How to effectively increase performance of condition-based maintenance: an investment choice between time to response and level of measuring accuracy

(1)

How to effectively increase performance of condition-based

maintenance: an investment choice between time to

response and level of measuring accuracy

University of Groningen

Faculty of Economics and Business

Master thesis: Technology of Operations Management

Author: Karsten Hoekzema (s3468453)

Supervisor: dr. Onur Kilic

Co-assessor: dr. Jasper Veldman

(2)

2 Abstract

(3)

3

1. Introduction

Renewable energy production has to grow, and fast. The European Union (EU) is legally bound to get at least 27% of its energy consumption from renewables by 2030 (Arantegui & Jäger-waldau, 2018) and has already reached a share of 17.52% in 2017 (COM, 2019). Wind energy has a central role in the energy transition as the wind itself is sufficient to meet the current global energy consumption (Adaramola et al., 2011). Wind energy can be produced on- and offshore. Stronger offshore winds, economies of scale and limited visual and noise pollution are benefits of offshore wind production. Due to these benefits, offshore wind is expanding rapidly (Arantegui & Jäger-Waldau, 2018). It is, however, more difficult to maintain offshore wind installations because of harsh offshore environment. Next to that, offshore maintenance is more costly as compared to onshore maintenance due to high logistics costs. As a result, operations and maintenance (O&M) account for 20-35% of the levelized cost of offshore wind energy (Schafiee, 2015). This figure is expected to increase even further as offshore wind farms are being built further into sea due to space limitations (De Regt, 2012).

Due to the importance of O&M in the offshore wind industry, using the right maintenance policy is essential. Maintenance policies can be divided into two main categories: corrective maintenance (CM) and preventive maintenance (PM). Based on the data used for the decision of when to perform maintenance, a further distinction in PM can be made between time-based and condition-based maintenance (CBM). Time-based maintenance uses time data (e.g. running hours, monthly maintenance etc.) to determine when to perform maintenance. Due to varying factors (e.g. operators, environmental conditions etc.) time-based maintenance often results in suboptimal maintenance, hence waste of useful component/system lifetime or failures. CBM is more complex, and combines data-driven reliability models and observation technology to approach optimal (just in time) maintenance performance (Alaswad & Xiang, 2017). Due to technological innovations in monitoring, storing and analysing conditions, the relative benefits of CBM over TBM are becoming more significant (De Jonge et al., 2017).

(5)

5

et al., 2016; Compare et al., 2015; De Jonge et al., 2017). Both response time and uncertainty in measuring the deterioration process can significantly influence the performance of CBM (De Jonge et al., 2017). In contrary to the other factors, both response time and uncertainty in measuring the deterioration process can be influenced by technological innovation, and will therefore be investigated in this thesis.

The response time is the time it takes from maintenance initiation until the actual performance of maintenance. It is a function of distance to the object and the speed with which the distance can be travelled. Moreover, response time is subject to uncertainty, especially in offshore maintenance due to weather conditions (e.g. Rademakers et al., 2003). If the response time is relatively high, the maintenance threshold needs to be set higher to guarantee a give level of reliability. Response time can be reduced by improving performance of maintenance vessels (De Regt, 2012) or implementing an offshore base of operation (Ambrose, 2019).

The level of uncertainty in measuring the deterioration process (also measuring accuracy) is based on the actual level of deterioration and the measured level of deterioration. The difference between an actual state and a measured state is also known as measurement error. In case of a low level of accuracy in measuring the deterioration process there will be a high level of uncertainty regarding the actual level of deterioration, hence the maintenance threshold needs to be set higher to guarantee a given level of reliability. Uncertainty in measuring the deterioration process can be reduced by advanced condition monitoring technologies (Tchakoua et al., 2014).

As discussed, reducing O&M costs in the offshore wind industry is critical in the energy transition. Response time and level of accuracy are factors that influence the maintenance threshold, and therefore the O&M costs, and can be improved through investments (offshore base of operation, capabilities of the maintenance vessel and sensor technology). By analysing the effects of different investment options through a numerical investigation, this thesis will enable the offshore wind industry to make effective investment choices. Next to that, this thesis will extend the work of De Jonge et al. (2017) by incorporating uncertainty into the influence of response time on CBM performance. Lastly, this thesis contributes to literature by analysing the effects of measurement noise in a CBM context for offshore wind turbines.

(6)

6

individually and combined) on the performance of CMB. Additionally, in Section 3 the model and scenario parameters are provided. Next, the results of the numerical study are presented. In Section 5, the key findings, implications, limitations and further research possibilities are discussed. Lastly, in Section 6 we provide concluding remarks.

2. Theoretical background

We start this section with background information on CBM to motivate the timing of the research and position the research question in current literature about CBM models. Next, in Section 2.2, we discuss the influence of both maintenance response time and measuring accuracy on the performance of CBM in the context offshore wind.

2.1 Condition-based maintenance

Condition monitoring can be defined as “a technique or a process of monitoring the operating characteristics of machine in such a way that changes and trends of the monitored characteristics can be used to predict the need for maintenance before serious deterioration or breakdown occurs, and/or to estimate the machine’s health.” (Han & Song, 2003).

CBM was introduced in the late 1940s by the Rio Grande Railway Steel Company (Noman et al., 2019). However, it was not until around the year 2000 that CBM ‘caught’ the attention of scholars, as can be seen in Figure 2.1. To perform CBM it is essential to be able to measure the deterioration process of certain components/systems. It is therefore logical that the trend of increasing publications regarding CBM is in line with the technological innovations which are facilitators of CBM (Schmidt & Wang, 2018).

Figure 2.1. Number of publications per year in CBM (Noman et al., 2019). Blue denotes the amount of published articles and red indicates the ratio (NCBMP/TNP)

(7)

7

with continuous monitoring, therefore only continuous monitoring will be considered (Alaswad & Xiang, 2017). Furthermore, the simplest form of deterioration is a deterministic linear process. In this case, the maintenance threshold can be set just under the failure threshold. However, in most cases the deterioration process is stochastic due to the influences of a variety of factors. For example, weather conditions (e.g. temperature and humidity) or personnel (e.g. operator) can influence the deterioration process. Saassouh et al. (2007) describe the deterioration process as an infinite number of small shocks. The characteristics of these shocks influence the maintenance threshold. The deterioration process is often modelled using different types of stochastic processes. The best fit, between stochastic process and the actual deterioration process, depends on a number of factors. Discrete-state deterioration is often modelled by Markov processes, and continuous-state deterioration processes usually as Wiener, Gamma or Inverse Gaussian processes (Alaswad & Xiang, 2017). The Wiener process can be used to model deterioration which in- and decreases over time. For instance, Elwany et al. (2011) use the Wiener process to model vibration-based deterioration of bearings. The Gamma process can be used to model deterioration which only increases or decreases over time. For instance, Do et al. (2015) use the gamma process to model deterioration with the example of the evolution of crack length. The Inverse Gaussian process was only recently introduced in degradation modelling by Wang & Xu (2010). They find that the Invers Gaussian process can be used to model deterioration which only increases or decreases over time.

2.2 Offshore maintenance

2.2.1 Response time

(8)

8

decrease in response time and total maintenance costs. These results can be explained by two reasons. Firstly, the shorter sailing distance decreases downtime in case of a failure. Secondly, the shorter sailing distance also decreases the needed good weather window, which also results in less downtime in case of a failure. It should be noted that De Regt (2012) and Ambrose (2019) consider a CM policy. Therefore, by considering a CBM policy, which will decrease the amount of failures, it is likely that the effect of an offshore base of operation will be smaller. Besnard et al. (2012) investigate the effect of different means of transportation for the maintenance crew, such as helicopters and vessels with varying capabilities/characteristics. They compare two maintenance vessels with different values for the maximum wave height in which they can still operate. Based on their results it can be derived that increasing the maximum wave height in which a maintenance vessel can still operate reduces the maintenance costs. This result can be explained by the fact that increasing the capability to operate in higher waves also increases the good weather windows. Therefore, in case of a breakdown, the chance will be lower that the vessel has to wait for a good weather window, which reduces downtime costs. Besnard et al. (2012) consider PM to reduce CM, however they still assume that approximately 50 percent of the maintenance actions are corrective. Therefore, considering a CBM policy, which is likely to decrease the CM percentage, it is likely that the effect of increased capabilities to operate in higher waves will be smaller.

(9)

9

found by balancing personnel costs versus extra induced revenue losses due to unavailability of personnel.

Weather conditions can cause a significant uncertainty in the maintenance response time of offshore wind turbines. Offshore weather conditions, such as wave hight and wind speed, can limit the time in which maintenance can be performed due to limited equipment capabilities. The time in which the weather conditions are within the operational capabilities of the maintenance equipment are called (good) weather windows. The time needed for a maintenance action (including sailing both ways) needs to be less than the time of a certain weather window to be able to perform the maintenance action. If the weather window is smaller than time needed to perform maintenance, the operation has to be postponed resulting in a delay.

2.2.2 Measuring accuracy

As discussed before, measuring accuracy (or the level of uncertainty in measuring) is based on the difference between the actual level and the measured level of deterioration. The difference between the measured and observed level of deterioration is called measurement noise/error. Measurement errors in condition monitoring can occur due to instrument measurement margins (Goto et al., 2007). For instance, Liu et al. (2019) find increasing measurement noise over time in the observation of a deterioration process due to sensor degradation. A variety of sensors (vibration, acoustic, thermal imaging) can be applied for the continuous condition monitoring in the offshore wind industry (Ciang et al., 2008; Dutton, 2004; Iliopoulos et al., 2015).

Technological innovation has improved sensor accuracy in many applications (Dinh et al., 2016; Rodbard, 2017; Tosi, 2017). However, Kamei and Takai (2010) find that small measurement errors can result in premature preventive maintenance or breakdowns as a result of performing preventive maintenance too late.

(10)

10

true positive, false negative and true negative, for a wind turbine component. Due to the generalized expressions, this model is not limited to certain stochastic processes and type of measurement noise. In contradiction to the previous discussed studies, this model can therefore be used for various deteriorating systems.

To conclude, response time equals the summation of time to travel, delay due to unavailability of resources and delay due to operational weather windows. Response time causes maintenance initiation at a lower deteriorating level, and therefore reducing CBM performance, due to the fact that condition information cannot be utilized during the response time (De Jong et al., 2017). By incorporating uncertainty, through time windows for the offshore maintenance, the influence of response time on CBM performance is expected to be even greater. Next to that, measurement noise negatively influences CBM performance through loss of value of condition information (De Jonge, 2017). By comparing different investment options (offshore base of operations, improved vessel capabilities and improved sensor technology) in a CBM context, future investment can be guided towards the most effective way of reducing O&M costs.

3. Approach

In this thesis, we address different possible improvements of a CBM operation, in the offshore wind industry context, from an investment perspective. The impact of the different investment options on the CBM operation performance are analysed through a scenario analyses using simulation.

3.1 Research method

(11)

11

Scenario planning enables organizations to compare possible future events, and therefore makes them more innovative (Amer et al., 2013). Due to the goal of this thesis we therefore take a scenario analysis approach in conjunction with the described simulation study. By comparing the base case scenario with scenarios with possible future investments, causal relationships between the control variables and performance indicators can be identified.

3.2 Simulation model

We consider maintenance operations of an offshore wind turbine that require a service operation vessel (SOV). Following several other papers in the CBM domain, we assume the offshore wind turbine to be a single-unit system which can only be maintained in its entirety (Alaswad & Xiang, 2017). Furthermore, we consider a CBM policy where the deterioration of the wind turbine is continuously monitored. Corrective maintenance is initiated when the wind turbine fails. Preventive maintenance is planned if the level of deterioration exceeds a certain threshold level.

3.2.1 Deterioration process & failure modelling

A number of stochastic processes are used to model the deterioration process in the context of condition based maintenance (Alaswad, 2017). In this thesis we use a stationary gamma process to model the deterioration process. According to Van Noortwijk (2009) the gamma process is an appropriate stochastic process to model a variety of monotonic and gradual deterioration processes. A recent example where the gamma processes was successfully applied to model deterioration in the offshore wind domain can be found in Zhang & Tee (2019). Furthermore, Van Noortwijk (2009) lists some studies, where the gamma process fits well to data of for example fatigue crack growth and corrosion of steel, which are also common for offshore wind turbines.

The density function f with the respective shape (α) and scale (β) parameters of the gamma distribution can be found in Equation 1. Following the example of Van Noortwijk (2019) we use gamma increment sampling to simulate the gamma process. The deterioration process is modelled over time, with step sizes denoted as ∆t, with random increments from the gamma distribution.

𝑓_{; ,} =

( ) (1)

(12)

12

for the wind turbine to fail. The standard deviation of the gamma process, given by σ = √𝛼𝛽, is a measure of variability. For this thesis we use the coefficient of variation (CV), given by CV = , as a measure of variability since it is considered a more generalizable measure.

3.2.2 Response time modelling

The response time is modelled based on the time it takes to get the SOV in the harbour/to the offshore base of operation, the time it takes to prepare the SOV for the operation and the time it takes to sail to the wind turbine.

Due to the relatively low number of annual failures of an offshore wind turbine, and the high costs of SOV’s, it is likely a SOV serves multiple wind farms (e.g. in IRO, 2019). This introduces a certain delay since the SOV potentially first has to finish other jobs and sail to the base of operation before maintenance can be initiated. After the vessel has arrived in the base of operation a delay is introduced for the mobilisation of the vessel. During the mobilisation of the vessel, spare parts are loaded on to the vessel and safety checks are performed.

After the SOV is mobilised, a weather window check has to be performed. Sperstad et al. (2014) found that a single-parameter analysis, based on the significant wave height (𝐻 ), give similar results as multi-parameter analysis for maintenance models such as developed in this thesis. Therefore, the weather window check will be modelled based solely on the significant wave height. If during the planned maintenance operation time, given by sailing distance (both ways) divided by the speed of the SOV and the repair time, the significant wave height does not exceed the maximum significant wave height of the SOV, maintenance can be initiated. If the maximum significant wave height is exceeded during the planned operation time, the maintenance initiation is delayed until this is no longer the case. There are multiple methods to model weather conditions. For instance, Hagen et al. (2013) use Markov chain model to produce statistics for weather windows. For an extensive overview of stochastic models used for simulating weather conditions we refer to Monbet et al. (2007). However, for this thesis we will follow the example of Uit het Broek et al. (2019), and model the weather condition using actual weather data. We use 10 years (2010-2019) of weather data obtained from the K13 Alpha offshore platform (RWS, 2020). In the simulation, whole years of data are randomly drawn from the data set to preserve seasonality effects.

The total time to response is given by Equation 2.

(13)

13

3.2.3 Repair modelling

After the response time is completed, maintenance will be performed. The type of maintenance, either preventive or corrective, depends on the level of deterioration of the system at the time maintenance is performed. If the level of deterioration exceeds the failure level, corrective maintenance is performed and otherwise preventive maintenance is performed. Both a preventive and corrective maintenance action will bring the system to a good-as-new state assuming perfect maintenance. Furthermore, to simplify the simulation model the repair time is assumed to be negligible in the deterioration process. Therefore, if either preventive or corrective maintenance is performed, the level of deterioration immediately goes back to 0. It should be noted that the repair time is taken into account for the weather window, and the calculation of the wind turbine availability.

3.2.4 Measuring accuracy modelling

As discussed in Section 2.2.2, measurement noise can be modelled using various methods. Following the example of Wu & Law (2004), we assume random measurement noise to be uniformly distributed with a mean of 0 and a variance depending on the sensor measuring accuracy. We therefore consider the observed level of deterioration at time t, denoted as 𝑋 ( ), as the actual level of deterioration at time t, denoted as 𝑋 ( ), multiplied by the measurement noise at time t. Given that e denotes the maximum percentage error, the observed level of deterioration is given by Equation 3.

𝑋 ( )= 𝑋 ( ) ∗ 𝑈( , ) (3)

3.2.5 Cost modelling

In the simulation model two different costs are considered, namely costs associated with (corrective/preventive) maintenance and downtime.

(14)

14

denoted as 𝑣 , is the minimum wind speed needed for a wind turbine to be able to produce electricity. The maximum production speed, denoted as 𝑣 , is the wind speed at which the wind turbine is producing maximum electricity. The cut-out speed, denoted as 𝑣 , is the maximum wind speed at which the wind turbine can produce electricity. In case the maximum wind speed is exceeded, the wind turbine will be shut down to as a precaution measure. The electricity produced for a certain wind speed, denoted as 𝑃(𝑣), is calculated with Equation 4, in which 𝑢 denotes the nameplate capacity of the wind turbine.

𝑃(𝑣) = ⎩ ⎪ ⎨ ⎪ ⎧ 0 ∗ 𝑢 𝑓𝑜𝑟 𝑣 ≤ 𝑣 ( cos ( ( ))𝜋 + 1) ∗ 𝑢 𝑓𝑜𝑟 𝑣 ≤ 𝑣 ≤ 𝑣 1 ∗ 𝑢 𝑓𝑜𝑟 𝑣 ≤ 𝑣 ≤ 𝑣 0 ∗ 𝑢 𝑓𝑜𝑟 𝑣 ≥ 𝑣 (4)

The total costs can therefore be found with Equation 5.

𝑇𝑜𝑡𝑎𝑙 𝑐𝑜𝑠𝑡𝑠 = ∑ 𝑃𝑀 𝑐𝑜𝑠𝑡𝑠 + ∑ 𝐶𝑀 𝑐𝑜𝑠𝑡𝑠 + ∑ 𝐷𝑜𝑤𝑛𝑡𝑖𝑚𝑒 𝑐𝑜𝑠𝑡𝑠 (5)

3.3 Input parameters

In this section the input parameters of the simulation model are presented and motivated. An overview of all model parameters can be found in Table 3.1.

Table 3.1. Main model parameters including values for the sensitivity analysis

Value Parameter Sensitivity

analyses Sensitivity analyses All scenarios Sensitivity analyses Sensitivity analyses

Failure level 1000

Average deterioration ~0.57

# failures per year 2.5 5 7.5

Coefficient of variation 0.0 0.5 1.0 2.0 5.0

Mean planning time SOV 42 h 168 h 672 h

SD planning time SOV 12 h 48 h 192 h

Mobilisation time SOV 0 h 12 h 24 h

Repair time 10 17.5 h 25

CM costs €82,500

PM costs €20.625 €41,250 €61.875

Electricity price €50/MWh €100/MWh €150/MWh

Cut-in wind speed 3 m/s

Max production wind

speed 14 m/s

Cut-out wind speed 25 m/s

Wind turbine nameplate

capacity 2.5 MW 5 MW 7.5 MW

Step size ∆t 1 h

Simulation length 75 years

(15)

15

The value for the failure level is arbitrarily set to be 1000 for the whole duration of the simulation, hence assuming a fixed failure level.

According to Carroll et al. (2016) an offshore wind turbine has approximately 5 breakdowns per year which would require a SOV. Given that a year has 8766 hours on average, the average deterioration increment is set to be ~0.57. Different components of a wind turbine will have different deterioration paths with respective variations. Since no studies report on coefficients of variation for the gamma process which represent values found in practise for the offshore wind industry, the CV is arbitrarily set to be 1 and will be subject to a sensitivity analysis with ranging values from 0 to 5. With a failure level of 1000 and 5 failures per year on average, the yearly average cumulative deterioration is 5000. Since the CV reflects on the variation of a single deterioration increment, it should be noted that the annual variation will be significantly lower. Through the analysis of 100 deterioration sample paths with an CV set to 1 and 5, the variations in failures per year will between 4.89 – 5.11 and 4.36 – 4.91 respectively. See Appendix 1 for the distributions and sample paths.

The delay associated with SOV availability is arbitrarily set to be normally distributed with a mean of 168 hours and standard deviation of 48 hours. In case of a high SOV availability, due to for instance a low number of offshore wind turbines serviced by the SOV, the delay will decrease. In case of a low SOV availability, due to a chartering contract or high number of offshore wind turbines served by the SOV, the delay will increase. Since the value for the delay strongly depends on the amount of wind turbines served by the SOV, and the respective maintenance contract, it will be subjected to a sensitivity analyses.

Following the example of Uit het Broek et al. (2019), the mobilisation time is assumed to be constant. Since a SOV carries smaller spare parts and is used for more standardized operations compared to a jack up vessel, we set the mobilisation time to 12 hours instead of the 24 hours used by Uit het Broek et al. (2019).

(16)

16

wind turbine components is approximately 0.25. However, since we assume that preventive and corrective maintenance actions are both performed by a SOV this ratio is considered to be too low. We therefore arbitrarily set the PM/CM cost ratio to be 0.5, hence setting the costs of a PM action to be €41,250. However, due to the lack of consensus the PM costs, and therefore the PM/CM ratio, will be subjected to a sensitivity analysis.

For the electricity output of the wind turbine we follow the example of Uit het Broek et al. (2019). We therefore set the electricity price to be €100 per megawatt hour, and assume a cut-in wcut-ind speed of 3 m/s, a maximum production speed of 14 m/s, a cut-out speed of 25 m/s, resulting in a wind turbine power curve as presented in Figure 3.1.

Figure 3.1. Wind turbine power curve (5 MW)

Lastly, we set the simulation step size (∆t) to be 1 hour. Furthermore, to find the optimal PM threshold we set the simulation length to be 75, and the number of runs 14. This number of iterations was found to be sufficiently large to obtain stable results in the total costs as a function of the PM threshold.

3.4 Key performance indicators

The considered key performance indicators of the system include mean cost per year, wind turbine availability, percentage CM actions average level of deterioration.

(17)

17

by dividing the total costs of the maintenance operation (Equation 5) by the amount of years simulated.

𝑀𝑒𝑎𝑛 𝑡𝑜𝑡𝑎𝑙 𝑐𝑜𝑠𝑡𝑠 𝑝𝑒𝑟 𝑦𝑒𝑎𝑟 = [€]

[ ] (6)

The availability is defined as the percentage of time that the wind turbine is operational. Availability can be computed by dividing the total operational time by the total time, see Equation 7.

𝐴𝑣𝑎𝑖𝑙𝑎𝑏𝑖𝑙𝑖𝑡𝑦 = [ ]

[ ] (7)

The percentage of CM actions is an important performance indicators for offshore wind turbines (Gonzalez et al., 2017). The CM percentage is defined as the ratio of CM actions with regards to the total amount of maintenance actions. Since CM actions have a relatively high associated costs, a low percentage of CM actions is considered an indicator for a good maintenance operation.

𝐶𝑀(%) = (8)

The average level of deterioration performance indicator is an important performance indicator to evaluate the overall quality of CBM system. The average level of deterioration indicator is defined as the sum of the levels of deterioration at the time of performing maintenance over the amount of maintenance actions.

𝐴𝑣𝑒𝑟𝑎𝑔𝑒 𝑙𝑒𝑣𝑒𝑙 𝑜𝑓 𝑑𝑒𝑡𝑒𝑟𝑖𝑜𝑟𝑎𝑡𝑖𝑜𝑛 =∑ (9)

3.5 Scenarios

(18)

18

Apart from the base case, we consider 17 scenarios where each scenario is representing an investment option, or a combination of investment options. Each scenario has three scenario specific settings: base of operations, SOV capabilities and sensor accuracy. Next to the onshore base of operations considered in the base case, a base of operations close to the offshore wind turbine is considered as an investment option. By investing in an offshore base of operation, the distance between base of operations and the offshore wind turbine will be reduced to 5 kilometres. Next to the SOV considered in the base case, two other SOV’s are considered with improved capabilities. For scenarios 7-12 we consider a SOV with the capabilities to operate in waves with a significant wave height of 3.6 meters. For scenarios 13-18 we consider a SOV with a sailing speed of 30 kilometres per hour. Lastly, next to the sensor with ± 10 % measuring

accuracy considered in the base case, two different sensors with ± 5 % and ± 0 % measuring accuracy are considered. The improved measuring accuracy of those sensors are in line with possible future innovations as a result of investments in sensor technology. In Table 3.2 all scenarios with their respective model parameters are presented.

Table 3.2. Scenario model parameters

Measuring

accuracy Response time

Scenario Sensor accuracy

[%] Distance base of operation - wind turbine [km]

Service operation vessel

Sailing speed [km/h] Max wave height [m]

1 ± 10 5 25 3.0 2 ± 5 5 25 3.0 3 ± 0 5 25 3.0 4 – base case ± 10 50 25 3.0 5 ± 5 50 25 3.0 6 ± 0 50 25 3.0 7 ± 10 5 25 3.6 8 ± 5 5 25 3.6 9 ± 0 5 25 3.6 10 ± 10 50 25 3.6 11 ± 5 50 25 3.6 12 ± 0 50 25 3.6 13 ± 10 5 30 3.0 14 ± 5 5 30 3.0 15 ± 0 5 30 3.0 16 ± 10 50 30 3.0 17 ± 5 50 30 3.0 18 ± 0 50 30 3.0

4. Results

(19)

19

of measuring accuracy and response time on CBM performance are presented and analysed. Lastly, the base case results are tested in a sensitivity analysis.

4.1 Base case analysis

In Figure 4.1 the mean costs per year as a function of the PM threshold are visualized.

Figure 4.1. Mean annual total cost as a function of the PM threshold for the base case

(20)

20 Figure 4.2. Number of CM & PM actions as a function of the PM threshold for the base case.

Figure 4.3. Total mean costs divided into downtime, CM and PM costs as a function of the PM threshold for the base case.

The availability of the wind turbines follows an opposite path, with an optimal of 97%, compared to the mean costs per year in Figure 4.1. This can logically be explained by the relation between downtime, CM actions, PM action and availability. Availability and costs are both determined by the amount of downtime and the amount of CM and PM actions. Less downtime and maintenance actions results in a higher availability and lower total costs.

(21)

21

average level of deterioration increases slightly faster due to the chance of a failure. The graph stabilizes to an average level of deterioration of 1000 which is the failure threshold, and therefore the maximum level of deterioration.

Figure 4.4. Average level of deterioration at time of maintenance as a function of the PM threshold. 4.2 Scenario analyses

In Sections 4.3.1 – 4.3.4 the results of the scenario analysis are presented. With the scenario analysis the impact of different investment options, and a combination thereof, on the maintenance operation are investigated. We will mainly reflect on the financial performance indicator to investigate financial viabilities of the investment options. Since the investment options would benefit at least a whole wind farm, we consider the cost benefits of the investments for a wind farm of 150 turbines over a lifetime period of 25 years.

4.2.1 Sensor technology

(22)

22 Table 4.1. Results of scenario 4 ,5 and 6 with measurement noise as the experimental model parameter.

Scenario Measurement noise Total annual costs

Optimal PM threshold

Total costs life time Cost benefit

4 ± 10 % €243,793 880 €914,222,352 €-

5 ± 5 % €243,379 850 €912,672,578 €1,549,774

6 ± 0 % €243,124 810 €911,713,865 €2,508,487

The cost reduction as a result of the increased measuring accuracy can be explained by the fact that measurement noise makes the condition information less valuable. Since the exact level of deterioration is unknown, it becomes more difficult to find the optimal PM threshold which causes an increases in total costs.

Opposed to the cost reduction, the finding that the optimal PM threshold decreases with an increased measuring accuracy is found to be counter intuitive. We would expect the CBM model, with CM costs greater than PM costs, to hedge for the uncertainty introduced by the measurement noise by decreasing the optimal PM threshold. To be able to explain this finding we use an example of a deterministic deterioration path, with deterioration increments of 2, as shown in Figure 4.5. It can be observed that as the level of deterioration increases over time, the measurement noise also increases. This can be explained by the fact that the measurement noise is modelled as a percentage, of the actual level of deterioration.

(23)

23

If we consider the response time to be fixed at 100 hours, and the failure level to be 1000, the optimal PM threshold would be 799 indicated in Figure 4.6 with the black dashed line. Since we assume a deterministic system, PM would always be performed just before failure, resulting in an optimal cost when PM is planned at time 400 hours. However, with the introduction of measurement noise, the CBM model is no longer able to initiate PM at time 400 hours. This is due to the fact that PM is only planned if the PM threshold is exceeded. Therefore, to find a near-optimal solution, the model will increase the PM threshold as indicated with the red and green dashed lines.

Figure 4.6. Deterministic deterioration path (level of deterioration as a function of time) with 5 & 10 % measurement noise and the optimal PM thresholds.

4.2.2 Service operation vessel

(24)

24 Table 4.2. Results of scenario 4 ,10 and 16 with SOV capabilities as the experimental model parameters.

Scenario Max significant wave height [m] Sailing speed [km/h] Total annual costs Optimal PM threshold

Total costs life time

Cost benefit

4 3.0 25 €243,793 880 €914,222,352 €-

10 3.6 25 €241,808 910 €906,780,004 €7,442,348

16 3.0 30 €243,434 890 €912,878,926 €1,343,426

The cost reduction as a result of the increased SOV sailing speed can be explained by two reasons. Firstly, the SOV will be able to reach the wind turbine faster in case of a failure, resulting in less downtime costs and therefore reducing the total costs. Secondly, the increased sailing speed will reduce the needed weather window to complete a maintenance action, reducing the effect of the weather uncertainty and hence reducing total costs. Additionally, the increase in the optimal PM threshold can be explained by a combination of the reduced weather uncertainty, and the reduction in total response time. Since CM is more expensive than PM, the model hedges for uncertainty by reducing the PM threshold to prevent excessive CM actions. Therefore, reducing uncertainty will increase the optimal PM threshold. Next to that, in a deterministic system the optimal PM threshold solely depends on the total response time. Increasing the response time in a deterministic system, will logically decrease the PM threshold in a linear fashion. The effect of response time on the PM threshold of course also holds for a stochastic system, hence increasing the PM threshold for the response time reduction as a result of increased sailing speed.

The cost reduction as a result of the increased SOV capability to operate in higher waves can be explained by the increased weather windows. Opposed to the effect of a faster SOV, which decreases the needed weather window, the capability to operate in higher waves will increase the weather window. Similarly to the reduction of the needed weather window, increasing the weather window will reduce the effect of the weather uncertainty and therefore reduce costs. Furthermore, the increase in the optimal PM threshold can, similarly as with the faster SOV, be explained by the reduction of weather uncertainty.

4.2.3 Service island

(25)

25 Table 4.3. Results of scenario 1 and 4 with the distance from the base of operation to the wind turbine as the experimental model parameter.

Scenario Distance to base of operation [km]

Total annual costs

Optimal PM threshold

Total costs life time

Cost benefit

4 50 €243,793 880 €914,222,352 €-

1 5 €242,493 900 €909,349,734 €4,872,617

The cost reduction as a result of an offshore base of operation can, similarly to a faster SOV, be explained by two reasons. Firstly, since the distance from the base of operation to the wind turbine is smaller, the SOV will be able to reach the wind turbine faster in case of a failure, resulting in less downtime costs and therefore reducing the total costs. Secondly, the reduced sailing distance will reduce the needed weather window to complete a maintenance action, reducing the effect of the weather uncertainty and therefore reduce costs. Furthermore, the increase in the optimal PM threshold can, similarly as with the faster SOV, be explained by a combination of the reduced weather uncertainty, and the reduction in total response time.

4.2.4 Investment combinations

In Figure 4.7 the cost benefits of all scenarios are visualized. On the y-axis, the scenarios are identified with the respective scenario settings. The base of operation is indicated by either offshore or onshore. The measuring accuracies are indicated with sensor 1, 2 or 3 with measuring accuracies of ±10 %, ±5 % and ±0 %, respectively. Lastly, the SOV capabilities are indicated with SOV 1 as base case, SOV 2 with increased capabilities to sail through higher waves and SOV 3 with increased sailing speed capabilities.

(26)

26

Similarly to the results in Section 4.2.1, we observe an increase in cost benefit if only the sensor accuracy increases (e.g. scenario 1, 2 and 3). Furthermore, similarly to the results in Section 4.2.2 we observe an increase in cost benefit for all scenario comparisons where only SOV capabilities change from SOV 1 to SOV 2 or 3 (e.g. scenario 1 and 7 or 1 and 13). Also, similarly to the results in Section 4.2.3 we observe an increase in cost benefit for all scenario comparisons where the base of operation changes from onshore to offshore (e.g. 13 and 16). By finding the same trends as identified in Sections 4.2.1 – 4.2.3, the reliability of the results are further improved. Lastly, it can be observed that combining investments increases the cost benefits. However, it has to be noted that the cost benefits of combining an offshore base of operation with an SOV with increased sailing speed capabilities are negligible. This can logically be explained by the fact that with using an offshore base of operation, the sailing distance to a wind turbine is relatively small, and therefore increasing the sailing speed does not have a significant effect.

4.3 Sensitivity analysis

In Figure 4.8 the results of the sensitivity analyses on the base case parameters are visualised. The results shown are the total costs for the optimal PM threshold. It can be observed that both the amount of failures per year, and the PM costs have a relatively large impact on the total costs compared to the other model parameters.

Figure 4.8. Sensitivity analysis for main model parameters of the base case.

(27)

27

trend can easily be explained by the direct relationship between the number of failures per year and all cost parameters (downtime, CM and PM costs). Additionally, the large impact on the optimal PM threshold can easily be explained by the fact that de-/ increasing the amount of failures per year changes the average deterioration increment.

Table 4.4. Results of sensitivity analysis for the number of failures per year. # of failures per year Total annual costs Optimal PM threshold

2.5 €116,836 990

5.0 €243,874 890

7.5 €375,718 800

In Table 4.5 the results of the sensitivity analysis of the PM costs are shown. It can be observed that the PM cost has a lower impact on total costs than the number of failures. This can be explained by two reasons. Firstly, opposed to the number of failures, PM costs are not directly linked to the downtime costs. Secondly, due to the change in failure severity (CM costs with respect to PM costs) the model finds a new optimal CM percentage. The combination of the lack of relation with the downtime costs, and the change in optimal CM percentage causes for a non-linear relationship between PM costs and total costs.

Table 4.5. Results of the sensitivity analysis for the PM costs. PM costs Total annual

costs Optimal PM threshold Annual # CM actions Annual # PM actions CM (%) €20,625 €132,824 880 0.04 5.37 0.7 €41,250 €243,874 890 0.08 5.27 1.5 €61,875 €350,537 910 0.23 5.01 4.4

(28)

28 Table 4.6. Results of the sensitivity analysis of the coefficient of variation of the gamma (deterioration) process.

Coefficient of variation Total annual costs Optimal PM threshold

0 €241,385 900 0.5 €242,269 890 1 €243,874 890 2 €245,541 890 5 €255,737 860

5. Discussion

In this section the key findings are discussed with regards to the investment viability, and are compared to similar studies. Additionally, both the managerial and academical implications are discussed. Lastly, the limitations of this study and the future research possibilities are discussed.

5.1 Key findings

The results of the developed model highlight the impact of an offshore base of operation, sensors accuracy of measuring the deterioration process and SOV capabilities on the performance of CBM for offshore wind turbines. It was derived that implementing the proposed investment options (offshore base of operations, increased sensor accuracy and increased SOV capabilities) all have a beneficial impact on the CBM performance of offshore wind turbines. By modelling the measurement noise as a uniformly distributed percentage of the actual level of deterioration, we find that an increase in measurement noise decreases the financial performance, and increases the optimal PM threshold. The finding that financial performances decreases as measurement noise increases is in line with the findings of De Jonge et al. (2017). However, De Jonge et al. (2017), and any other paper to the best of the authors knowledge, do not report on the optimal PM threshold for varying measurement noises in the CBM domain. Compared to the study of Bernard et al. (2013) the cost benefits for increasing the capabilities of an SOV to operate in bigger waves are significantly smaller in this thesis. This can be explained by the difference in considered maximum significant wave heights. Besnard et al. (2012) consider an increase in the maximum significant wave height of 1.5 to 2.0 meters compared to the 3.0 to 3.6 meters considered in this thesis. Given that 3.0 meters already causes relatively large weather windows to perform maintenance, the potential benefit is relatively low compared to the situation with 1.5 meters.

(29)

29

per wind turbine of €171,703 compared to the €1,300 found in this study. This difference can be explained by three main reasons. Firstly, De Regt (2012) considers a corrective maintenance policy where this study considers a CBM policy. A CBM policy reduces the effect of an offshore base of operation through PM. Due to PM less failures occur resulting in less downtime, and therefore less potential for improvement. Secondly, De Regt (2012) considers a maximum significant wave height for the maintenance vessels between 1.5 and 2.0 meters compared to the 3.0 meters considered in this study. A lower value for the maximum significant wave height in which the maintenance vessel can operate causes a significant higher waiting times due to weather conditions. Since an offshore base of operation decreases the needed weather window, taking a higher value for the maximum significant wave height will reduce the effect of an offshore base of operation. Lastly, in this thesis we took a single-unit approach where De Regt (2012) took a multi-system approach. With a multi-system approach an additional potential waiting time is introduced due to the fact that a maintenance crew can only serve 1 wind turbine at the same time. Given that De Regt (2012) only modelled 1 maintenance crew for 728 wind turbines, the waiting time was found to be very high. And since an offshore base of operation reduces the sailing distance, and therefore the waiting time, downtime was significantly reduced in the study of De Regt (2012) resulting in a large cost benefit.

5.2 Implications

The developed model can be used by the offshore wind industry to determine their most effective investment option. Given the model parameters settings, as discussed in this thesis, the following implications should be considered by the offshore wind industry while considering investment options to reduce O&M costs.

Implementing increased sensor technology, and therefore reducing measuring accuracy, can result in a cost savings of €2.5 million for the total lifecycle of an offshore wind farm. Given that such technology, once available, could be implemented to many offshore wind farms without significant additional costs, this is found to be an interesting investment. Additionally, since technological innovation already has improved sensor accuracy in a variety of applications, this could be used to improve sensor accuracy in the offshore wind industry with relatively low costs.

(30)

30

the capabilities of an SOV to operate in bigger waves is found to be more interesting than investing in increased sailing speed capabilities.

Implementing an offshore base of operation can results in a cost benefit of €4.9 million for the total lifecycle of an offshore wind farm. It can therefore be concluded that building a new structure to serve as an offshore base of operation is not a viable business case. However, repurposing an exciting structure, as for example proposed by Ambrose (2019) in the form of an offshore oil and gas platform, could still be an interesting investment.

5.3 Limitations and future research

One limitation of this study is the way we modelled the measurement noise in combination with the relatively high measuring frequency and the CBM control limit strategy. Given the relatively high measuring frequency in combination with a uniformly distributed percentage of the actual level of deterioration as the measurement noise, the chance of exceeding the PM threshold while the actual level of deterioration is still lower is very high. As the control limit strategy initiates maintenance when the observed level of deterioration exceeds the PM threshold, high measurement noise will always initiate maintenance too early with a fixed PM level. However, in practise the decision maker would probably be basing its decision based on multiple measurements (e.g. using averages). Therefore, it is recommended that further research investigates the differences in modelling methods for measurement noise in a CBM context for systems with high measuring frequencies.

Another limitation of this study is the model parameter setting of the gamma process coefficient of variation. Since there was no data available of the CV which reflects a good resembles of practise in the offshore wind domain, the CV was arbitrarily set to a value of 1. Therefore, it would be interesting to extend the findings of our sensitivity analysis on the base case with regards to the impact of different coefficients of variation on the proposed investments (scenarios). Since all investment options reduce some kind of uncertainty, increasing the uncertainty in the deterioration process (which is logically linked to all other forms of uncertainty in the model) is likely to increase the effect of all investments. Additionally, it would be interesting to investigate the impact of implementing different stochastic processes for modelling the deterioration process.

(31)

31

downtime, taking a multiple-system approach is likely to increase the effect of those investments.

6. Conclusion

(32)

References

Adaramola, M. S., Paul, S. S., & Oyedepo, S. O. (2011). Assessment of electricity generation and energy cost of wind energy conversion systems in north-central Nigeria. Energy Conversion and Management, 52(12), 3363-3368.

Alaswad, S., & Xiang, Y. (2017). A review on condition-based maintenance optimization models for stochastically deteriorating system. Reliability Engineering & System Safety, 157, 54-63.

Ambrose, M. (2019). Assessing the Impact of Onsite Maintenance on Offshore Wind Farms : The Case of the Repurposed Oil Platform Master Thesis. Thesis University of

Groningen. Accessed on 10-02-2020: https://scripties.feb.eldoc.ub.rug.nl/root/MScTOM/ Amer, M., Daim, T. U., & Jetter, A. (2013). A review of scenario planning. Futures, 46,

23-40.

Arantegui, R. L., & Jäger-Waldau, A. (2018). Photovoltaics and wind status in the European Union after the Paris Agreement. Renewable and Sustainable Energy Reviews, 81, 2460-2471.

Besnard, F., Fischer, K., & Tjernberg, L. B. (2012). A model for the optimization of the maintenance support organization for offshore wind farms. IEEE Transactions on Sustainable Energy, 4(2), 443-450.

Carroll, J., McDonald, A., & McMillan, D. (2016). Failure rate, repair time and unscheduled O&M cost analysis of offshore wind turbines. Wind Energy, 19(6), 1107-1119.

Chen, N., Ye, Z. S., Xiang, Y., & Zhang, L. (2015). Condition-based maintenance using the inverse Gaussian degradation model. European Journal of Operational Research, 243(1), 190-199.

Ciang, C. C., Lee, J. R., & Bang, H. J. (2008). Structural health monitoring for a wind turbine system: a review of damage detection methods. Measurement science and

technology, 19(12), 122001.

COM. 2019. Communication from the Commission to the European Parliament, the Council, the European Economic and Social Committee and the Committee of the regions. Renewable Energy Progress Report, COM (2019) 225 final, 09 April 2019. European Commission: Brussels.

Compare, M., Martini, F., & Zio, E. (2015). Genetic algorithms for condition-based maintenance optimization under uncertainty. European journal of operational research, 244(2), 611-623.

(33)

33

De Regt, D. N. (2012). Economic feasibility of offshore service locations for maintenance of offshore wind farms on the Dutch part of the North Sea. Delft University of Technology. Dewan, A., & Asgarpour, M. (2016). Reference O & M Concepts for Near and Far Offshore

Wind Farms. Petten: ECN.

Ding, F., & Tian, Z. (2012). Opportunistic maintenance for wind farms considering multi-level imperfect maintenance thresholds. Renewable Energy, 45, 175-182.

Dinh, T. V., Choi, I. Y., Son, Y. S., & Kim, J. C. (2016). A review on non-dispersive infrared gas sensors: Improvement of sensor detection limit and interference correction. Sensors and Actuators B: Chemical, 231, 529-538.

Dinwoodie, I., Endrerud, O. E. V., Hofmann, M., Martin, R., & Sperstad, I. B. (2015). Reference cases for verification of operation and maintenance simulation models for offshore wind farms. Wind Engineering, 39(1), 1-14.

Do, P., Voisin, A., Levrat, E., & Iung, B. (2015). A proactive condition-based maintenance strategy with both perfect and imperfect maintenance actions. Reliability Engineering & System Safety, 133, 22-32.

Dutton, A. G. (2004, November). Thermoelastic stress measurement and acoustic emission monitoring in wind turbine blade testing. In European Wind Energy Conference London (pp. 22-25).

Elwany, A. H., Gebraeel, N. Z., & Maillart, L. M. (2011). Structured replacement policies for components with complex degradation processes and dedicated sensors. Operations research, 59(3), 684-695.

Foxwell, F. (2018). Service ship of the future will be greener and much more cost-effective. Offshore wind journal. https://www.rivieramm.com/opinion/opinion/service-ship-of-the-future-will-be-greener-and-much-more-cost-effective-24789

Gonzalez, E., Nanos, E. M., Seyr, H., Valldecabres, L., Yürüşen, N. Y., Smolka, U., ... & Melero, J. J. (2017). Key performance indicators for wind farm operation and maintenance. Energy Procedia, 137, 559-570.

Goto, R., & CG, N. M. T. (2007). Precision of measurement as a component of human variation. Journal of physiological anthropology, 26(2), 253-256.

Hagen, B., Simonsen, I., Hofmann, M., & Muskulus, M. (2013). A multivariate Markov weather model for O&M simulation of offshore wind parks. Energy Procedia, 35, 137-147.

Han, Y., & Song, Y. H. (2003). Condition monitoring techniques for electrical equipment-a literature survey. IEEE Transactions on Power delivery, 18(1), 4-13.

(34)

34

Iliopoulos, A., Shirzadeh, R., Weijtjens, W., Guillaume, P., Van Hemelrijck, D., & Devriendt, C. (2016). A modal decomposition and expansion approach for prediction of dynamic responses on a monopile offshore wind turbine using a limited number of vibration sensors. Mechanical Systems and Signal Processing, 68, 84-104.

IRO. (2019) Keel laying ceremony for DEME’s first dedicated service operation vessel for offshore wind farm maintenance. Accessed on 07-06-2020: https://iro.nl/news-and- press/keel-laying-ceremony-for-demes-first-dedicated-service-operation-vessel-for-

offshore-wind-farm- maintenance/?utm_source=rss&utm_medium=rss&utm_campaign=keel-laying- ceremony-for-demes-first-dedicated-service-operation-vessel-for-offshore-wind-farm-maintenance

Kamei, M., & Takai, O. (2010). Influence of sensor information accuracy on condition-based maintenance strategy for GIS/GCB maintenance. IEEE transactions on power

delivery, 26(2), 625-631.

Karlsson, C. (Ed.). (2016). Research methods for operations management. Routledge. Liu, B., Do, P., Iung, B., & Xie, M. (2019). Stochastic filtering approach for condition-based

maintenance considering sensor degradation. IEEE Transactions on Automation Science and Engineering.

Lydia, M., Kumar, S. S., Selvakumar, A. I., & Kumar, G. E. P. (2014). A comprehensive review on wind turbine power curve modeling techniques. Renewable and Sustainable Energy Reviews, 30, 452-460.

Maples, B., Saur, G., Hand, M., Van De Pietermen, R., & Obdam, T. (2013). Installation, operation, and maintenance strategies to reduce the cost of offshore wind energy (No. NREL/TP-5000-57403). National Renewable Energy Lab.(NREL), Golden, CO (United States).

Mercier, S., & Castro, I. T. (2013). On the modelling of imperfect repairs for a continuously monitored gamma wear process through age reduction. Journal of Applied

Probability, 50(4), 1057-1076.

Monbet, V., Ailliot, P., & Prevosto, M. (2007). Survey of stochastic models for wind and sea state time series. Probabilistic engineering mechanics, 22(2), 113-126.

Nicolai, R. P. (2008). Maintenance models for systems subject to measurable deterioration (No. 420). Rozenberg Publishers.

Noman, M. A., Nasr, E. S. A., Al-Shayea, A., & Kaid, H. (2019). Overview of predictive condition based maintenance research using bibliometric indicators. Journal of King Saud University-Engineering Sciences, 31(4), 355-367.

(35)

35

Poppe, J., Boute, R. N., & Lambrecht, M. R. (2018). A hybrid condition-based maintenance policy for continuously monitored components with two degradation

thresholds. European Journal of Operational Research, 268(2), 515-532.

Rademakers, L. W. M. M., Braam, H., Zaaijer, M. B., & Van Bussel, G. J. W. (2003, June). Assessment and optimisation of operation and maintenance of offshore wind turbines. In Proc. EWEC.

Raza, A., & Ulansky, V. (2019). Optimal Preventive Maintenance of Wind Turbine

Components with Imperfect Continuous Condition Monitoring. Energies, 12(19), 3801. Rodbard, D. (2017). Continuous glucose monitoring: a review of recent studies demonstrating

improved glycemic outcomes. Diabetes technology & therapeutics, 19(S3), S-25. RWS. (2020). 10 years of weather data from platform K13 Alpha. Retrieved from

https://waterinfo.rws.nl/#!/kaart/golfhoogte/.

Saassouh, B., Dieulle, L., & Grall, A. (2007). Online maintenance policy for a deteriorating system with random change of mode. Reliability Engineering & System Safety, 92(12), 1677-1685.

Schmidt, B., & Wang, L. (2018). Cloud-enhanced predictive maintenance. The International Journal of Advanced Manufacturing Technology, 99(1-4), 5-13.

Shafiee, M. (2015). Maintenance logistics organization for offshore wind energy: Current progress and future perspectives. Renewable Energy, 77, 182-193.

Sperstad, I. B., Halvorsen-Weare, E. E., Hofmann, M., Nonås, L. M., Stålhane, M., & Wu, M. (2014). A comparison of single-and multi-parameter wave criteria for accessing wind turbines in strategic maintenance and logistics models for offshore wind farms. Energy Procedia, 53, 221-230.

Smith, R. D., & Chief Scientist, M. (1999). Simulation: The engine behind the virtual world. GEN, 1, 72.

Tchakoua, P., Wamkeue, R., Ouhrouche, M., Slaoui-Hasnaoui, F., Tameghe, T. A., & Ekemb, G. (2014). Wind turbine condition monitoring: State-of-the-art review, new trends, and future challenges. Energies, 7(4), 2595-2630.

Tian, Z., Jin, T., Wu, B., & Ding, F. (2011). Condition based maintenance optimization for wind power generation systems under continuous monitoring. Renewable Energy, 36(5), 1502-1509.

Tosi, D. (2017). Review and analysis of peak tracking techniques for fiber Bragg grating sensors. Sensors, 17(10), 2368.

(36)

36

Van Horenbeek, A., Buré, J., Cattrysse, D., Pintelon, L., & Vansteenwegen, P. (2013). Joint maintenance and inventory optimization systems: A review. International Journal of Production Economics, 143(2), 499-508.

Van Noortwijk, J. M. (2009). A survey of the application of gamma processes in maintenance. Reliability Engineering & System Safety, 94(1), 2-21.

Walsh, C. (2019). Offshore Wind in Europe–key trends and statistics 2018. Wind Europe, Brussels.

Wang, X., & Xu, D. (2010). An inverse Gaussian process model for degradation data. Technometrics, 52(2), 188-197.

Wu, D., & Law, S. S. (2004). Damage localization in plate structures from uniform load surface curvature. Journal of Sound and Vibration, 276(1-2), 227-244.

(37)

37

(38)

How to effectively increase performance of condition-based maintenance: an investment choice between time to response and level of measuring accuracy