uniform temperature loading with OpenSees . . . 55

35 Stress XXB from uniform temperature loading with AxisVM . . . 56

(12)

LIST OF FIGURES

37 Y translation due to self-weight with AxisVM . . . 58

38 Y translation due to self-weight with OpenSees . . . 58

39 Calculation of the response to two axles . . . 59

40 Screenshot of animation of responses . . . 61

41 Strain sensors available in simulation . . . 62

42 Positions of sensors for question 5, sensors placed 1 m apart . . . 65

43 Minimum Y translation responses over 100 years . . . 66

44 Y translation responses over 1 and 100 years at X = 41.675 m and Z = -4.2 m . . . . 67

45 Smallest sensor thresholds that do not result in a false positive when comparing the diﬀerence between longitudinal sensors. . . 68

46 Change in strain XXB from a 0.5 m x 16.6 m crack zone . . . 69

47 Change in strain ZZB from a 0.5 m x 16.6 m crack zone . . . 70

48 Change in Y translation from a 0.5 m x 16.6 m crack zone . . . 71

49 Receiver operating characteristic curves for 2 m crack zone . . . 73

50 Requirements for damage detection of concrete slab bridges . . . 77

51 Convergence of Y translation as a function of ULS . . . 89

52 Convergence of Y translation as a function of nodes in FE model . . . 90

53 Convergence of strain XXB as a function of nodes in FE model . . . 91

54 Truck 1 positions in the experimental campaign . . . 94

55 Noise from Y translation sensors in experimental campaign . . . 96

56 Noise from strain XXB sensors in experimental campaign . . . 97

65 Influence lines of Y translation at sensors U28 and U29 . . . 106

66 Influence lines of strain XXB at sensors T1-T3 . . . 107

72 Influence lines of strain XXB at sensors T19 and T20 . . . 113

(13)

LIST OF FIGURES

74 Y translation from a 100 kN point-load at B with AxisVM . . . 115

75 Y translation from a 100 kN point-load at B with OpenSees . . . 115

76 Stress XXB from a 100 kN point-load at B with AxisVM . . . 116

77 Stress XXB from a 100 kN point-load at B with OpenSees . . . 116

78 Y translation from a 100 kN point-load at C with AxisVM . . . 117

79 Y translation from a 100 kN point-load at C with OpenSees . . . 117

80 Stress XXB from a 100 kN point-load at C with AxisVM . . . 118

81 Stress XXB from a 100 kN point-load at C with OpenSees . . . 118

82 Y translation from a 100 kN point-load at D with AxisVM . . . 119

83 Y translation from a 100 kN point-load at D with OpenSees . . . 119

84 Stress XXB from a 100 kN point-load at D with AxisVM . . . 120

85 Stress XXB from a 100 kN point-load at D with OpenSees . . . 120

86 Y translation from settlement of pier 5 with AxisVM . . . 121

87 Y translation from settlement of pier 5 with OpenSees . . . 121

88 Stress XXB from settlement of pier 5 with AxisVM . . . 122

89 Stress XXB from settlement of pier 5 with OpenSees . . . 122

90 Strain XXB convergence at increasing distance from pier 5 . . . 123

91 Y translation from linear temperature loading with AxisVM . . . 124

92 Y translation from linear temperature loading with OpenSees . . . 124

93 Stress XXB from linear temperature loading with AxisVM . . . 125

94 Stress XXB from linear temperature loading with OpenSees . . . 125

95 Stress XXB due to self-weight with AxisVM . . . 126

96 Stress XXB due to self-weight with OpenSees . . . 126

97 Change in strain XXB from a 1 m x 16.6 m crack zone . . . 127

98 Change in strain ZZB from a 1 m x 16.6 m crack zone . . . 128

99 Change in Y translation from a 1 m x 16.6 m crack zone . . . 129

106 Change in strain XXB from a 20 m x 16.6 m crack zone across 2 piers . . . 136

107 Change in strain ZZB from a 20 m x 16.6 m crack zone across 2 piers . . . 137

108 Change in Y translation from a 20 m x 16.6 m crack zone across 2 piers . . . 138

109 Change in strain XXB from a 14 m x 16.6 m crack zone across 1 pier . . . 139

110 Change in strain ZZB from a 14 m x 16.6 m crack zone across 1 pier . . . 140

(14)

List of Listings

1 Definition of the Bridge model for Python . . . 25

2 Instance of the Bridge model for bridge 705 . . . 25

3 Response to a vehicle from unit point-load simulations . . . 60

4 Contour plot of a point-load with bridge-sim . . . 78

5 A script with scheduling parameters for the Lisa cluster . . . 88

(15)

List of Tables

1 Structures in the Dutch national main road network . . . 4

2 Types of damage from short-term and long-term events . . . 6

3 Classification of existing literature based on the data used . . . 15

4 Material properties for crack zones . . . 36

5 Material properties for AxisVM comparison . . . 47

6 Loading positions for comparison with AxisVM . . . 48

7 Greatest diﬀerence in response after introducing a crack zone . . . 72

8 Area where diﬀerence in response is 80% of greatest diﬀerence in Table 7 . . . 72

9 Material properties of bridge 705 . . . 92

10 Y translation sensors in the experimental campaign . . . 93

(16)

1 INTRODUCTION

1 Introduction

1.1 Motivation and Problem Statement

The probability of a deteriorating bridge to fail increases over time until it is no longer considered safe for use. The Nanfang’ao bridge collapsed in Taiwan on 1 October 2019 killing 6 people and injuring 12, the cause is unclear but preliminary evidence points to corrosion of suspension cables. The Morandi bridge in Genoa collapsed in 2018 killing 43 people, possibly due to corrosion or construction error but an article [1] on the matter also hints strongly to neglect of maintenance as a cause, it is not clear. More recently another bridge collapsed in Italy, in Tuscany, injuring 2 people, a crack had developed the previous year and had been investigated but the bridge was cleared for further use. Maintenance of a bridge is typically carried out when something goes wrong or according to a preventative maintenance schedule based on expert knowledge, neither approach making the best use of limited maintenance resources.

Many bridges in Europe are reaching the end of their design life because of the post-war surge in bridge construction, yet maintenance can extend the life of a bridge. Sensor data can be used to detect damage1 _{in real-time without the delay or cost of a manual maintenance check. The research} question of this thesis is: what sensor installation is appropriate on bridge 705 for damage detection and further research, and what extent of damage can we expect to be detected? In order to answer this question we must first answer the question of how we can collect sensor data for the purpose of damage detection? Only then can we answer the question of how we can detect damage in the acquired data. Related questions include: what types of damage are bridges subjected to? And how is maintenance and damage detection of bridges currently carried out?

1.2 Approach

In order to classify data as coming from either a damaged or healthy bridge it is of course necessary to have sensor data corresponding to both the damaged and healthy state of the bridge, otherwise we could not evaluate the accuracy of the classification techniques applied. Due to the fact that bridges are expensive structures, it is unsurprisingly not permitted in the general case to apply damage to a bridge. The use of FE software allows us to create a simulation of a bridge to which damage can be applied. The responses of the structure collected from simulation can be used for the purpose of damage detection.

In this thesis a FE model of bridge 705 in Amsterdam is built. The reason that bridge 705 is chosen is because measured sensor responses of this bridge from an experimental campaign are avail-able for the purpose of validating the FE model. Bridge 705 is a non-prestressed non-posttensioned reinforced concrete slab bridge in Amsterdam, close to Lelylaan station. While the methods in this thesis are limited in scope to one concrete slab bridge, bridge 705, the methodology is believed to have a wider application. A photo of the bridge 705 is shown in Figure 1.

(17)

1 INTRODUCTION 1.2 Approach

A finite element model is a mathematical representation of a mechanical model of a structure, that can be simulated via computation. All models are wrong but some are useful, a FE model is an idealized representation of reality, some errors are associated with the mathematical model and some with the numerical solution, known as errors of idealization and errors of discretisation respectively. The sensor data that is generated with FE software will be diﬀerent from sensor data collected from a real bridge, in part due to environmental and operational eﬀects. However the simulated sensor data should be close enough to real data to provide confidence that the damage detection applied can also work with real sensor data, while acknowledging that some additional work would likely be needed to tune the analysis techniques once real sensor data is available. The mechanical behaviour of the FE model built in this thesis is validated against experimental measurements in order to provide confidence in the model.

Figure 1: Bridge 705 at Cornelis Lelylaan, Amsterdam. This photo is courtesy of TNO. SHM (structural health monitoring) is a term used to describe a range of systems implemented for the purpose of informing about the “fitness of purpose” of a structure. A major part of any SHM system has to be geared towards long-term evaluation of the “healthy” state of the system [2]. This is because environmental and operational effects such as temperature can affect the response of the structure enough to signal a departure from the structure’s healthy state. In this thesis a number of effects are modeled including traffic, temperature, pier settlement, concrete creep and shrinkage, and crack zones.

Finally I note that all of the model components used in this thesis are provided for re-use. This is achieved by not depending on expensive proprietary software, by “lifting” parameters to the boundaries of the system, and by publishing a system of composable functions that present the problem domain at a high-level of abstraction. After reading this thesis the large amount of work that went into data collection will become clear. It is my sincere hope to facilitate further research in the area of damage detection of concrete slab bridges and to prevent any duplication of eﬀort.

(18)

1 INTRODUCTION 1.3 Scope and Limits

Any interested party should be able to download this work and swiftly move to the application of damage detection methods.

1.3 Scope and Limits

Any automated damage detection system for bridge maintenance consists of a sensor installation that is hooked upto a software system that provides the user of the system with information on the current state of a bridge. The provided information should enable the operator of the system to make a more informed decision about when and/or where maintenance should be carried out. The provided information can include real-time sensor data and an analysis thereof. Intelligence augmentation, where human expertise is augmented by artificial intelligence (AI) techniques is a natural step in the cyclical development of fully automated AI-based systems. An example of this step in development is the car company Tesla’s “auto-pilot” for self-driving on highways which is only a step in the direction to full self-driving but has shown enough value to be produced and will allow for the collection of huge amounts of data for the future improvement of the system. While AI can extend human’s cognition with computational processing capacity, humans can at this point still oﬀer a more holistic, intuitive approach in decision making [3]. In this thesis we hope to motivate the installation and collection of sensor data from bridges, for further study and development of an automated SHM system.

Evaluation of a SHM system for bridges requires that the SHM system has been installed prior to the occurrence of damage. However for the economical decision of installing and maintaining a long-term SHM system to be taken, someone needs to be convinced of the value of the system, thus installation, development and evaluation of an SHM system will be a gradual cyclical process. The economical constraint into research of damage detection of bridges is also why research is largely based on numerical simulations of bridges, in which case environmental and operational eﬀects are typically missing in the analysis. Or damage is applied to a bridge before it is decommissioned, in which case the bridge can no longer be put under operational load. For these reasons any SHM system that is installed will likely be used in combination with operator expertise to prioritise maintenance or damage investigation, as the SHM system’s long-term accuracy is evaluated and the system improved. According to [4], while a SHM system should be capable of a minimal amount of damage assessment the more likely scenario is that an additional investigation is triggered by the system.

1.4 Thesis Structure

This thesis continues in Section 2 with an overview of the existing literature. In Section 3 some general and background information on methods used in this thesis is given. Section 4 details the approach used in this thesis. In Section 5 an analysis of the generated sensor data is provided and finally in Section 6 a conclusion of the work and some recommendations is presented.

(19)

2 LITERATURE REVIEW

2 Literature Review

The goal of this Section is to summarise the existing body of work related to the subject area of this thesis. This Section thus outlines the existing work on damage identification and SHM of civil infrastructure, with a particular focus on bridges (including some information on bridges in the Netherlands). While the amount of literature related to SHM is vast with numerous books written on the topic, the literature related to SHM of bridges is a little smaller, and the focus on concrete slab bridges is smaller again. In particular there seems to be much more research on SHM of expensive bridges such as the Sydney Harbour Bridge (SHB) rather than less expensive but numerous concrete slab bridges. The literature is vast therefore a complete literature review is not presented but the papers that are visited are believed to be representative.

2.1 Bridges in the Netherlands

The Dutch national main road network consists of 3,200 km of road. Assets in the road network are divided into four categories: pavements, structures, traﬃc facilities and environmental assets. Each structure is categorized into a type that has its own maintenance characteristics. Table 1 outlines the categorization of the 3,283 structures in the Dutch network. The aim of bridge maintenance is eﬀective management of bridge stock and realistic budgeting, taking into account the entire bridge stock.

Bridges are a type of structure that require a large investment, though they also have a long service life of 50 to 100 years. Annual maintenance costs are relatively small compared to the initial investment cost (<1%), however over the lifetime of the bridge the maintenance costs are on the order of the initial investment [5]. The annual maintenance cost of concrete bridges in the Dutch stock is €37 million. In an ageing bridge stock the cost of maintenance can be assumed constant, averaged over the large number of structures. After a long time the cost of replacement will be approximately €85 million, approximately half the cost of annual maintenance of concrete bridges at €37 million [5].

Table 1: Structures in the Dutch national main road network. Each type of structure has its own maintenance characteristics. The table lists for each structure type the total number in the Dutch national main road network and the total deck area. This table is from [5].

Structure type Number Deck Area (m2) Concrete bridge 3,131 3,319,002 Steel bridge (fixed) 88 301,997 Movable bridge 43 347,876

Tunnel 14 475,228

Aqueduct 7 86,491

Total 3,283 4,530,593

The Dutch national road network contains over 3,000 highway bridges. Of these, most are 30 or more years old. A significant amount of bridges were constructed in the 1970s, which is typical

(20)

2 LITERATURE REVIEW 2.2 Damage Types of Concrete Bridges

for many Western European road networks. The demolition of many Dutch bridges has been due to functional or economical requirements, before the design life of the bridge has expired, rather than due to technical failure. Design codes in the Netherlands require a design lifetime between 50 and 100 years. From the combination of the design life of bridges and the fact that may bridges were built post war, then a large peak can be expected in the maintenance and construction cost of Dutch bridges [5].

2.2 Damage Types of Concrete Bridges

Damage to a bridge can be classified as short-term or long-term. Short-term damage events are defined as a change in the properties of structural materials and elements, and of the behaviour of the whole structure, due to eﬀects that occur during a very short period of time such as a collision or fire. Long-term damage events are those that occur gradually over a longer period of time. Table 2 outlines some of the predominant types of damage due to short-term and long-term scenarios respectively.

When damage is detected based on sensor measurements another possibility is of course that the sensor is itself faulty. Sensors can become faulty for a number of reasons, increased noise, bad installation, battery issues, harsh environment et cetera [6]. Maintaining a healthy sensor network is important because faulty sensors can cause not only permanent loss of data but also inaccurate damage detection, if for example the detection system is being trained on the faulty sensor data.

2.3 SHM and Damage Detection

2.3.1 Overview

The question of damage detection is to identify if a system has departed from normal (i.e. un-damaged) condition, simply “is there damage or not?”. The more sophisticated problem of damage identification seeks to determine a greater level of information on the damage status, even to predict the future of the situation. The problem of damage identification can be considered as a hierarchy of levels (Rytter’s hierarchy) as described in [8].

Current state of the art health monitoring methods only indicate whether damage has occurred in a bridge, not determining the location or severity of the damage. It is suggested that another level could be added to Rytter’s hierarchy, level 2.1, the type of damage that occurs. If a certain type of damage is detected at a location then that will likely help with the assessment of the structure’s health (level 3).

(21)

2 LITERATURE REVIEW 2.3 SHM and Damage Detection T able 2: T yp es of damage from sh ort-term and long-term ev en ts. The first column giv es the name of the dama ge typ e, and wh ether it is a short-term or lon g-term damage ev en t. The second column giv es examples of the typ e of damage and the third column indicates the part of the bridge that w ou ld be aff ected under suc h an ev en t. Thi s T able is adapted from [7 ]. Short term ev en t Examples Comp onen t Collision Impact by vehicle or boat in the ri ver Pier Blast Impact by vehicle follo w ed by explosion Pier Fire Impact by vehicle follo w ed by explosio n and fire All Prestress loss Sudden failure of a pres tress tendon D ec k girder Abnormal loading cond itions Loading concen trat ion and/or ov erloading Dec k girder Excessiv e vibrati on E arthquak e Pier Impact Impact pr essure by w ater and debris during flo ods Substruct ure Long term ev en t Examples Comp onen t Corrosion Degra dation of bearings Dec k Loss of cross-sectiona larea in prestressing tendons Dec k Time-dep enden t prop er ties of Excessiv e creep & shrink age deformations Dec k the structural mater ials Concrete de terioration All Fatigue (lo w stress -high freque ncy) High freq uency and magnitude of tra ffi c loads Dec k Fatigue (hi gh stress -lo w freque ncy) T emp erature induced cyclic loading Abut men t En vironmen tal eff ects Freezing w ater leading to concrete expansion All W ater infilt ration/Leaking Deterioration of the expansion join ts; concrete Dec k degradation in the zone of the tendon anc horages Dec k Pier settlemen t Change in soil prop erties Dec k

(22)

2 LITERATURE REVIEW 2.3 SHM and Damage Detection

• Level 1. (Detection) indication that damage might be present in the structure. • Level 2. (Localization) information about the probable position of the damage. • Level 3. (Assessment) an estimate of the extent of the damage.

• Level 4. (Prediction) information about the safety of the structure.

In the USA the Federal Highway Administration (FHA) requires that the condition of bridges be evaluated every two years. Such an inspection typically takes the form of a tap test. A tap test is a test where the surface of the bridge is tapped in order to find variations in the sound response. However the tap test is limited to finding damage near the surface and in cases, significant cracks. And consider that in the USA there are over 500,000 highway bridges with a span length of over 7 m, inspection of all these bridges with a limited budget and staﬀ is not always achievable. Non-destructive evaluation (NDE) refers to methods of determining the location of damage without damaging the structure, for example with guided ultrasonic waves. NDE can be time consuming and expensive, and access to a location on the bridge may be diﬃcult or not even possible.

Many global health monitoring techniques are based on finding changes in resonant frequencies or mode shapes, as discussed later in this literature review. However for concrete structures the deterioration of reinforced steel has little eﬀect on natural frequency. Some methods attempt to find the location and length of cracks based on natural frequency, however these methods assume the only damage is cracking, an assumption that will certainly not always be true in a real-life setting. In the USA two bridge management systems (BMS) are used for the prioritization of funds and for the recommendation of actions for each bridge, these are PONTIS [9] and BRIDGIT [10]. Both of these deployed systems depend on the manual inspection of bridges for the acquisition of data. In [11] a decision support system is developed for concrete bridge deck maintenance is proposed, however the system has self-proclaimed flaws including a lack of bridge deck problems analyzed, and data from the inspection of a bridge is required.

2.3.2 SHM Applications

Much of the early research into damage identification of civil infrastructure was based on identifying modal properties, detecting damage by classifying changes in natural frequency or mode shape. Model-updating methods are based on having a model available and attempt to minimize the error between the model and real measurements by modifying model parameters with an optimization algorithm, in order to determine the state of the structure. Research into damage detection has over the years turned to the use of machine learning, in particular to unsupervised learning methods that do not require a model.

Damage was applied to the I-40 bridge, a 130 m girder bridge over the Rio Grande river, before it’s demolition, and data recorded from ambient vibration tests. The damage was intended to simulate fatigue cracking and was inflicted with torch cuts in a girder. In the fourth and most

(23)

severe damage state the web of the girder contained a 1.8 m cut and the flange was completely cut through. In [12] it is noted that changes in dynamic properties were only observed in the fourth damage state. Furthermore, changes of similar magnitude were observed from repeated ambient vibration tests on the undamaged structure.

In [13] introduced the use of the curvature of mode shapes which is obtained by diﬀerentiating the displacement mode shape twice. Changes in the curvature of the mode shape are localized to the damage and furthermore the absolute diﬀerence of the curvature mode shapes of the damaged and undamaged structures increase with damage severity [14]. However the [13] study was on a computer model of a beam, and did not consider robustness to noise.

In [15] changes in mode shapes, from the same I-40 experimental data as in [12], were shown to be statistically diﬀerent from the undamaged state for all damage states, however the analysis could not discriminate whether the source of the change was structural damage. The damage in the fourth damage state was localized, however at this point the bridge was sagging by 2 cm at the damage location, and according to [16] the bridge would have collapsed under a live load.

The 64 m concrete Dogna bridge in Italy was built in 1978 and suﬀered from a strong flood in 2003. In 2008, prior to demolition, an experimental campaign was carried out where six damage configurations were applied to the bridge in the form of notches cut with a hydraulic saw. In [17] changes in modal curvature were successfully used to identify the location of the damage. However the dynamic tests were all carried out under similar environmental conditions, thus the robustness to noise was not investigated.

In [18] a model-updating approach was applied which minimizes the diﬀerence in mode shapes. This approach was validated on the Z24 highway bridge in Switzerland, which is a 14 m 30 m -14 m pre-stressed concrete bridge. The damage scenario considered was the lowering of one of the supporting piers (originally at a height of 44 m) by 95 mm. In this study only a single damage scenario was considered and environmental eﬀects such as temperature which could represent a false positive damage scenario were not considered.

Model-updating approaches compare measurement data with responses from an analytical model and attempt to minimize the diﬀerence by updating model parameters. One problem with opti-mization algorithms used to update model parameters is that they may find a local rather than a global optimum. Evolutionary algorithms are good candidates for such problems and in [19] the particle swarm optimization algorithm is used as a model-updating approach using vibration data. The approach was experimentally verified against data from a 129 m railway viaduct.

In [20] a model-updating procedure is used to determine the location and the extent of damage on a rigid frame and then to assess the safety of the structure. This method is presented as part of a methodology for the continuous assessment of “the safety of civil engineering structure e.g. bridge, frames and oﬀshore platforms”. However this paper highlights two issues common in the literature: data corresponding to a baseline “healthy” state is required, and second, robustness to noise is not addressed in the work. The requirement of “baseline” data could be addressed ideally

(24)

if 1) the baseline state comes from sensor measurements taken for newly built structures, or 2) existing structures could be monitored for changes after sensor installation, not knowing whether the structure was already damaged or not. The robustness to noise is a more crucial problem because civil structures will be subjected to environmental factors such as temperature changes and ambient vibration. In the research [20] it simply states “the existence of noise in the data processing should be addressed”.

Health monitoring based on a numerical model imposes a challenge simply because the numerical model is required, and the necessary data for building a numerical model is not always available. This is because civil infrastructure is not always built precisely to the original design, due to changes in orders or due to on-site construction constraints. Moreover, in the case of concrete, uniform material properties are not guaranteed.

A Bayesian probabilistic approach was applied in a laboratory test to a reinforced-concrete bridge column [21], this method compared the relative damage probabilities of diﬀerent damage events based on data from vibration tests. The method has the potential advantage of not requiring an accurate numerical model, yet the study was only on a single column of a bridge and it was a laboratory experiment that did not account for environmental noise.

2.3.3 SHM Applications using Machine Learning

Machine learning can broadly be split into two variants, supervised and unsupervised learning. Supervised learning methods map inputs to outputs based on previously given input-output pairs known as labeled training data. Thus for damage detection, supervised learning methods require the existence of data corresponding to damage states, which is unlikely in the case of expensive civil infrastructure such as bridges. Unsupervised learning methods attempt to find previously unknown patterns in data set without pre-existing labels. One-class classification is a method of outlier detection, sometimes referred to as a semi-supervised learning method, where only one class of the data is available for training.

In [22] a number of damage identification experiments were applied that attempted to identify damage on an aircraft wing. The study showed damage localization and assessment to be possible with machine learning methods however the experiments were in a controlled laboratory setting without any environmental factors present. In the same paper it is argued that “damage prediction cannot be addressed by machine learning methods in general”. In the paper’s conclusion they reason about this saying “[prediction] is likely to be very context specific and to rely critically on understanding the physics of the damage progression”, “For example . . . the constants of the Paris–Erdogan equation [to predict crack development] are strongly dependent on micro-structure and would probably have to be treated as random variables in a given prediction”, continuing “Another major stumbling block in the application of prognosis is that most realistic situations will not be backed up by applicable theory, i.e. the laws of damage progression are not known for simple materials with complicated geometry, or for ‘complicated’ materials like composite laminates

(25)

even with simple geometries.”. Any reader interested in damage detection using machine learning methods is urged to read this informative paper, at least the conclusion.

In [16] a FE model of the 214m Clifton suspension bridge in Bristol, England is used to generate data corresponding to healthy and damaged states, namely damage to the girders. Environmental factors were considered by heating one side of the model by 30 C. In order to generalize the classification problem, data was generated by simulating a vehicle moving at 3 diﬀerent speeds. The vehicle was simulated using 2 concentrated loads, one per axle. Features were extracted from simulated vibration data and given as input to two unsupervised neural networks. The better-performing of the two was DIGNET [23] with a damage detection rate of 70%.

An ANN is used to detect damage from dynamic responses from a FE model of a railway bridge in [24]. To accomplish this an ANN is trained on past acceleration responses from the healthy bridge and then used to predict future values, the difference between predicted and measured data are used as a damage indicator. While prediction of subsequent acceleration data was possible, the loading applied (a train) was appropriate for the conditions (a railway bridge) but it is a much simplified case in contrast to a highway bridge that may have multiple lanes of traffic. Furthermore the authors suggest further work regarding the effect of environmental and operational effects.

The Sydney Harbour Bridge is a steel-reinforced concrete bridge built in 1932. The SHB consists of 800 jack arches in longitudinal direction. Each jack arch was fitted with 3 accelerometers. It was known apriori that one of the arches was cracked. Damage detection from acceleration data collected from the sensors on the SHB was examined in a few papers. In two of these papers, unlike any of the works discussed so far, damage detection methods made use of knowledge of the geometry of the bridge.

[25] uses the idea that if an arch on the SHB is healthy then accelerometers would move together, if there is a crack then they would move diﬀerently, see Figure 2. An SVM was trained using labeled data from features combining data from the sets of 3 accelerometers on each arch. A one-class SVM (OCSVM) which is an unsupervised variant of the SVM that is trained only on the healthy data, was also tested. The supervised variant achieved an accuracy of approximately 97% and the unsupervised approximately 71%. In [26] an algorithm is suggested to improve selection of the Gaussian model parameter of the OCSVM, which improved damage detection accuracy on the same SHB data set. Two methods were applied in [27] using the idea that similar substructures should behave simi-larly. K-means clustering was applied to the features collected from each arch. With k=2 k-means clustering, considering only 6 arches, including the one known damaged arch, a cluster was formed containing features primarily from the damaged joint. This method did not perform well when the amount of arches considered was increased to 71. The other method applied in [27] considered a “joint representative”, a feature that is the mean of the features from one arch. Then a pairwise map was created using the Euclidean distance between each pair of joint representatives. This method detected the known damaged arch, another arch with a known faulty sensor and a third arch with unknown damaged state.

(26)

2 LITERATURE REVIEW 2.4 Lessons Learned in Practice

Figure 2: Accelerometers on the Sydney Harbour Bridge. Sensors spanning a severe crack are expected to move diﬀerently (relative to each other) than those on a solid piece of concrete. This image is from [25].

In [26] a OCSVM with modified Gaussian kernel function is shown to have significantly improved performance on benchmark datasets than previously used kernels, however on the dataset from the SHB the accuracy improvement was ⇡ 2%. Research into damage detection of levees also leveraged the OCSVM in [28], and a multivariate gaussian method [29]. The OCSVM had a 91% accuracy and the multivariate Gaussian technique an accuracy of 97%. This research on levees however was collected under “artificial conditions” and does not consider the environmental conditions that may trigger anomalous conditions in real life.

2.4 Lessons Learned in Practice

2.4.1 Noise

Any structural health monitoring system that is deployed on a real-life structure must consider the environmental and operational effects that will affect the responses of the bridge. Temperature can change the stiffness properties of a bridge resulting in different responses throughout a day or year, and noise from traffic on another lane will also make damage identification more difficult.

A regression analysis was applied to acceleration data from the Alamosa Canyon Bridge in New Mexico in [30]. The natural frequency varied approximately 5% during the 24-hour interval when measurements were taken and the frequency was well correlated with bridge deck temperature from two spatially separated sensors. Temperature measurements from one of the bridge deck sensors ranged from 24 C to 45 C in one of the two datasets (collected during a 24-hour interval) and the eastern and western sides of the bridge showed a large temperature gradient, because the bridge is oriented north to south.

In [31] a linear relationship is shown between the 1st_{and 2}nd_{eigenfrequencies of the Z24 bridge in} Switzerland and air temperature above 0 C, and a separate linear relationship with air temperature below 0 C. The bi-linear relationship was related to the presence of the asphalt on the bridge.

In [32] a number of models are proposed to show the relationship between natural frequency and temperature recorded from sensors on the 44 m Dowling Hall footbridge, these included a bi-linear model and 4th _{polynomial order models with and without cross terms, all models performed well.}

(27)

In [33] a regression based approach, Regression-Based Thermal Response Prediction (RBTRP), is suggested to be integrated in a SHM system for bridges, the paper suggests a data collection reference period of 1 year and a minimum of 24 samples per day in order to remove the daily temperature eﬀect. Experiments in the paper measures strain and “ambient temperature in the vicinity of the test-best”. This paper from 2016 also suggests a method for anomaly detection based on computing the diﬀerence between measurements and predictions using RBTRP. The authors state that “further work will also focus on defining flexible thresholds that account for long-term shifts in response due to expected and normal changes in material behaviour such as from creep”.

An integrated machine learning algorithm, combining techniques including PCA, is presented in [34] for separating the individual components of the deflection signal into components with separate frequencies. When the noise level was under 10%, each component (temperature, live load, structural damage) was successfully separated based on data from a computer model of a long-span bridge. A linear relationship between temperature and deflection was assumed. Temperature was decomposed into two sinusoidal components, daily and annual. An auto-associative neural network is employed for separating the eﬀect of damage in extracted features from responses caused by environmental variations of the system [35]. The authors admit that several issues are to be addressed before the approach can be used on real structures.

2.4.2 Faulty sensor

In any deployed SHM system the possibility that a sensor has developed a fault and that the received signal is incorrect must be considered, in the work on the SHB [27] one of the sensors was faulty, which was detected as damage. Damaged sensors can be detected via sensor data reconstruction. In this approach sensor data is reconstructed based on spatial and temporal correlations among the sensor network. If there are discrepancies between the measurement data and reconstructed data then the sensor may be faulty. Spatial correlations are used to reconstruct sensor data via PCA [36], minimum mean square error estimation [37], and support vector regression [38]. A recurrent neural network (RNN) was used in that includes both spatial and temporal correlations [39]. This method outperformed a number of existing methods on their test set, however the test data was from numerical simulation of a not validated model.

2.4.3 Deployments

A number of bridges which were lucky enough to receive an installation of sensors have already been visited in this literature review. In the majority of these cases the bridge being researched was scheduled to be demolished [12, 17, 18], which enabled diﬀerent damages to be applied to the bridge in the period prior to demolition. The SHB however had sensors installed on lane 7 in 2014 and they were still in use in 2016 [27]. In this Section we will visit some of the operational SHM installations on bridges around the world.

(28)

A number of SHM systems were deployed to monitor distinct behaviours on bridges in Kentucky. From 2005 - 2011 a system was deployed to monitor impact damage from over-height vehicles to the eastbound I-64 bridge and to identify the vehicles [40]. Vehicles were recorded by ultrasonic height sensors and video cameras, accelerometers and strain gauges compared impact to responses from sensors on a second bridge. One of the drawbacks that led to the system being decommissioned was the cost of replacement of the data acquisition system (once due to vandalism and once due to lightning strike) and time the system was down. Two SHM systems were deployed in Kentucky on bridges over the Ohio river, a barge impact detection system on the US 41 bridge and a crack growth monitoring system on the I-275 bridge [41]. The barge impact detection system (2006 - 2015) was poorly configured. Threshold limits were set based on [42] and did not detect an impact that should have been detected. The impact caused a response less than 0.25 times the set threshold – a lower threshold was required. Data transfer was also sporadic due to a change in communication protocol by the cellular provider. The crack detection system consisted of a vibrating wire micro crack meter transmitting data to a server by cell modem every 10 minutes and data is viewable via a website. This was a successful low cost installation which began in 2012, and was still operational in 2018, and through it additional costly repairs were deemed unnecessary.

A number of SHM systems have been installed on bridges in Sweden. According to [43] many communities are responsible for the maintenance of their own bridges, but new constructions are dependent on political decisions that delay projects, thus a SHM system is deployed to provide confidence in the health of an old structure and the safety of users. The 9 x 78 m span New Arsta Railway Bridge highlights a number of practical issues relating to the SHM system installed during construction in 2003 [44]. Data collection was interrupted due to interruption in power delivery and internet connectivity, water damage to the data logger due to freezing of a drainage hole, and damage to sensors including “violent treatment after the installation like hitting the sensors with heavy re-bars”. In [45] a distributed fibre optic based SHM system was installed on the 950 m steel-beam concrete-deck Gotaalv bridge between Gothenburg and Hisingen. In an on-site crack test, the SHM system detected 4 of 7 cracks. The fibre optic is bonded to the structure with gluing to transfer the strain distribution from the structure to the fibre optic. The system is designed to operate for 15 years however it’s eﬀectiveness over that period of time remains to be proven.

A long-term study was undertaken on predicting responses of the North Halawu Valley viaduct on the Hawaiian island of Oahu [46]. The viaduct was instrumented extensively during construction in 1994 and data collected over a 9 year period. Short-term deflections and strains were successfully predicted using a three-dimensional linear elastic beam model, long-term deflections and strains also compared well after some adjustments to the original model. A method is proposed for the prediction of upper and lower bounds for the long-term response of long-span prestressed concrete bridges.

(29)

2 LITERATURE REVIEW 2.5 Concluding Remarks

2.5 Concluding Remarks

The trend in damage detection is to employ machine learning, with particular use of unsupervised methods such as the OCSVM because they do not require having a model available. Feature extraction is arguably the most important and diﬃcult step in ML-based health monitoring [22]. Two works that successfully detected apriori known damage on the SHB combined machine learning techniques with knowledge about the behaviour of the structure, in these works data from multiple sensors was compared.

Most of the existing research suggests promising results but in a simulated or laboratory set-ting, and does not consider the difficulty that environmental or operational effects provide: that anomalous conditions may be the result of these effects and not the result of damage. In many cases approaches still need to be validated for operational civil structures, because data corresponding to real structural faults is difficult to acquire.

Another criticism is that most of the research investigated focuses on the discussion of perfor-mance in terms of the accuracy metric (accuracy is defined as the number of true positive and true negative classifications divided by the total number of classifications). However a discussion of the false positive rate and false negative rate are crucial. Consider a paper that presents a false positive rate of 1%, does this mean that an anomaly detection system would incorrectly signal that damage has occurred for every 1 in 100 samples taken? Furthermore, when environmental eﬀects are taken into account the false positive rate would be expected to rise sharply.

One problem related to damage detection of reinforced concrete structures is as follows: in concrete structures with steel rebars, once the steel has corroded and failed the concrete bridge is liable to collapse, however the stiﬀness of the bridge is mostly contributed by the concrete, the corrosion of the steel has little influence on the dynamics, until the reinforcing steel bars and bridge have failed [47].

The biggest problem with damage detection of bridges concerns the acquisition of data corre-sponding to the damage state of a bridge. The following “flow chart” in bullet-point format below is intended to illustrate this problem. This “flow-chart” is a re-structuring of a few paragraphs from the conclusion of [22].

(30)

2 LITERATURE REVIEW 2.5 Concluding Remarks

• Requirement: data from damage states to find damage-sensitive features within. – Avoid this requirement, collect data only in the healthy state.

∗ Results in a large normal condition within the dataset (due to environmental factors) which is insensitive to certain types of damage.

– Accept this requirement, then where do we get the data from damage states? ∗ From real structures?

· Expensive and we cannot damage bridges while they are in operation. ∗ From a model?

· Data is still required for validation.

In the following table some of the papers in this literature review are classified based on three criteria. The criteria are whether the data includes temperature eﬀects, the type of traﬃc loading, and whether the data comes from non-linear analysis.

Table 3: Classification of some of the literature in this literature review based on three criteria: whether the data includes temperature load, traﬃc load, whether the data comes from a non-linear analysis (or potentially even real data). Duplicate papers related to the same dataset are not included here.

Papers Temperature eﬀect Traﬃc Linear/Non-linear/Real Damage

[12, 17] No No Real Yes

[16] Heated to +30 °C 2-axle vehicle Linear Yes

[24] No Train Non-linear Yes

[25] Real temperature Real traﬃc Real Yes [18] Real temperature No Non-linear and real Yes [46] Real temperature Real traﬃc Real No

Based on this literature review a recommendation is made that future research should focus on the central problem related to damage detection of concrete slab bridges, namely the problem of data collection. It is not permitted to damage a real bridge and also have real traﬃc driving over it. Yet research is only as good as the data it is built on, therefore research should attempt to be based on more realistic data (meaning more realistic models that generate the data). In the long-term this means a “digital twin” (a digital replica of a physical asset) of a bridge and its environment should be made available to researchers, to alleviate the problem of data collection.

Furthermore, toward the goal of more realistic models it is important to include environmental effects as they can significantly affect the false positive rate. Good model-based temperature effect is notably missing in Table 3. Finally, we suggest that any work should include a discussion of how a damage detection system can work in a real installation, including a discussion of false positive rate.

(31)

3 METHODS

3 Methods

This Section provides some general and background information on the methods by which work in this thesis was accomplished. An overview of the FE method is provided including the FE software used in this thesis and how simulations were run on high performance computing infrastructure, and then the importance of creating reusable models for research is argued. After this Section, in Section 4, a detailed approach of this thesis work is presented.

3.1 Finite Element Method

The finite element method is a numerical technique that can be applied for solving diﬀerential equations: structural, thermal, electrical and more. The FE method is suitable for structures where a simple analytical formula representing a geometry, material models or boundary conditions is unavailable. Geometries are decomposed (meshed) into discrete regions (finite elements) and material properties are assigned to the elements. Element-level equations are assembled and solved to determine the response to a set of boundary conditions and applied forces. A brief introduction to the FE method is now given.

Consider a simple spring system (representing a single 1D (line) finite element) with two nodes (i and j) and forces acting on these nodes as in Figure 3, the stiffness of the spring is denoted by k. The spring stiffness equation dF = kdu relates nodal displacements to the forces via the spring’s stiffness. The force equilibrium equations can be written in terms of two degrees of freedom (ui and uj) as in Equation 1 or in matrix form as in Equation 2.

Figure 3: A simple spring system, representing a single finite beam element. This image is from [48]. kui kuj = Fi kui+ kuj = Fj (1) 2 4 k k k k 3 5 2 4ui uj 3 5 = 2 4Fi Fj 3 5 ₍₂₎

(32)

Com-3 METHODS 3.1 Finite Element Method

patibility equations ensure displacements at shared nodes of adjacent elements are equal, equilibrium equations ensuring that elemental forces and external forces are in equilibrium, and constitutive equations that govern the relationship of materials ( = E"). For a system consisting of multiple finite elements the global system of equations for displacements is given as [K][u] = [F ]. An exam-ple of applying a boundary condition is setting one of ui = 0, indicating that no displacement can occur for node i.

3.1.1 Computational Time

The computational time is approximately proportional to the degrees of freedom. The models in this thesis consist of ⇡ 24, 000 nodes, each with 6 degrees of freedom (DOF), for a total of ⇡ 100, 000 DOF. The run-time of a simulation, including time to build a FE model file and parse the results is approximately 30 minutes. In this work all of the simulations are run under linear assumption, such that results from multiple simulations can be composed together, as explained in more detail later in Section 4.7.

1227 linear simulations are run for the bridge in either healthy or cracked condition, just once, and the results can be combined together relatively quickly to create a large amount of time series of sensor responses. If non-linear time-dependent analyses had been performed to generate time series of responses, then if ⇡ 100⇥24 minute time series were generated at 100 Hz over the course of development then a very crude estimation of the total run-time would be ⇡ 100⇥60⇥24⇥100⇥30 = 432, 000, 000minutes or 7,200,000 hours or ⇡ 821 years. However some of these time series can be run in parallel, but not all, since during development changes will need to be made. It is very diﬃcult to predict the amount of time that would be taken to perform such non-linear time-dependent analyses, a lot of factors are involved. Even though this is an incredibly crude estimation it is still evident that it is significantly expensive (primarily because of the loss of parallelism). In contrast the 1227 simulations of the bridge in healthy condition (under linear assumption) can be run in parallel, in practice they are run with ⇡ 200 parallelism which takes about 3 hours to run.

3.1.2 The Lisa Cluster

This work was carried out on the Dutch national e-infrastructure with the support of SURF Coop-erative. In particular simulations were run on the SURFsara Lisa cluster [49] which oﬀers a large amount of multi-core nodes for use. The bridge-sim software was run on the Lisa cluster via containerisation, by first creating a Docker image which is then converted to the Singularity format that is supported by Lisa.

Regarding the 1227 point-load simulations that are run for the bridge in the healthy condition, first a shell script is used that schedules 60 jobs on the Lisa cluster, one job per node. In each job the bridge-sim Singularity container is copied to the node’s local storage and then 20 (of the 1200) point-load simulations are run with 7x parallelism. The amount of nodes that are running

(33)

3 METHODS 3.2 Software for Research

simulations at a time depends on the availability of resources on the cluster, 30 nodes running at once would result in 30 ⇥ 7 = 210x parallelism. Examples of scripts used to run simulations on the Lisa cluster are given in Listings 5 - 6 in Appendix B.

3.1.3 OpenSees

The FE program chosen is OpenSees [50]. The primary reason for this is that OpenSees is open source software. OpenSees stands for the Open System for Earthquake Engineering Simulation, it is “an open source software framework for creating applications for the nonlinear analysis of structural and soil systems using either a standard FE model or an FE reliability analysis. It is object-oriented by design and—in addition to achieving computationally eﬃciency—it’s designed to be flexible, extensible, and portable” [51].

FE simulation commands are written in the TCL programming language into a .tcl file which is then read and executed by OpenSees. OpenSees cannot generate a mesh or plot any results, however it was chosen because it is free software. Depending on proprietary software limits access to research to those who can aﬀord it, similar to submitting research to paywalled journals and should not be encouraged. OpenSees is free and open source and supported on multiple operating systems.

3.2 Software for Research

In recent years there has been an increased emphasis on re-producible research within the sci-entific community. Re-producing research allows for the verification of another’s work, however re-producing research may come at the cost of one or more researcher’s time. In order to pre-vent the waste of time that comes with re-producing another’s work, the “dynamic document” has gained much popularity, a single document that contains both code and results, an example being the Jupyter Notebook.

However I argue that we can do even better. There may be components of another’s research that you are eager to use but that are diﬃcult to extract. To avoid this, the software which forms the research should be presented as composable functions, allowing for the re-use of the research, allowing a researcher to compose some of the functions in a manner which was not done in the original research. Jupyter Notebooks don’t compose well.

Software, including research built on software, can simply be thought of as Lego® _blocks. Research that oﬀers new and exciting Lego® _{blocks will allow for faster iteration in that research} domain by others. In order to prevent a duplication of eﬀort regarding generation of sensor data, all of the model components in this thesis are provided via a software package bridge-sim for the Python programming language. A paper has also been written about this package and is being reviewed by the Journal of Open Source Software which is a peer-reviewed open-access journal that publishes articles about research software.

(34)

4 APPROACH

4 Approach

This Section describes the approach to building a model that is followed in this thesis for the purpose of data collection. Following a brief overview of the modeling approach, a description of the generated FE model of bridge 705 is provided in Section 4.2 including listing the modeling assumption that are made. Models of different environmental and operational effects are outlined in Section 4.3. The validation of the FE model in healthy state and verification of different loading is provided in Section 4.5. Models of two damage states are detailed in Section 4.4. In Section 4.6 sensor noise is briefly discussed. In Section 4.7 the process of combining responses from multiple FE simulations to generate sensor data is described, and in Section 4.8 some selected parameters are described.

4.1 Overview

The mechanical behaviour of bridge 705 is modeled and in addition the major components that a bridge will be aﬀected by are modeled: live load (traﬃc), pier settlement, temperature, creep, shrinkage, and also crack zones, areas of the bridge deck where cracks occur. All of these model components together from a model that can be considered a “digital twin”, a digital replica of a physical asset, namely a concrete slab bridge in this case.

As will be discussed in Section 4.2.1 an assumption of linearity is made such that responses from many “unit” FE simulations can be combined to generate time series of sensor responses, (Section 4.7). These “unit” simulations include point-load simulations placed along where vehicles will drive, and 1 °C temperature and 1 mm pier settlement simulations. Vehicles are sampled from a distribution (Section 4.3.2) and driven along the bridge on a traﬃc lane in discrete time steps.

All of these unit simulations are run on a high performance computing cluster (Section 3.1.2) with high parallelism. Once the unit simulations have run, time series of responses can be generated of the bridge 705 over a longer operational period such as one year where seasonal temperature eﬀect, creep and shrinkage come into play, or for a shorter period where a crack zone is introduced.

4.2 Finite Element Model

This Section outlines the FE model of bridge 705 which was generated for simulation with the OpenSees FE software package. The geometry, element types, material models and boundary con-ditions of the model of bridge 705 are outlined. Finally the parameterized computer code model bridge-sim which allows for generating FE models for OpenSees is described.

4.2.1 Model Assumptions

The model components in this thesis are an idealization of reality. Furthermore, in order to simplify modeling a number of assumptions are made during the modeling process, they are listed here:

Towards an Autonomous Damage Detection of Concrete Slab Bridges

University of Amsterdam

Masters Thesis