Framework for extracting and structuring raw data to strengthen
proactive decision making in maintenance
A thesis for the degree of Master of Science: Technology and Operations Management
Faculty of Economics and Business University of Groningen
The Netherlands
Author Ivar Luijmers
i.luijmers@gmail.com +31612789805
First supervisor dr. B. de Jonge Second supervisor dr. J. Veldman External supervisors Mr. E. Eiten - Stork
Mr. H. Assink - Nouryon
III
ABSTRACT
CONTEXT Condition-based maintenance (CBM) is becoming more popular because of increasing possibilities with respect to monitoring, storing and analysing data about the conditions of assets. However, many firms still follow time-based maintenance policies. This is because of incorrect data collection approaches, perceived complexity of applying statistical or mathematical models, and the difficulty to determine the actual business case to justify exploring raw data sources. Moreover, certain monitoring parameters are only suitable in experimental or theoretical settings.
OBJECTIVE The primary objective of this research is to facilitate the implementation of CBM in practice. This is done by designing a framework to structure the approach to move from raw data to measurements of degradation. In addition to this, it aims to explore new combinations of different types of data to provide insights in the degradation of an asset.
METHOD A three-phase design science approach is used. The first phase, problem investigation, consists of a literature study and a case study. This phase provides the basic constructs of the framework. The framework itself is developed in the second phase. The third phase consists of implementing the framework into practice in a case study at Nouryon and Stork. In the case study, process data of six cooling water pumps were considered. After that, the framework is validated in a workshop at Stork with experts in the Asset Management field. FINDINGS The application of the framework showed that the clogged coarse filter of the discharge cooling water pump potentially influences the degradation. A delay-time model is used to design and recommend a new condition-based maintenance policy. The experts in the validation workshop indicated that the framework is likely to be applicable in other contexts. VALUE AND ORIGINALITY This research contributes to filling the gap between theoretical CBM works and the application of these models in practice. First, by proposing a combination between the use of less accurate process data and the more exact offline inspections as a measurement of degradation. Second, by designing a framework to move from raw process data towards measurements of degradation. Furthermore, the application of this framework resulted in potential cost savings for Nouryon. Moreover, Stork could apply this framework as a roadmap for other clients.
IV
PREFACE
This thesis is the final chapter of the master’s degree Technology and Operations Management at the University of Groningen. This master program greatly contributed to my understanding of how organizations and their operations work. It enabled me to develop my academic skills and specialize myself in the Operations Management field.
During my master’s thesis project, I had the support of several people who were always available in case I needed them. Therefore, I would like to show my gratitude. First of all, I would like to thank dr. de Jonge for guiding me throughout my thesis project. The combination of his positive approach, critical and quick feedback was very pleasant to work with. Second, I would like to thank dr. Veldman for giving me the opportunity to propose my own thesis topic, by proposing my topic to dr. de Jonge and the more active role as second supervisor. Third, I would like to thank my external supervisors Henk Assink and Edwoud Eiten for their advice and the opportunity to collect data at Nouryon and Stork. Next, I would like to thank all the participants at Stork and Nouryon for their input and enthusiasm. Last but not least, I want to thank my girlfriend Ivana, family, and friends for their support throughout my time as a Bachelor and Master student.
Before you proceed with reading my thesis, I would like to give a brief analogy which perfectly captures the essence of my research project. When I started my thesis at Stork and Nouryon, I had an interesting conversation with a colleague about my research topic. After a while, we touched upon the subject of predictive maintenance. He immediately started laughing and said:
“Predictive maintenance!? We have been doing that for years. We put on our wizard suits and look into our glass ball and ask nicely when the equipment is going to break down”. In my
experience, this joke perfectly illustrates the perception of some about condition-based maintenance. It is a concept that is difficult to grasp without a clear example and is still sometimes perceived as fiction. By doing this research project, I would like to show that condition-based maintenance is not fiction and with the right example, is relatively easy to understand, and implement. I hope you enjoy reading my thesis.
Ivar Luijmers
V
ABBREVIATIONS
CBM Condition-based maintenance
CRISP-DM Cross-industry standard process for data mining P-F curve Potential failure – functional failure curve PHM Proportional Hazards Model
PM Preventive Maintenance
ROC Receiver operating characteristics RUL Remaining useful life
VI
LIST OF FIGURES AND TABLES
Figure 1: Typology of CBM types (Veldman et al., 2011a) ... 13
Figure 2: CBM stochastic deterioration models (Derived from Alaswad & Xiang, 2017 and De Jonge, 2017) ... 14
Figure 3: P-F curve (De Jonge, 2017) ... 15
Figure 4: Conceptual model – typical phases in a CBM application ... 19
Figure 5: Research design – three-phase design science approach ... 22
Figure 6: Condition-based maintenance framework ... 28
Figure 7: Erratic behaviour P1452B - dirt accumulation in the sensor ... 31
Figure 8: Erratic behaviour P1452A – recalibration of the sensor ... 31
Figure 9: Illustration of the coarse filter ... 32
Figure 10: Coarse filter (left) and discharge cooling water pumps (right) ... 33
Figure 11: P1452A – Wood in propeller on 12-05-2015 ... 33
Figure 12: P1452C – Broken scoop on 30-07-2016 ... 34
Figure 13: P1452C – High vibrations on 28-06-2018 ... 34
Figure 14: P1452C - High vibrations 28-06-2017 – Clogged coarse filter (blue) and increase in power usage (red) ... 34
Figure 15: Coarse filter with mud at the bottom ... 35
Figure 16: Expected deterioration process of P1452 before the historical failures ... 37
Figure 17: No deteriorated state ... 38
Figure 18: Deteriorated state ... 38
Figure 19: Total water flow output of the inlet cooling water pumps in 2015 to 2018 ... 52
Figure 20: P1452A – Level difference coarse filter and power usage (2015-2018) ... 53
Figure 21: P1452B – Level difference coarse filter and power usage (2015-2018) ... 53
Figure 22: P1452C – Level difference coarse filter and power usage (2015-2018) ... 53
Figure 23: P1452A – Coarse filter alarm threshold (2015-2018) ... 54
Figure 24: P1452B – Coarse filter alarm threshold (2015-2018) ... 54
Figure 25: P1452C – Coarse filter alarm threshold (2015-2018) ... 54
Table 1: Typical phases in the execution of CBM ... 11
Table 2: Overview of failure modes and indicators in centrifugal pumps (derived from McKee et al., 2011) ... 18
Table 3: Interviewees in the case study – expert knowledge ... 25
Table 4: Participants in the validation workshop ... 25
Table 5: Recent failures with respect to the cooling water pumps ... 30
Table 6: Process data with respect to the cooling water pumps ... 30
Table 7: ROC matrix – Alarm threshold of 300mm ... 39
Table 8: ROC matrix – Alarm threshold of 400mm ... 39
Table 9: ROC matrix – Alarm threshold of 250mm ... 40
VII
INDEX
ABSTRACT ... III PREFACE ... IV ABBREVIATIONS ... V LIST OF FIGURES AND TABLES ... VI
1. INTRODUCTION ... 8
2. THEORETICAL BACKGROUND ... 10
2.1 EXECUTION OF CBM ... 10
2.2 DATA COLLECTION ... 11
2.3 DATA ANALYSIS IN CBM ... 12
2.4 MODELLING THE DETERIORATION ... 13
2.5 DETERMINE THE OPTIMAL MAINTENANCE POLICY ... 15
2.6 PROCESS DATA FOR CONDITION MONITORING ... 17
2.7 FAILURE MODES IN CENTRIFUGAL PUMPS ... 18
2.8 CONCEPTUAL MODEL ... 19 3. METHODOLOGY ... 20 3.1 RESEARCH DESIGN ... 20 3.2 PROBLEM INVESTIGATION ... 22 3.3 TREATMENT DESIGN ... 23 3.4 TREATMENT VALIDATION ... 23 3.4.1 Framework application ... 24
3.4.2 Framework validation workshop ... 25
4. FRAMEWORK DESIGN ... 26 4.1 DATA COLLECTION ... 26 4.2 DATA ANALYSES ... 26 4.3 DECISION-MAKING ... 27 4.4 IMPLEMENTATION ... 27 5. FRAMEWORK APPLICATION ... 29 5.1 DATA COLLECTION ... 29 5.2 DATA ANALYSES ... 30 5.2.1 Failure diagnosis ... 30 5.2.2 Failure prognosis ... 37 5.3 DECISION MAKING ... 40
6. FRAMEWORK VALIDATION WORKSHOP ... 41
7. DISCUSSION AND CONCLUSION ... 43
7.1 MAIN FINDINGS ... 43 7.2 THEORETICAL IMPLICATIONS ... 44 7.3 PRACTICAL IMPLICATIONS ... 45 7.4 LIMITATIONS ... 45 7.5 FUTURE RESEARCH ... 46 7.6 CONCLUSION ... 47 REFERENCES ... 48 APPENDICES ... 52
APPENDIX A:TOTAL OUTPUT OF WATER OF THE INLET COOLING WATER PUMPS (2015-2018) ... 52
APPENDIX B:LEVEL DIFFERENCE AROUND THE COARSE FILTERS AND POWER USAGE (2015-2018) ... 53
8
1.
INTRODUCTION
Maintenance is a significant contributor to the total operational costs of manufacturing firms (Waeyenbergh & Pintelon, 2002; Jardine, Lin, & Banjevic, 2006). Modern technology is developing rapidly and higher quality and reliability is becoming more important (Jardine et al., 2006). Therefore, Condition-based maintenance (CBM) is becoming more popular. In this policy, maintenance actions are scheduled based on the conditions of the assets. This popularity is caused by increasing technical possibilities with respect to monitoring, storing and analysing data about the conditions of assets. Despite this development, many firms still follow time-based maintenance (TBM) policies (De Jonge, Teunter, & Tinga, 2017). In this policy, preventive maintenance actions are scheduled based on the time the asset is in service. The actual health of the asset is not considered, and any remaining useful life is wasted. There are various theoretical studies devoted to CBM, but the number of practical applications is limited (Jardine et al., 2006; Ahmad & Kamaruddin, 2012a; Bousdekis, Magoutas, Apostolou, & Mentzas, 2015; Tiddens, 2018). This is caused by incorrect collection approaches and the perceived complexity of implementing advanced statistical models. Furthermore, collecting data may not be practical, as certain monitoring parameters are only suitable in experimental or theoretical settings. Future research could study how various data sources could be combined and applied in practice to better estimate the condition of an asset and how to drive the development of more practical applications of CBM.
9
of data acquisition and cleaning techniques is beyond the scope of these studies. However, for CBM applications to be successful, it is essential that it is as simple, realistic, and practical as possible (Ahmad & Kamaruddin, 2012b). Tiddens (2018: 48) states that future research should focus on “methods that assist companies in determining which parameters need to be
monitored and which types of analytics are best suited to the firm for their specific application”.
Therefore, the second aim is to design a practical framework to move towards CBM and to provide a practical application of this framework.
This research project also has major practical relevance as a similar issue was identified by Nouryon, one of the largest specialty chemicals manufacturers in the world. They indicated that there is a relatively large amount of process data available with respect to various assets, currently used to operate the asset. This data could potentially be used to identify failures. However, it is unclear how to structure, analyse, and apply this raw data. In addition, it is difficult to determine the actual business case to justify exploring these data sources.
This research will therefore explore the following question: “How should potential
measurements, for instance process data and failure data, related to the condition of assets be collected, structured and applied to make proactive maintenance decisions?”. This question is
addressed by using a design science approach (Wieringa, 2014). The aim is to design a framework for extracting, structuring, and analysing raw data to move towards condition-based maintenance. This framework is designed based on a literature study and input from Nouryon and Stork. To test the validity of this framework, it is subsequently applied to a case study and validated by experts in the Asset Management field. In the case study, six cooling-water pumps were considered. Multiple data sources that could serve as potential measurements for degradation for these pumps were used. Examples are process data and offline inspection data with respect to vibration levels. These data sources were subsequently extracted, combined, and analysed and used to estimate the overall degradation process of the pumps.
10
2.
THEORETICAL BACKGROUND
This section presents the knowledge base of this study. The aim of this section is to increase the understanding of CBM as a concept and to position this research project in the current body of literature. Section 2.1 starts with the steps required to execute CBM. Section 2.2 discusses current literature with respect to data collection/processing in CBM. After that, Section 2.3 presents techniques for the analysis of process and failure data. Next, Section 2.4 elaborates on incorporating degradation measurements into a deterioration parameter. Furthermore, Section 2.5 discusses techniques to determine the optimal maintenance policy. Section 2.6 presents some examples of studies that use process data for condition-monitoring. After that, Section 2.7 briefly discusses the most common failure modes in centrifugal pumps. Finally, Section 2.8 presents the conceptual model and reflects on the most important theoretical concepts.
2.1 Execution of CBM
11
these cornerstones, Sillitti et al. (2019) incorporated different phases from the cross-industry standard process for data mining (CRISP-DM). The CRISP-DM defines a methodology for data mining. This methodology contains the following phases: business understanding, data understanding, data preparation, modelling, evaluation, and deployment. An overview of the typical phases in the execution of CBM is shown in Table 1.
Table 1: Typical phases in the execution of CBM
Jardine et al. (2006) Peng et al. (2010) Veldman et al. (2011a) Silliti et al. (2019)
1. Data acquisition 1. Data collection 1. Data collection 1. Root cause analysis (business understanding, data
understanding, data preparation, modelling)
2. Data processing 2. Diagnostics 2. Data analysis 2. Identification of RUL 3. Maintenance
decision making
3. Prognosis 3. Decision making 3. Alerting and predicting failures (modelling, evaluation,
deployment)
Overall, the phases and cornerstones identified by aforementioned studies correspond to each other. These steps are generally in line with the OSA-CBM (Open System Architecture for Condition-Based Maintenance) (Lebold et al., 2012), which is a general industry reference with respect to the software behind the CBM programs. However, the studies of Jardine et al. (2006), Peng et al. (2010), and Veldman et al. (2011a) only provide a list of the general steps and occasionally include a brief example. Silliti et al. (2019) provides a comprehensive list of tools that could be used in the different phases. However, they only provide brief examples. Examples that implement all these different phases step-by-step and in detail and apply them in practice are not provided. This means that the problem of perceived complexity of implementing advanced statistical models identified by Ahmad and Kamaruddin (2012a), is not addressed. To overcome this, Tiddens (2018) developed a framework for the selection of the optimal preventive maintenance approach. However, this study does not discuss the steps of extracting and structuring raw data that could potentially be used for condition monitoring. This study complements these studies by comprising all the steps into a framework and applying them to a real industrial case study.
2.2 Data collection
12
maintenance, and oil changes). Condition data relates to measurements of the health or state of the asset.
Veldman, Wortmann, and Klingenberg (2011b) identified two types of condition data, namely failure data, also identified by Jardine et al. (2006), and process data. Failure data is defined as direct expressions of the failure mode of an asset (e.g. vibrations or number, type, and size of metal particles in lubrication oil). Process data relates to the output characteristics of the asset (e.g. pressure, flow, and temperature). It could only be used to indirectly identify the failure mode (Tsang, 1995). It is therefore expected that the process data could only identify that there is an issue with the asset. Jardine et al. (2006) stress the importance of collecting both event and condition data. It could be used to compare the condition data before and after an event. Event data is important in assessing the performance of the condition indicators or features. 2.3 Data analysis in CBM
If the data is collected, the next step is to clean the data. Data cleaning increases the change that error-free data is used (Jardine et al., 2006). This is generally done graphically. It is necessary because plants or assets could show erratic behaviour during start-ups and shutdowns (Veldman et al., 2011a). According to Jermyn et al. (1999), the processing of data, requires 60 to 80 percent of the time involved in the process of extracting data. This is caused by outliers (i.e. data deviating from the normal situation) and missing data (i.e. lack of some values), that require correction (Larose, 2005). Outliers could be identified by looking at the inter-quartile ranges (IQR). Three approaches could be used to handle missing data: case deletion, parameter estimation, and simple imputation. Case deletion means that only complete observations are used. It should be used in situations where the number of missing data is relatively small. In parameter estimation, maximum likelihood procedures could be used. In simple imputation, missing data will be replaced with substituted values.
The next step is to analyse the data. Two types of models are generally used to identify trends of anomalies/deviations in the data, namely: analytical and statistical models (Veldman et al., 2011b). Veldman et al. (2011a: 42) defines these models as follows: “Analytical models are
cause-effect type of expressions of failure, whereas statistical models need historical data to calculate the probability of failure, along with its expected time to failure”. Veldman et al.
13
analysing the data. Examples of these tools are also presented in Figure 1. The framework presented in Figure 1 could be used to determine the approach required in the data that is available in the case study.
Figure 1: Typology of CBM types (Veldman et al., 2011a)
2.4 Modelling the deterioration
In TBM models only the lifetime distribution needs to be modelled, whereas CBM models requires the modelling of the deterioration of an asset (De Jonge, 2017). The condition of equipment can either be monitored continuously or by carrying out inspections. Stochastic deterioration can be modelled by various stochastic processes. The literature review of Alaswad and Xiang (2017) identified different deterioration models. These different models are presented in Figure 2. When the system conditions are directly observable, stochastic deterioration models are generally classified based on either discrete or continuous deterioration states. If the deterioration is caused by multiple factors, the proportional hazard model (PHM) is commonly used to model the multivariate failure models. According to De Jonge (2017), a model should be selected based on the practical situation (i.e. available condition data) and analytical traceability (i.e. simpler model might be more appropriate for more complicated problems).
The study of Sikorska, Hodkiewicz, and Ma (2011) categorize models into four main groups, namely knowledge-based models, life-expectancy models, artificial neural networks, and physical models. The study of Silliti et al. (2019) also uses this categorization. Knowledge-based models identify similarities between an observed situation and data of a historical failure. Based on these historical data, a life-expectancy is determined. Life-expectancy models determine the life expectancy of the asset based on the expected risk of deterioration under known operating conditions. Artificial Neural Networks compute an estimated output for the
Process data and statistical modelling E.g. Principal component
analyses
Failure data and statistical modelling
E.g Proportional Hazards Model
Process data and analytical modeling E.g Linear dynamic modelling with vibration
indeces
Failure data and analytical modeling
E.g. Use of parity relation to monitor outflow
14
remaining useful life of an asset from a mathematical representation, that has been based on observation data. Physical models estimate an output for the remaining useful life of an asset from a mathematical representation of the physical behaviour of the degradation processes.
Figure 2: CBM stochastic deterioration models (Derived from Alaswad & Xiang, 2017 and De Jonge, 2017)
A common approach in the literature is the delay-time model, which adds one deterioration state between the operating state and the failed state (Wang, 2012). The delay-time model is related to the P-F (Potential Failure – Functional Failure) curve, introduced by Mourbay (1997). This curve is illustrated in Figure 3. The P-F interval is the time between the moment that a potential failure (P) becomes detectable and the moment where it deteriorates into a functional
failure (F1,2) (De Jonge, 2017). In the delay-time model only the functioning, the deteriorating,
and the failed state are observable (i.e. good, deteriorated, and failed). When a continuous deterioration process is used, preventive maintenance is typically initiated when a certain deterioration threshold level M is exceeded. Preventive maintenance should be performed when the level of deterioration exceeds a certain threshold value (M). This strategy is called the control-limit policy. If failures do not occur at a constant deterioration level, randomness in the failure level should be modelled.
CBM modelling Discrete state deterioration Delay-time model Markov process Semi-markov process
PHM Continuous state deterioration
Brownian motion
Gamma process
15
Figure 3: P-F curve (De Jonge, 2017)
2.5 Determine the optimal maintenance policy
If the data considered in the case study contains the deterioration level of an asset, the optimal maintenance policy can be determined. Various aspects of the asset and its environment need to be considered in determining the optimal maintenance policy. This subsection presents the most important aspects.
The first aspect to consider is to determine the criterion that will be used for optimality. Alaswad and Xiang (2017) identified three optimality criteria: cost minimization, availability/reliability maximization and safety. These different criteria are often used individually and some of them could potentially be in conflict. When multiple objectives should be achieved simultaneously, multi-criteria methods should be used.
Next, it is important to determine what the effect of the preventive and corrective maintenance action is on the condition of the asset. For instance, preventive maintenance (PM) actions are generally imperfect. PM could return an asset to a state somewhere in between as-good-as-new and as-bad-as-old (Alaswad & Xiang, 2017). This could be modelled in two ways. The first method could be applied when the deterioration level moves through stages. When the system is in a certain deterioration stage, repair improves the system deterioration by, for instance, one stage. In the second method, the imperfect maintenance action could reduce the maintained system degradation by a random amount.
16
economic (Thomas, 1986). Economic dependence means that the “…joint maintenance of a
group of components does not equal the total cost of individual maintenance of these components.” (Nicolai & Dekker, 2006: 4). Stochastic dependence occurs if the condition of a
unit influences the lifetime of other units, whereas structural dependence occurs if the components structurally form a part of a system. The more recent study of Keizer, Flapper, and Teunter (2017) extended this categorization by distinguishing structural dependence from a technical and performance point of view. In most literature with respect to multi-unit systems, the optimal CBM policy is firstly determined per unit (Alaswad & Xiang, 2017). After that, the conditions of all the units will be combined to determine the optimal CBM policy of the whole system. In some cases, multi-unit systems could be considered as single-unit systems. This is because some units will fail more often than others and not all the unit failures will result in the failure of other units or the whole system.
17 2.6 Process data for condition monitoring
With respect to condition monitoring, prior research by Bloch and Geitner (1983) state that 99 percent of equipment failures are preceded by certain signs, conditions, or indicators that a failure is going to occur. Multiple potential sources could be used as parameters for condition monitoring, for instance vibration data, acoustic data, oil analysis data, temperature, pressure, moisture, humidity, weather, and environment data (Jardine et al., 2006; Ahmad, & Kamaruddin, 2012a). However, to the best of our knowledge, examples of studies that use process data to obtain information about the condition/degradation of equipment are limited. The study of Veldman et al. (2011b) provided a few examples of using process data for CMB, e.g. using voltage to monitor an electric point machine (i.e. equipment to operate railway turnouts). In addition, the study of Shin and Jun (2015) provides an example in which process data such as temperature, voltage, and pressure is used to estimate the condition of a locomotive. The limited number of examples of process data used for condition monitoring is in line with the study of Veldman et al. (2011b), in which they explain that process data can only be used to indirectly identify the failure modes.
18 2.7 Failure modes in centrifugal pumps
In the case study at Nouryon, six cooling water centrifugal pumps are considered. The study of McKee, Forbes, Mazhar, Entwistle, and Howard (2011) identified thirteen main failure modes that could arise in centrifugal pumps. These failures modes and their indicators are presented in Table 2.
Table 2: Overview of failure modes and indicators in centrifugal pumps (derived from McKee et al., 2011)
Failure modes Indicators, diagnosis and/or detection
Cavitation Erosion, noise, vibrations, efficiency reduction of the pump. Pressure
pulsations
Vibration of suction or discharge piping, instability of pump controls, and fatigue of internal pressure-containing component.
Radial thrust Hard to detect. Temperature rises in the bearing may or may not be a symptom of excessive radial loading.
Axial thrust Proximity-type sensors should be used to determine the axial movement of the shaft relative to the bearing housing.
Suction and discharge recirculation
Monitoring the pressure pulsations.
Bearing failure Bearing temperature, shock pulses, vibrations.
Seal failure Temperature gauges. Many fluids are affected by a change in their temperature, which may eventually lead to seal failure.
Lubrication Non-contaminated oil has a useful life of about thirty years at 30 degrees Celsius. This life span is halved for every 10 degrees Celsius rise in the temperature of the oil. Infrared thermography of the bearing.
Excessive vibrations
There are various overviews of vibration frequencies that can be found in a centrifugal pump, and the possible causes of each vibration.
Fatigue Three stages (1) crack initiation, (2) pre-existing defects, (3) final failure. These three stages of fatigue cracking can generally be observed on the fracture surface.
Erosion Cavitation erosion, adhesive wear, fretting, abrasive wear, or erosion by solid particle impingement.
Corrosion General, dealloying, galvanic, stress corrosion cracking, hydrogen
embrittlement, microbiologically induced corrosion, intergranular corrosion. Excessive power
consumption
19 2.8 Conceptual model
To conclude this section, the literature shows that there is an overall consensus about the typical steps and phases required in the execution of CBM. It also shows that there are multiple tools and techniques available in the literature to extract raw data and subsequently analyse it. However, little research is focussed on applying these tools and techniques step-by-step on potential measurements of degradation.
The combined steps presented by Jardine et al. (2006), Peng et al. (2010), Veldman et al. (2011a) and Silliti et al. (2019) are illustrated in the conceptual model (Figure 4). However, the in-depth steps required to move from raw data to actual measurements of degradation are not provided and applied in an industrial application. This study addresses this gap by combining these different steps into one framework and by applying it in practice.
Figure 4: Conceptual model – typical phases in a CBM application
20
3.
METHODOLOGY
This section presents the methodology of this study. The aim is to provide the information required to replicate this study and what is done to ensure that the study is reliable and valid. It is divided into four subsections. Section 3.1 describes the research design and why the design science approach fits with the aim of this study. Section 3.2 elaborates on the problem investigation phase. Next, Section 3.3 describes the design of the framework. The aim is to design a framework that treats the problem identified in the problem investigation phase, presented in the introduction, and the theoretical background. After that, Section 3.4 discusses the treatment validation phase. The aim is to determine validate the framework in practice. The research question is as follows:
“How should potential measurements, for instance process data and failure data, related to the condition of assets be collected, structured and applied to make proactive maintenance decisions?”
To properly answer this question, it is divided into the following subquestions:
1) How to determine whether certain measurements (e.g. process data) are related to the degradation of equipment and could be used for condition monitoring?
a) How to extract and structure raw data?
b) How to convert this raw data to actual measurements of degradation?
2) How could these measurements of degradation be integrated into a single deterioration parameter and how should the occurrence of failure be modelled?
3) How could the single deterioration parameter contribute to improving the maintenance policy?
3.1 Research design
21
In this study, the artefact is the framework to determine whether potential condition measurements are actual indicators of degradation. The performance of this artefact is researched in the case study at Nouryon and Stork. The design science approach focusses on making a clear contribution in the application environment (Hevner, 2004). It is pragmatic in nature and has a strong emphasis on relevance. This perfectly fits the aim of this research, as applying CBM should be as simple, realistic, and practical as possible (Ahmad & Kamaruddin, 2012b).
22
Figure 5: Research design – three-phase design science approach
3.2 Problem investigation
In the first phase of this research, the problem is investigated. This is done by the means of a literature study. The aim is to obtain a better and in-depth understanding of the problem. The literature served as a knowledge base. The results of this literature study are presented in the previous section, the theoretical framework. The main research question and the subquestions are used as a basis of the literature study. Based on this, initial scientific papers were found. These initial papers were used for a forward and backward search (Karlsson, 2016).
The application environment of this framework is Nouryon and Stork at the Chemical Park in Delfzijl, the Netherlands. Nouryon (previously AkzoNobel Specialty Chemicals) produces chemicals to manufacture everyday products such as paper, plastics, building materials, and personal care items. Nouryon manufacturers salt, chlorine, mono-chloroacetic acid, and operates a cogeneration plant in Delfzijl. These different plants and processes utilize a wide range of different assets that require maintenance. This maintenance is partially executed by on-site third-party maintenance contractor Stork. Stork provides services with respect to maintenance, modifications, and asset integrity solutions.
Literature study Case study
Develop framework for extracting, structuring and
analysing raw data to move towards
condition-based maintenance
Workshop with experts to validate the
framework
Apply the framework in the case study
23
In this context, there are several stakeholders with different interests. Nouryon indicated that there is a relatively large amount of process data available with respect to various assets (e.g. rotating equipment) currently used to operate the asset. This data could potentially be used to identify deterioration levels and failures. However, it is unclear how to structure, analyse and apply this raw data. Time-based maintenance is based on the years of experience with the assets. In addition, it is difficult to determine the actual business case to justify exploring these data sources. The framework could guide and structure this. In addition, Stork indicated that there are various other customers with similar issues. Stork could use this framework to optimize maintenance policies for Nouryon and other customers.
3.3 Treatment design
The literature study shows that there is a consensus about the general steps required in the application of CBM. The phases presented by Veldman et al. (2011a), are used as a basis for the framework. The steps in each of these phases are derived from Jardine et al. (2006), Veldman et al. (2011), Alaswad and Xiang (2017), and Silliti et al. (2019). The input from the case study is used to further complement the framework. The framework could be categorized as an emergent system.
3.4 Treatment validation
24 3.4.1 Framework application
Nouryon has three inlet and three discharge cooling water pumps. These pumps supply and drain water to and from the Wadden Sea. These pumps are used to cool the machines and evaporators in the production of salt. Furthermore, these pumps are critical for multiple production plants of Nouryon. For instance, a by-product in the production of salt is chlorine. Chlorine is subsequently used to produce mono-chloroacetic acid and by other third parties on the chemical park. Therefore, if the cooling water pumps fail, it results in the loss of production in multiple production plants of Nouryon and other firms.
Nouryon currently follows a time-based maintenance policy. Initially, the pumps were maintained and revised every five years, but this has been extended to six years because of budget cuts. Because of the high criticality, all the pumps are completely revised, regardless of the actual health state.
Despite these full revisions, there have been three major failures in the last four years. For instance, the scoop of the discharge drainage pump broke. This resulted in the need for corrective maintenance. Furthermore, there was a decrease in the production of salt and chlorine. In addition to this, a temporary pump had to be installed to maintain the production levels. This resulted in a relatively high amount of costs. Therefore, Nouryon aims to move towards a CBM policy.
There is continuous process data available with respect to the pumps. In addition, Stork performs three-monthly inspections including a vibration analysis. Currently, there are no historical analyses performed with respect to this data. The question of Nouryon is: firstly, whether more predictability about potential failures can be achieved based on the data and expert knowledge? Secondly, is there a difference between the data prior to a failure in comparison to the “healthy” state? Thirdly, is there a relationship between the failure and the measured data? To answer these questions, the framework designed in the previous phase will be applied in this case study.
25
The qualitative aspect relates to interviews with experts to supplement these insights. Furthermore, detecting abnormalities in the data will be done in consultation with experts. The interviewees are presented in Table 3. According to Karlsson (2016) and Yin (2003), a case study is very applicable when (1) situations need to be studied in a natural setting (i.e. practical situation) and (2) to answer why, what and how questions. Therefore, this method fits perfectly with the aim of this research project and the research questions.
Table 3: Interviewees in the case study – expert knowledge
# Organization Function
1 Stork Machine Diagnostics Engineer 2 Nouryon Reliability Engineer Rotating 3 Nouryon Production Assistant 4 Nouryon Project Manager
3.4.2 Framework validation workshop
The framework is applied in the narrowly defined case study presented in the previous subsection. Due to time constraints, it was not possible to apply the framework in different contexts. It is argued by Karlsson (2016) that confirmatory interviews could be performed to test the validity of theoretical models. To realise this, a validation workshop is performed with experts in the Asset Management field. These experts are listed in Table 4. The aim was to “self-destruct” the framework to discover any weaknesses or expected issues when applying this framework in different contexts.
Table 4: Participants in the validation workshop
# Organization Function Area of expertise
1 Stork Account Manager Implemented and worked with CBM-programs primary in the oil & gas industry.
2 Stork Operations Manager Machine Diagnostics
Manager of the department that performs condition monitoring.
3 Stork Machine Diagnostics Engineer
In depth mechanical knowledge of various types of rotating equipment. Performs vibration analyses.
26
4.
FRAMEWORK DESIGN
This section provides the development process of the framework and the framework itself. The aim is to develop a framework that treats the problem described in Sections 1 and 2. The conceptual model in the theoretical background is used as a basis for the framework. The framework is further complemented with literature and input from the case study. Each section briefly summarizes the steps required in each of the four phases of CBM. The overall framework is presented in Figure 6 and designed as a flowchart. The framework is illustrated as a strict sequential model. However, the actual application is meant to be iterative.
4.1 Data collection
The first step is the selection of an asset for which CBM could be beneficial (Jardine et al., 2006). The second step is to collect and extract the data. There are several inputs required to determine whether some parameters are actually indicators for the condition of an asset: event data and condition data (i.e. process and failure data) (Jardine et al., 2006; Bousdekis et al., 2015). The condition data should be collected around the event data.
4.2 Data analyses
The data analysis is divided into a diagnosis and prognosis phase (Peng et al., 2010). The main aim of the diagnosis phase is to identify the abnormalities in the condition data around the different events in the historical data. The first step is to clean the data. If there are a relatively large numbers of sensors producing a large number of variables, a principle component analysis (PCA) could be performed to reduce some of the variables that show the same behaviour (Jardine et al., 2006). To identify any erratic behaviour, one could graphically plot the data and see if there any outliers. One could also look at the interquartile ranges of the data.
27
statistical correlation, a relatively large amount of event data and patterns in the condition data should be available. If there are too little number of patterns or events available to make statistical correlations, expert knowledge could be used to strengthen the correlations.
The main aim of the prognosis phase is to estimate the RUL of the assets and which patterns in the data show that the health is degrading. Some deterioration parameters could correspond with a single failure mode, while some failure modes will need multiple deterioration parameters (Lu et al., 2007). There are multiple prognostic models available in the literature that could be used to estimate the RUL. The studies of Sikorska et al. (2011), Alaswad and Xiang (2017) and Silliti et al. (2019) both provided an overview of the available prognostic models. A model should be selected based on the practical situation (i.e. available condition data) and analytical tractability (i.e. a simpler model might be more appropriate for more complicated problems) (De Jonge, 2017).
4.3 Decision-making
The prognostic model should be incorporated in the maintenance policy (Jardine et al., 2016). The maintenance policy must be consistent with the optimization criteria (e.g. cost minimization, availability/reliability maximization and/or safety) (Alaswad & Xiang, 2017). In the CBM policy, appropriate threshold levels should be determined to make certain maintenance actions. Optimal thresholds could be computed mathematically or approximated with simulation techniques. After the new maintenance policy is determined, the internal decision-making process should be followed.
4.4 Implementation
28
If a decision is made, proceed to the implementation phase. Estimation of the remaining useful life (RUL).
Determine threshold(s) The
condition level of the parameter at which there is a high failure probability.
Number of variables?
Structure the data
Compare data before and after the events and with "healthy"/ normal behaviour. Identify any patterns or abnormalities.
Clean the data
Identify outliers (min, max & interquartile ranges). Plot the data and identify erratic behaviour during events like start-ups or shut-downs.
Phase 2: data analysis - diagnosis (Jardine et al., 2006; Veldman et al., 2011)
Identify event data and failure modes
Failures, breakdowns, overhauls, revisions, corrective and preventive maintenance, start-ups, and/or shut-downs.
Identify condition data
1) Failure data - e.g. vibrations or number, type, and size of metal particles in lubrication oil 2) Process data - e.g. pressure, flow, temperature, and current.
Collect/extract data
Collect the event data. Collect/extract condition data around the event data. Collect/extract data during a "healthy behaviour"
Phase 1: data collection (Jardine et al., 2006; Veldman et al., 2011)
Type of correlation?
Analytical
Statistical
Clean the data
Statistical techniques Principle Component Analyses (PCA) to reduce number of variables showing similar behaviour.
Identify analytical or statistical correlations between trends, abnormalities, and patterns in the condition data and events.
High
Low
Phase 2: data analysis - prognosis (Silliti et al. 2019)
Establish statistical expressions between explanatory and response variables.
Phase 3: decision making (Alaswad and Xiang, 2017)
Select asset to monitor
Health, safety, environmental, customer-related, financial, and/or objectives and issues.
Establish cause-effect type of expressions between explanatory and response variables.
Select deterioration parameter(s). Determine the historical trend of deterioration.
Select model based on the practical situation (i.e. available condition data) and analytical traceability (i.e. simpler model might be more appropriate for more complicated problems). Consult e.g. Sikorska et al. (2011) or Silliti et al. (2019).
Determine the criteria that makes the maintenance policy optimal Three optimization
criteria: cost minimization, availability/reliability maximization and safety
Which actions would resolve certain events? Determine the optimal moment when to perform these actions.
Follow the internal decision-making process.
29
5.
FRAMEWORK APPLICATION
This section presents the application of the framework (Figure 6) in the case study of Nouryon and Stork. The aims are to discover if the framework works, to identify any issues with the framework, and to determine if any elements or activities are missing. This section is structured based on the framework.
5.1 Data collection
As explained earlier, there have been three major failures of cooling water pumps in the last four years. The event data are listed in Table 5. The failures have been costly in terms of maintenance and production losses. Nouryon collects process data with respect to these pumps, but these are currently only used to operate the assets. Nouryon wants to research if, based on this data, a CBM policy could be developed and implemented.
There are three inlet cooling water pumps (P1451A, P1451B, P1451C) and three discharge cooling water pumps (P1452A, P1452B, P1452C). The inlet cooling water pumps did not have any major issues in the last four years. There were only some balancing issues during the start-up after time-based maintenance of one of the pumps. Nouryon performs certain corrective and preventive maintenance actions. Every five years, Nouryon fully revises the pumps. This is done separately for each pump. If the coarse filter is clogged, the filter could be removed and cleaned. According to the Project Manager, the total cost of each failure of the cooling water discharge pumps range from around 700.000 to 1.000.000 euros. These costs are mainly caused by production loss, the costs of installing, and operating a backup pump, and repair of the failed the pump.
30
Table 5: Recent failures with respect to the cooling water pumps
# Event type Date Action Cause
1 Failure pump P1452A 12-05-2015 Corrective maintenance Wood in propeller 2 Failure pump P1452C 10-06-2016 Corrective maintenance Broken scoop 3 High vibrations P1452C 28-06-2017 In-depth inspection High vibrations 4 Failure during start-up P1451A 11-05-2018 Corrective maintenance Balancing issue
Nouryon collects both process and failure data and is continuously measuring multiple types of process data. These different types are listed in Table 6. The process data is extracted from the database with an interval of ten minutes in a period of five years, from 2015 up to and including 2018. With respect to failure data, Stork is executing three-monthly offline inspections to measure the vibration levels.
Table 6: Process data with respect to the cooling water pumps
Parameters Inlet cooling water pump Discharge cooling water pump P1451A P1451B P1451C P1452A P1452B P1452C Recorded motor current x x x x x x Bearing temperature x x x
Blade position x x x x x x
Vacuum exhaust x x x
Flow lubricate water supply (bearing) x x x x x x Pressure lubricate water supply (bearing) x x x
Flow cooling water exhaust x x x Pressure cooling water exhaust x x x Level difference band-pass filter x x x
Level difference coarse filter x x x x x x Level before coarse filter x x x x x x
5.2 Data analyses
Section 5.2.1 contains the failure diagnosis phase, in which the abnormalities are identified and analysed. Section 5.2.2 contains the failure prognosis phase, in which faults and degradation are predicted in the future.
5.2.1 Failure diagnosis
31
current alarm policy and threshold, and the current time-based maintenance policy. Lastly, an overall diagnosis is provided.
Data cleaning
The first step is to clean the data. This is done graphically because of the relatively small number of sensors. Three main types of erratic behaviour are identified. First from July 2015 to the beginning of December 2015, the sensors showed some high variability in the level difference around the coarse filter. This is shown in Figure 7. According to the Production Assistant, it was likely caused by a sensor failure due to dirt accumulation. Second, when the pumps are not in use, the level difference around the coarse filter occasionally exhibit extreme values or high variability. Third, the production assistant indicated that in 2018 there were some re-calibrations of the sensor. This is shown in Figure 8.
Figure 7: Erratic behaviour P1452B - dirt accumulation in the sensor
32 Identifying patterns
Patterns were found in the collected failure and event data by plotting the data. The main finding is that a few weeks before every failure of the cooling water discharge pumps, the level difference in the coarse filter increases. The main purpose of the coarse filter is to avoid that any coarse pieces of waste, damages the components of the pump (e.g. the scoop). Each of the three cooling water discharge pumps, have an individual water basin. Before these individual basins, there is a shared water basin in which the used water is stored. The coarse filters are located between these individual and shared water basins. The level difference around the coarse filters is computed by measuring and subtracting the level of water before and after the filter (1). The level difference indicates to what extent the filter is clogged. This is illustrated in Figure 9. Pictures of the cooling water discharge pumps and the coarse filter are shown in Figure 10.
Level difference coarse filter = water level shared basin - water level individual basin (1)
33
Figure 10: Coarse filter (left) and discharge cooling water pumps (right)
A few weeks before each of the failures with respect to the cooling water discharge pumps (Table 5), the coarse filter was clogged for three to five weeks. This is shown in Figure 11, 12 and 13. This could cause the pump to suck up air instead of water. According to the machine diagnostics engineer, it could result in an increase of shocks/vibrations and the power consumption. The increase in power consumption is visible before the failures of P1452C (Figure 14). If the coarse filter is clogged for a relatively long period, these shocks could create more freedom of movement of the pump and could also increase the vibrations after the clogged coarse filter is cleaned. This could potentially result into metal fatigue in some of the components of the pump.
34
Figure 12: P1452C – Broken scoop on 30-07-2016
Figure 13: P1452C – High vibrations on 28-06-2018
Figure 14: P1452C - High vibrations 28-06-2017 – Clogged coarse filter (blue) and increase in power usage (red)
Relating patterns to failures
35
Engineer expects that a pump sucking up air would present a similar pattern. Shortly after the inspections, wood came into the propeller and caused the pump to fail. According to the Reliability Engineer, it is expected that the wood fell into the individual water basin from above. This is therefore not related to the increasing level difference of the coarse filter. However, the vibration analysis and the expert knowledge of the Machine Diagnostic Engineer further correlates the level difference around the coarse filter to the vibrations.
Potential causes of the pattern (i.e. clogged filter)
The clogging of the coarse filter has two main causes. First, according to the Production Assistant, the clogged coarse filter is mainly caused by the growth of algae and mussels. A sudden increase in the level difference around the coarse filter is generally occurring somewhere in the spring. The Production Assistant indicated that there is an increase in the growth of mussels and algae when the temperature is rising (generally above ten degrees Celsius). To counter this, bleach is added to the cooling water. In fall and winter, the bleach is added one hour per twenty-four hours. In spring and summer, the bleach is added for ten minutes with an interval of ten minutes. This increase of the bleach dosage causes the algae and mussels that are attached to walls of the pipelines to let go and move into the shared water basin. It is expected that this could be one of the causes of the sudden increase in the level difference of the coarse filter in the spring. Second, according to the Project Manager, there could be an accumulation of mud around the filter which causes the level difference around the coarse filter to increase. An example is this is shown in Figure 15.
36 Threshold to clean the coarse filters
The current alarm threshold for the level difference around the coarse filter is 300mm. The filter can be cleaned by lifting out the filter with a crane. After the filter has been removed, it is cleaned with a high-pressure cleaner. Next to that, the mud in the water basins is pumped out and cleaned by divers. According to the Production Assistant, the total costs amount to approximately 5000 euro. According to the Production Assistant, the current planning interval is around ten days. However, cleaning the filters is currently not the highest priority. The historical data of the level difference around the coarse filter shows that cleaning is generally done after three to six weeks. The level difference around the coarse filter in 2015 to 2018 and the threshold is depicted in Appendix C.
Current time-based maintenance policy
The demand for cooling water is higher during the summer. This is because of a higher water temperature and less cooling capacity. This total output of cooling water inlet pups is shown in Appendix A. During the level increase around the coarse filter before the high vibrations of P1452C on 28-06-2017, the level also increased for P1452B. The level increase lasted around five weeks for both pumps. During this increase, pump 1452A was under revision. This in combination with the higher demand for cooling water during spring and summer, hindered the possibility to clean the filters. After the pump failed on 28-06-2018, the level started to increase again for P1452C. Briefly after the failure of P1452C, the revision of P1452A was finished and the pump is utilized again.
Overall diagnosis
It remains difficult to statistically correlate the water level around the coarse filter to failures listed in Table 4 due to the limited number of failures in the last years. However, the Project Manager, Machine Diagnostics Engineer, Production Assistant, and the Reliability Engineer stated that this could hypothetically be one the causes of the two failures of P1452C. The researcher would like to emphasize that this does not mean that these failures are definitely caused by the clogged coarse filter. The expert knowledge and the level difference around the coarse filters before the two failures of P1452C, only indicate that a clogged coarse filter is positively increasing the degradation of the pumps.
37
the pump broke down, it can be observed that the pump has failed. Results show that with respect to the two failures of the pump P1452C (events #2 and #3 – Table 5), the alarm went off and the pump likely entered into a deteriorating state. Because cleaning the filters was not a priority, the filters were cleaned after five weeks. This means that the pumps were deteriorating for five weeks. Around four weeks after the filter was cleaned, the pump failed due to vibrations and/or metal fatigue. During this deteriorated state there was no vibration analysis due to the three-monthly planning interval. This is illustrated in Figure 16.
Figure 16: Expected deterioration process of P1452 before the historical failures
5.2.2 Failure prognosis
The diagnosis phase shows that it is likely that the clogged coarse filter is resulting in a health deterioration of the cooling water discharge pumps. However, the actual impact of this on the health of the pump is difficult to quantify due to the lack of historical failures. Moreover, the behaviour and shape of level difference around course filter over a time period is difficult to mathematically model. Therefore, it is difficult to mathematically compute the optimal threshold for cleaning the filters. It is however possible to approximate the optimal threshold with a sensitivity analysis.
The clogged filter only indicates that the pump is deteriorating, it does not give any information at the current health state of the pump. Meanwhile, the three-monthly vibration analysis does provide an assessment of the current health state of the pump. A combination of these two data sources could possibly provide a better prognosis of potential failures. If an alarm threshold of the level difference around the coarse filter is reached, the filters should be cleaned. According to the Production Assistant, cleaning the filters should be scheduled and performed in ten days. During these ten days the pump will potentially be deteriorating. If the filter has been cleaned, a vibration analysis should be performed to assess the health of the pump.
38
This assessment can have two outcomes. First, the clogged filter did not cause freedom of movement of the pump and the state of the pump can be assessed as “good”. Second, the clogged filter did cause freedom of movement of the pump and potentially metal fatigue. In this case, the current state of the pump can be assessed as “deteriorated” (Figure 17). Based on this assessment preventive maintenance could be scheduled, and this will expectedly return the state of the pump back to “good” (Figure 18).
Figure 17: No deteriorated state
Figure 18: Deteriorated state
To determine whether the current threshold of 300mm for the alarm is suitable, a receiver operating characteristics (ROC) analysis could be used to depict the trade-off between hit rates and false alarm rates. Results of the ROC analysis are shown in Table 7. In this analysis, the alarms that would occur while the pump is not running, and the erratic behaviour identified in the diagnosis are omitted. Expert knowledge indicates that if the procedure depicted in Figure 18 and 19 would be followed, it is likely that the two failures of the pump 1452C would be detected with a vibration analysis. At the same time, expert knowledge indicates that the failure of pump 1452A would not have been detected. With the current alarm threshold, there would have been four situations in which a vibration analysis would be performed while the deteriorated state was not reached.
Good Deteriorating Good
Alarm Coarse filter Clean filter Vibration analysis
39
Table 7: ROC matrix – Alarm threshold of 300mm
1452
True condition leading to a failure Positive Negative A B C A B C Predicted deterioration leading to a failure Positive A 0 - - 0 - - B - 0 - - 3 - C - - 2 - - 1 Negative A 1 - - 0 - - B - 0 - - 0 - C - - 0 - - 0
If the threshold would be increased to for instance 400mm (depicted in Table 8), the number of type I errors would decrease by one. However, the failure of P1452C on 30-7-2016 would be detected around one week later. This would result into a higher change of damaging the pump.
Table 8: ROC matrix – Alarm threshold of 400mm
1452
True condition leading to a failure Positive Negative A B C A B C Predicted deterioration leading to a failure Positive A 0 - - 0 - - B - 0 - - 2 - C - - 1 - - 1 Negative A 1 - - 0 - - B - 0 - - 0 - C - - 1 - - 0
40
Table 9: ROC matrix – Alarm threshold of 250mm
1452
True condition leading to a failure Positive Negative A B C A B C Predicted deterioration leading to a failure Positive A 0 - - 2 - - B - 0 - - 3 - C - - 1 - - 1 Negative A 1 - - 0 - - B - 0 - - 0 - C - - 1 - - 0 5.3 Decision making
Based on the diagnosis and prognostic phase, there are two main decisions to be made. First, it is recommended to schedule the time-based revisions of the pumps in the fall or winter. This is because results show that the coarse filter is more likely to become clogged during spring or summer. Moreover, the demand for cooling water is higher during spring and the summer. This means that if one of the pumps is in revision and in the meantime another pump failed, it could be difficult to meet the demand.
Second, it is recommended to implement a condition-based inspection policy next to the time-based inspection policy. This means that if the alarm threshold for level difference around the coarse filter is reached, the coarse filter should be cleaned within the ten days planning interval. After the coarse filter is cleaned, an inspection with a vibration analysis should be performed. By doing to this, the actual health state of the pumps can be determined, and any metal fatigue of the components could potentially be detected before complete failure.
41
6.
FRAMEWORK VALIDATION WORKSHOP
Before the design science cycle proceeds to implementation, it is important that the process of validation takes place to evaluate the soundness of the developed artefacts (Wieringa, 2014). The designed framework has been validated by the means of a validation workshop with three experts in the Asset Management field. The aim is to find any weaknesses or issues when applying the model into a different context.
The first step of the framework is to select an asset to monitor, this is mainly dependent on the internal organizations. According to the Account Manager: “…organizations generally have
an overview of their most critical equipment in terms of safety, health, environmental, compliance, and financial aspects. This list serves as a prioritization of the of equipment that would require improvements in the maintenance policy”. The Operations Manager indicated
that in the framework, it is only depicted as one small step. However, in practice, the type and nature of the equipment or asset is an essential success factor in the application of CBM. With respect to the data collection, there is an overall consensus among the experts that different types of data listed in framework would be sufficient to perform a diagnosis of historical failures. The Account Manager indicated that the costs related to the events are an important input for the decision-making stage. Furthermore, the Machine Diagnostics Engineer states that:
“...the Original Equipment Manufacturer is likely to have more information about the different failure modes of their equipment, which in turn could be useful in the diagnosis phase”.
The experts indicate that the depicted steps in the diagnosis phase of the data analysis are clear. The experts agree on if there is a large amount condition data available and there are sufficient events related to a certain equipment, these steps would enable them to establish possible correlation between the events and the condition data.
With respect to the prognosis phase of the data analysis, the experts indicate they would prefer a more detailed elaboration on the different prognostic’s models. According to the Operations Manager: “…an overview of the different available prognosis model and which information is
required, and which context, type of equipment is most suitable for each of this model, would enable the development of more practical applications of condition-based maintenance”. The
42
determine the deterioration of an asset, an overview is required of which model is suitable in which situation. This should be supplemented by examples of each of these models”.
The steps depicted in the decision-making phase are sufficient according to the experts. The Account Manager indicated that “…the decision-making is indeed very much context
dependent, for instance for some organizations safety is more important than availability. Furthermore, the internal procedures with respect to implementing is also different in every organization”.
Overall, the strengths and weaknesses according to the experts are summarized in Table 10. Recommendations for further research are listed in the discussion and might address these weaknesses.
Table 10: Strengths and weaknesses of the framework
Phases Strengths Weaknesses
Data collection All the required types of data are listed
Selecting a unit to monitor is more complicated than depicted and
influences whether CBM is possible or not
Data analysis The diagnosis phase is clear. Relating patterns to events is an approach that experts would also use
Selecting a type of model for the prognosis is unclear. Examples of each type of model are required
Decision-making The decision-making phase is context
43
7.
DISCUSSION AND CONCLUSION
In this final section, the discussion and conclusion are provided. The discussion reflects on the research objectives and the results, the theoretical and practical implications, the limitations, and the future research recommendations. This section ends with the conclusion.
7.1 Main findings
This research started by introducing the problem that, despite the increasing technical possibilities to monitor, store, and analyse data, many firms still use time-based maintenance instead of condition-based maintenance (CBM) policies (De Jonge, 2017). According to existing studies (Jardine et al., 2006; Ahmad & Kamaruddin, 2012a; Bousdekis et al., 2015), this is primary caused by incorrect data collection approaches, the complexity of advanced statistical models, and the difficulty to collect certain monitoring measurements as they are only suitable in experimental settings. To overcome these issues, the objectives of this research were (1) to drive the development of more practical applications of CBM and (2) to identify new combinations of data sources to better estimate the condition of an asset.
