This is an author’s copy of the paper, which has been accepted and published in Journal of Computers and

Industrial Engineering

Big data analytics architecture design—an

application in manufacturing systems
Mahdi Fahmideh, Ghassan Beydoun
Faculty of Engineering and Information Technology, University of Technology Sydney, Sydney 2007, Australia

Abstract

Context: The rapid prevalence and potential impact of big data analytics platforms have sparked an interest
amongst different practitioners and academia. Manufacturing organisations are particularly well suited to benefit
from data analytics platforms in their entire product lifecycle management for intelligent information
processing, performing manufacturing activities, and creating value chains. This requires re-architecting their
manufacturing legacy information systems to get integrated with contemporary data analytics platforms. A
systematic re-architecting approach is required incorporating careful and thorough evaluation of goals for data
analytics adoption. Furthermore, ameliorating the uncertainty of the impact the new big data architecture on
system quality goals is needed to avoid cost blowout in implementation and testing phases.
Objective: We propose an approach to reason about goals, obstacles, and to select suitable big data solution
architecture that satisfy quality goal preferences and constraints of stakeholders at the presence of the decision
outcome uncertainty. The approach will highlight situations that may impede the goals. They will be assessed
and resolved to generate complete requirements of an architectural solution.
Method: The approach employs goal-oriented modelling to identify obstacles causing quality goal failure and
their corresponding resolution tactics. It combines fuzzy logic to explore uncertainties in solution architectures
and to find an optimal set of architectural decisions for the big data enablement process of manufacturing
systems.
Result: The approach brings two innovations to the state of the art of big data analytics platform adoption in
manufacturing systems: (i) A systematic goal-oriented modelling for exploring goals and obstacles in integrating
manufacturing systems with data analytics platforms at the requirement level and (ii) A systematic analysis of
the architectural decisions under uncertainty incorporating stakeholders’ preferences. The efficacy of the
approach is illustrated with a scenario of reengineering a hyper-connected manufacturing collaboration system
to a new big data architecture.

Keywords: big data, big data analytics platforms, manufacturing systems, goal-oriented modeling, fuzzy logic

1 Introduction
Product lifecycle management is a data intensive process comprising market analysis, product design,
development, manufacturing, distribution, post-sale, and recycling (Stark, 2015). The process
involves a variety of voluminous data, e.g. customers’ comments on social media, product functions,
product configuration, and failure incidences reported by installed sensors to monitor parameters of
environment and products. Manufacturing organisations view such data as a valuable business asset to
achieve good performance and to reduce cost in the product lifecycle. They also regularly seek to
increase their productivity using new advanced information technologies that place further demand on
their data processing storage requirements such as Internet of Thing (IoT) and radio-frequency
identification (RFID) tags in their daily production. For example, Toyota automotive company equip
cars with smart sensors and continuously collecting data about its locks, location, ignitions, and tyres
which can be later used by the manufacturer assembly. Continuous product innovations lead to further
product data generation coupled with a great diversity of types, sources, meaning, and format.

1
Given its increasing volume and variety, manufacturing data is increasingly difficult to process using
common manufacturing data platforms be they computer aided design (CAD), supply chain
management (SCM) manufacturing execution system (MES), or enterprise resource planning (ERP).
Indeed, the high volume, velocity, variety, veracity, and value adding data requirement all point to the
need to complement manufacturing systems with big data platforms (McAfee, Brynjolfsson, &
Davenport, 2012). New platforms such as Apache Hadoop, Google’s Dremel, or S4 are promising
ways forward to address the abovementioned processing complexity (Lycett, 2013). They provide a
support for capturing, processing, and visualising large volume of data sets that organisational
systems may have collected over the years.
Taming big data has the potential added benefit to analyse real-time data across different phases of
product lifecycle management from receipt of a customer’s order, identifying promising customers,
collecting variable data about the quality of raw material, selecting a detailed design, procuring,
selecting suppliers and outsourcing policies, and product warehousing, maintenance, recycling, and to
identifying labor errors have been discussed (Li, Tao, Cheng, & Zhao, 2015), (Protiviti, 2017),
(Waller & Fawcett, 2013), (Bi & Cochran, 2014), (Dubey, Gunasekaran, Childe, Wamba, &
Papadopoulos, 2016).
Compared to others fields such as electronic commerce, financial trading, health care, and
telecommunication, the manufacturing field seems to be slow in pace in adoption big data analytics
platforms in their business processes (Li et al., 2015). This could be attributed to the high capital costs
associated with manufacturing systems which automate, process, and integrate data flows between
one or more above phases and typically composed of a number of software systems, machines,
transportation devices, and so on (Camarinha-Matos, Afsarmanesh, Galeano, & Molina, 2009).
Nevertheless, manufacturing organisations recognise the value of big data analytics and the fact that
their adoption failure poses a risk to their operating and financial performance (Protiviti, 2017).
Gartner reports that 60 percent of big data projects fail to get piloting and production due to reasons
such a lack of adequate IT skill set, inability to understand stakeholders requirements in utilising data
analytics, and disparate legacy systems (Gartner, 2015), (Wegener, 2013). Thence, reluctance of
manufactures in moving to these platforms is unsurprising. Some are also still figuring out what kind
of data is worthy for advanced data analytics and which stage of product lifecycle management is
suitable to utilise big data analytics platforms (Govindarajan, Ferrer, Xu, Nieto, & Lastra, 2016), (S.
Jha, Jha, O'Brien, & Wells, 2014), (Bi & Cochran, 2014).
It has been a long-standing acknowledgement that a poor system upgrade with a new technology can
have far reaching consequences in later stages that are costly to rectify. This continues to be
permeating theme in adoptions of big data analytics platforms in manufacturing systems. Particularly,
these may involve many competing goals, e.g. security, performance, reliability, scalability,
maintainability, and the development cost. There are also unforeseen risks. As articulated by Protiviti:
“Manufacturers should have clear and easily definable goals. As a part of that planning process,
companies need to determine whether the systems they have in place will achieve the desired results
and/or what enhancements might be required” (Protiviti, 2017). Manufacturing also tend to have their
own goals and preferred competitive dimensions with respect to taking advantages of data analytics
platforms to augment their systems (Wang, Xu, Fujita, & Liu, 2016). A system architect, who is
responsible to the design high-level big data enabled solution architecture, should meticulously
specify goals of multiple stakeholders, analyse potential risks, and make a right balance among
operationalisation of the goals in adopting these technologies (Lee, Kao, & Yang, 2014) (Wang et al.,
2016). For example, using a poor big data visualisation technology may negatively affect
performance, scalability, and real-time data processing coming from sensors. The choice of a data
mining algorithm to process sensor data across the product line may also impact the real-time
performance of control systems. An early stage analysis of big data adoption goals gives an
opportunity in exploring countermeasures to tackle probable risks in advance rather than drowning in
narrow aspects of these technologies.
Furthermore, uncertainties about the impact of decisions on data analytics adoption goals are
unavoidable, as in any other adoption endeavors. That is, a lack of complete knowledge about the
actual consequences of architectural decisions is a fact. For instance, the raw feedback data generated
by online customers about produced cars that are processed by data analytics platforms may produce

2
some uncertainties in terms of the interpretation of data. The choice of a data visualisation technique
may have an uncertain impact on the reliability of other generated diagnostic reports about a product
due to inherent uncertainties of data sources. On the other hand, the system architect is still expected
to make right choices in such uncertain circumstances. The quest for a risk-aware early stage analysis
of big data adoption in making critical and uncertain decisions remains a top priority as highlighted in
(Pal, Meher, & Skowron, 2015) and (C. P. Chen & Zhang, 2014).
This paper provides an approach aiding system architects for goal-obstacle analysis of big data
solution architectures and selecting architectural decisions using imperfect information. It provides a
step-by-step goal-obstacle analysis process to address uncertain risks. It then produces a complete set
of requirements, ranks candidate architectures based on the fuzzy logic, and ultimately to find an
optimum architecture. The contributions of the paper are thus two folds: (i) providing a goal-oriented
approach for reasoning about architectural requirements at the early stage of adoption (ii) dealing with
uncertainty about the impact of integrating data analytics platforms on manufacturing systems which
can be unforeseen at the requirement time. It should be noted that the approach is applicable outside
the context of manufacturing. The emphasis of the validation and the exemplars are manufacturing,
however the goal modelling and obstacle resolution approach can be easily applied to other contexts
(Fahmideh & Beydoun, 2018).
The rest of this article is organised as follows: Section 2 provides the background of this study
including a motivating scenario and fundamental concepts used in the proposed approach. Section 3
details the approach. Section 4 illustrates the approach in an exemplar scenario of re-architecting a
legacy car manufacturing system to utilise data analytics platforms. Section 5 outlines other related
studies. Finally, Section 6 summarises the article with a discussion of limitations and future research
directions.

2 Background
2.1 A motivating scenario of moving manufacturing systems to data
analytics platforms
We adopt and extend an exemplar reengineering scenario of a hyper-connected manufacturing
collaboration system (HMCS) (Lin, Harding, & Chen, 2016). HMCS provides a platform to enable
several partners of Toyota car manufacturing to collaborate and to share knowledge about car
products and parts across the product line. At the core of the HMCS architecture, a data Extract-Load-
Transform (ELT) processes and integrates data streams from multiple manufacturing parties and
different types of databases including those storing data coming from sensors in the product line or
from buyers. The process has a middleware layer that includes rules and logic for mapping data
between different formats. An additional data stream comes from the online Toyota buyer
conversations that appear in social media, such as Twitter, posts, Internet server logs, and blogs. This
stream provides feedback on recent purchase experiences, warranty claims, repair orders, service
reports and others which can be used to uncover actionable trends and to generate appropriate early-
warning signals to the manufacturing process. The stream is increasingly voluminous and it generates
millions of items per day.
The heterogeneity of data sources and the large volume of data strain the ETL. It often becomes a
bottleneck and incapable of identifying all patterns and generating statistical reports. To resolve this,
the IT department of Toyota aims to deploy multiple data analytics platforms. The aim is to enable
extraction and management of both sources of data, i.e. internal manufacture parties and the
unstructured data in the online Toyota buyer conversations. A system architect is appointed
specifically to design a solution to upgrade and integrate the ETL with services offered by the data
analytics platforms. An immediate concern of the architect is a cost benefit analysis of the adoption of
the platforms evaluating the risks and mapping the way forward. The following questions are
pertinent to the system architecture: What are ETL system quality goals? How these may be positively
or negatively affected if data analytics platforms are utilised? Will higher ETL performance be
attainable in all circumstances? What obstacles are likely to occur during and after re-engineering

3
ETL to data analytics platforms and what are their severity? What countermeasures can be added in
advanced to negate such obstacles and do they have any side effects? Answering these questions is a
challenging task as it involves reasoning with a long chains of ‘what-if’ scenarios and their uncertain
impacts on goals given by the stakeholders of HMCS. Additionally, these impacts are even imprecise
and may be hard to quantify. Human judgments are often too vague for using exact numerical values,
let alone in an early stage of a solution architecture design.
Towards designing a big data solution architecture, we offer an approach that explicitly relates
manufacturing system high-level quality goals to potential obstacles, highlights architectural
requirements in addressing them and assesses their impact on stakeholders’ goals, calculates the
uncertainty of the various impacts, and finally shortlists candidate architectures satisfying the goals.

2.2 Goal-oriented requirement modeling

Goal-oriented reasoning approaches such as KAOS and i* are means for the elicitation, elaboration,
and analysis of system requirements (Yu & Mylopoulos, 1994). We choose KAOS (Keep All Objects
Satisfied) modelling framework. It defines two components: (i) a modeling language including
concepts such as goal, obstacle, agent, operation, and domain objects and (ii) a method specifying a
series of steps to elaborate and analyse goals (Van Lamsweerde, 2009). In this article, we only use to
concepts goal and obstacles as defined in the following.
A goal is a prescriptive statement of intention that a system should satisfy through the cooperation of
agents. A goal has a name and a specification expressed using natural or formal languages. The
specification defines what the goal means and its satisfaction conditions. Goals may range from high-
level business objectives to fine-grained technical ones. All goals are continuously refined into sub-
goals until all sub-goals can be assigned to a single agent, i.e. a user or a system component. In this
article, we refer to common system quality goals such as performance, security, and maintainability.
The method part of the KAOS framework provides a process to create a goal model through a
hierarchical refinement process.
A goal which is stated without considerations of unexpected conditions in a real environment that may
cause their failures is considered an optimistic goal (Letier, 2001; van Lamsweerde & Letier, 2000).
Taking a more realistic and a deeper look, it is prudent to consider these conditions and to construe
them as obstacles. Obstacles are duals of goals in the sense that as goals represent desired conditions,
obstacles represent undesirable conditions (Letier, 2001) that should be systematically and
concomitantly identified. They need to be assessed and tackled via defining resolution tactics at an
early stage of a system development to identify any needs to modify the goals (Letier, 2001). As
such, in KAOS, resolving obstacles includes steps for generating and selecting alternatives to resolve
obstacles. Selection of a subset of decision alternatives satisfying goals faces a multi-criteria decision
making problem (MCDM) with the possibility of different priority of goals in a view of stakeholders
needs. Any uncertainty in selecting options may also occur in terms of a range of impacts that options
may have on goals. It is often the case that stakeholders may express such impacts qualitatively or
using imprecise measures because their judgements are unavoidably vague and indescribable with
exact numerical values. For example, the impact of choosing a data mining algorithm for processing
data coming from sensors might be expressed in linguistic terms or a range of values rather than a
crisp and single number. Subsequently, each decision alternative may fall within a range. Comparing
two decision alternatives with overlapping in their impact on goals is not easy. There is often a need
to consider trade-offs amongst various alternatives. Whilst MCDM frameworks can help in
comparing, prioritising, and selecting the most suitable resolution tactics, they do not reflect the
uncertainty in human thinking style. Fuzzy logic can better handle the uncertainty. For instance,
expressions such high performance or low cost become usable. We make a synergy between the
KAOS approach and fuzzy logic to cope with various sources of uncertainties in integrating
manufacturing systems with data analytics platforms.

2.3 Fuzzy set theory

Fuzzy set theory analogises human judgment at the presence of proximate information and uncertainty
in decision making (Zadeh, Fu, & Tanaka, 2014). Whilst classic sets define crisp values, fuzzy sets
shows groups of data with boundaries that are not crisp. This provides a better capability to resolve

4
real-world problems, which unavoidably involve imprecise and noisy parameters. Accordingly,
linguistic expression of variables is the central aspect of fuzzy logic where general terms such a very
large, large, medium, small, too small are used to represent a range of numerical values.
The fuzzy set, originally proposed by Zadeh, is defined as follows (Zadeh et al., 2014): In a universe
of discourse U, a fuzzy subset A is characterised by a membership function F where each member of
x ∊ U is associated with a number of F in the internal [0,1], denoting the membership of x in A. The
impact linearly decreases from very low to very high. This range of impact is represented using a
triangular fuzzy value (Pedrycz, 1994). For example, the choice of a particular data mining algorithm
for processing sensor data may have Very High positive impact on the overall system performance.
Such values may be available from statistical data of similar architecture designs in other systems,
system architect’s experience, or expert judgment.
The ability to quantitatively analyse alternatives in a big data solution architecture manufacturing
systems can be achieved via representing uncertain parameters as fuzzy numbers which belong to
fuzzy sets. Instead of showing the anticipated impact of an architectural decision alternative on system
quality goals as a discrete point, we represent such impact as a range of values. This is more aligned
with human judgment in conceptualisation and representation of uncertainty.

3 The approach
Our approach, as shown Figure 1, has two steps: In the first step, high-level goals of adopting data
analytics platforms are identified. Their operational alternatives and potential corresponding obstacles
are then identified and assessed. Obstacles that are deemed to be severe are resolved through
generating resolution tactics. Step 1 engages the stakeholders and iterates over the sequence of
refinements akin to the one described in (Lim & Finkelstein, 2012). In the second step of the
approach, the impact on data analytics adoption goals is analysed to optimise the chances of success
in goal achievement. The overall output of the approach as shown in Figure 1 is a set of solution
architectures, ranked on the basis of likelihood of satisfying specified quality goals. One final selected
architecture from this set gets later incorporated to reengineer the existing manufacturing systems to
data analytics platforms. To illustrate the details the steps, the scenario presented in Section 2.1 is
used as an exemplar in what follows.

Figure 1. Proposed approach

3.1 Step 1. Analysing goal-obstacle of data analytics solution architecture

The first step is based on KAOS modelling framework and uses the notation shown in Table 1. Step 1
includes these two sub-steps:
(i) Step 1.1. Specifying data analytics adoption goals targeting by data analytics platforms
and possibly decision alternatives to operationalise the goals,
(ii) Step 1.2. Analysing data analytics adoption obstacles for identifying goal failure causes,
assessing their likelihood and criticality of consequence, defining resolution tactics and

5
 decision alternatives. This is a mitigation step aiming at reducing the likelihood of the
obstacles occurring or eliminating it altogether.

Table 1. Notations used for the goal modelling

Modelling element Definition Graphical notation
The overall goal of adopting data analytics platforms
Root (overall goal) contributing to systems.

A quality goal that is expected to be satisfied via utilising

Goal
data analytics platforms.
A technical/none-technical exceptional situation/condition
Obstacle preventing the goal satisfaction.

A generic architectural solution either to operationalise a

Architectural goal or to tackle an obstacle.
decision
A technique, tool, or technology taking to operationalise
Architectural
an architectural decision.
decision alternative

Like many IT projects that are inherently conducted cooperatively by a team, the rationale for
embedding modeling within the approach is to broaden stakeholder participation (e.g. system
architect, developers, and users) in the entire goal analysis. This will enable appropriate
documentation and argumentation surrounding goals, obstacles, architectural decision alternatives and
to coordinate the design effort, and to ensure convergence to potential architectures. The following
subsections provide technical details of Step 1.
Step 1.1 Specifying data analytics adoption goals. Seven goals are set for the integrating ETL with
data analytics platforms (Figure 2): g1.Achieve [Processed social media weekly under expected time],
g2.Achieve [Processed sensor data under expected time], g3.Achieve [Improved availability],
g4.Achieve [Maintained interoperability with other big data platforms], g5.Achieve [Improved data
visualisation], g6.Achieve [Maintained data security on big data platforms], and g7.Achieve
[Increased unstructured data storage capacity].

Figure 2. Goals of integrating ETL with data analytics platforms

ETL databases is rapidly growing in size and reaching several terabits of data collected by sensors in
the product line. An unlimited data storage capacity is required. The system architect documents the
specification of goal g7.Achieve [Increased unstructured data storage capacity] as follows:
Goal g7.Achieve [Increased unstructured data storage capacity]
Category scalability goal
Definition Batches of data records from manufacture product line provided by installed sensors
should be captured and stored continuously. These records consist of data monitored by robots
about assembling Toyota parts in the production line.
Quality Variable storageSize: Batch → Size

6
 Definition The required capacity in storing records of data collected by sensors in a working day.
Sample Space The set of daily cars is assembled and delivered to the end of the product line.
Objective Functions At least one gigabyte of records (e.g. images, events, logs, and errors) are
generated at the end of a working day. The database should be able to store this volume.
For the goal g1.Achieve [Processed social media weekly under expected time], the following
definition is documented:
Goal g1.Achieve [Processed social media weekly under expected time]
Category performance goal
Definition Relevant data of buyers experience should be collected from Twitter, posts, Internet
server logs, and blogs and then be processed every week and results be available on Monday at
12am. This data from customers can be in the form of feedback or complaints, maintenance
requests, part orders, or product comparisons.
Quality Variable ProcessedTime: Batch → Time
Definition The required capacity to storage sensor data at the end of the day.
Sample Space The set of daily cars that are assembled on the product line and delivered.
Objective Functions The processing of all the collected data and generating reports should be
started from 12 am on Saturday and finished at midnight on Sunday.
Goals are operationalised through different architectural decisions, i.e. tactics, techniques, and
technologies. As shown in Figure 3, to realise the goal g7.Achieve [Increased unstructured data
storage capacity], the system architect considers five mainstream big data storages namely
a17.MongoDB, a18.Accumulo, a19.HBase, a20.Cloudant, and a21.BigTable. Likewise, to implement
the goal g1.Achieve [Processed social media weekly under expected time], both technologies
d1.scheduler and d2.social media data processing can be used where each has different alternatives to
be employed (Figure 3).

Figure 3. Operationalisation of the goals through architectural decisions and related implementations
Step 1.2. Analysing data analytics adoption obstacles. Normally, goals neglect unexpected
situations that may cause their failures in operational environment (Letier, 2001; van Lamsweerde &
Letier, 2000). As mentioned earlier, these situations are referred to as obstacles. They should be
systematically identified, assessed, and mitigated against. This may lead to goal model elaboration. If
goals are not threatened by any obstacles, the system architect can skip this step and proceed to Step 2
(detailed in Section 3.2). An iterative identify-assess-resolve cycle for the obstacle analysis is required
as follows (steps 1.2.1 to 1.2.3).
Step 1.2.1. Identifying data analytics adoption obstacles. Obstacles may originate from intrinsic
characteristics of data analytics platforms or their operations. The system architect uses domain
information to iteratively refine the goal model identifying obstacles and any sub-obstacles (Letier,

7
2001). In the scenario, the candidate data store technologies for the operationalisation of goal
g7.Achieve [Increased unstructured data storage capacity] may obstruct the goal g4.Achieve
[Maintained interoperability] (Figure 4). The reason is that existing ETL’s databases are relational.
These are not compatible with no-SQL schema-free data storages such as a17.MongoDB,
a18.Accumulo, a19.HBase, a20.Cloudant, and a21.BigTable. Converting thousands of line of
complex T-SQL codes defined in ETL to this type of big data storages is not a simple task. In other
words, the alternative technologies raise the obstacle o1.Incompatibility of ETL and big data storages.
This obstacle is further decomposed into sub-obstacles o1.1.Incompatible datatypes,
o.1.2.Incompatible data operations, and o.1.3.Incompatible APIs. Data analytics platforms leverage
cloud computing servers which are often vulnerable to issues such as bandwidth capacity bottleneck,
performance variability or scaling latency, and security (Agrawal, Das, & El Abbadi, 2011). Given
that, the partial goal model in Figure 5 represents probable obstacles against the goals that are
identified by the system architect.

Figure 4. Obstacles to goal Achieve [Maintained interoperability with other big data platforms] in the case of
using big data store technologies
Step 1.2.2. Assessing data analytics adoption obstacles. The identified obstacles from Step 1.2.1 are
assessed to generate a new set of architectural requirements. The criticality of the obstacles is judged
based on their impact on the goals. Qualitative and quantitative techniques can be employed to
perform this step. However, our approach employs a common qualitative technique, Risk Analysis
Matrix (Franklin, 1996). This technique specifies the likelihood of an obstacle using a qualitative
scale ranging from Almost Certain, Likely, Possible, Unlikely, and Rare. It also indicates the obstacle
consequence as Insignificant, Minor, Moderate, Major, and Catastrophic. The risk of an obstacle is
defined as the product of its occurrence and severity, i.e. Risk = Likelihood × Consequences.
Estimating this risk relies on the availability of domain information sources such as statistics from
manufacturing systems, existing accounts on data analytics platforms, or the system architect’s
judgement. The system architect may conduct a voting technique involving all stakeholders to assess
the probability occurrence and severity of obstacles. A risk matrix highlights the risk zone as shown in
Table 2. For instance, the risk of an obstacle might be considered as moderate (M), however, it is still
tolerable. Whilst a High and Extreme obstacle may necessitate a countermeasure. The values
represented in Table 2 are exemplar values.

8
 Figure 5. Identified obstacles to goals

Table 2. Risk matrix for obstacle assessment

Consequence severity
Likelihood Insignificant Minor Moderate Major Catastrophic
Almost Certain H H E E V
Likely M H H E V
Possible L M H E E
Unlikely L L M H E
Rare L L M H H
V: Very extreme risk, E: Extreme risk; H: High risk; M: Moderate risk; L: Low risk
Step 1.2.3 Resolving data analytics adoption obstacles. Obstacles deemed with severe risk should
be resolved. This requires generating new architectural decision alternatives and selecting suitable
alternatives amongst them. We employ eight generic and platform independent KAOS’s obstacle
resolution tactics: goal substitution, agent substitution, obstacle prevention, goal weakening, obstacle
reduction, goal restoration, obstacle mitigation, and do-nothing (Letier & Van Lamsweerde, 2004;
van Lamsweerde & Letier, 2000). These are operators on a goal model to refine it to new or existing
modified goals, assumptions, and responsibility assignments. They are defined as follows.
(i) Substitute goal defines a new alternative goal which is still contributable by data analytics
platforms in a way that the obstacle is no longer present. Consider the performance goal g1.Achieve
[Processed social media weekly under expected time] obstructed by the obstacle Social media size
exceeds processing speed time. An instance of accommodating this tactic is to collect and process data
daily instead of weekly based.
(ii) Substitute data analytics platform removes the occurrence of an obstacle by replacing the
responsibility for an obstructed goal to a new platform. For example, the obstacle o5.Sensor data
processing exceeds expected time can be removed via transferring the assigned goal from an
overloaded server to another server with lower workload.
(iii) Prevent obstacle introduces new assertions to the goal model preventing the obstacle occurrence
via applying some factors or doing things in particular way. For instance, consider the security
obstacle o8.Malicious attack by tenants obstructing the goal g6.Achieve [Maintained security of
sensor data] (Figure 6). An application of this tactic is to encrypt the batch data collected from
sensors prior storing them on big data storages. As such, batch data cannot be read or processed by

9
malicious tenants that are in performing on the same cloud servers. The system architect considers
three architectural decision alternatives o31.Obfuscate data, o30.Redact data, and o32.Mask data.
Furthermore, to prevent the occurrence of obstacles o1.1.Incompatible datatypes and
o.1.2.Incompatible data operations, architecture decisions a25.Adapt data and a26.Develop adaptor
are considered. It should be noted that employing architecture decision alternatives may cause another
set of obstacles against goals. For instance, on the one hand the system architect considers a25.Adapt
data and a26.Develop adaptor. On the other hand, these alternatives may negatively influence the
goal g2.Achieve [Processed sensor data under expected time]. These dependencies are modelled in
Step 2 of the approach described in Section 3.2.
(iv) Reduce obstacle introduces agents such as human or servers to behave in certain ways to lessen
the occurrence likelihood of an obstacle. Consider the obstacle o5.Sensor data exceeds expected
processing time to the goal g2.Achieve [Processed sensor data under expected time]. An example of
applying this tactic is to reduce server workload by prioritising upcoming data that are sent by sensors
installed in the product line. The data from highly important sensors that are collected and processed
take precedence over those sensors providing supplementary data or do not need a real-time
processing. In addition, to reduce the likelihood occurrence of the root obstacle o4.Performance
variability of big data platform, the architectural decision is a23.Refine network topology. Finally, as
mentioned earlier, data analytics platforms may be vulnerable to issues such as server latency as
represented by the obstacle o3.Big data analytic platform latency (Figure 6). To reduce this, the
architectural decision a22.Acquire more resources (e.g. virtual machines) is chosen.
(v) Weaken goal suggests degrading the goal definition to make it more liberal and relaxed in a way
that the obstruction no longer occurs. This can be applied in two ways:
-relaxing assumptions of an obstructed goal so that its original form does not needs to be
satisfied in all situations. The goal g2.Achieve [Processed sensor data under expected time]
obstructed by o5.Sensor data size exceeds expected processing time is modified to one that the
goal is not required to be satisfied in all situations, particularly when data analytics server is not
in its fully capacity.
-relaxing the required level of goal satisfaction condition, meaning that there is no further need
to full satisfaction. One example of applying this tactic to goal g2.Achieve [Reduced processing
social media weekly under expected time] is to soften its definition to maximum acceptable time
to be achieved by increasing the time/date, i.e. quality variable processedTime.
(vi) Restore goal and mitigate obstacle are two tactics usable when the avoidance of all obstacles is
too costly and tolerating or mitigating consequences of obstacles becomes more practical. In the goal
restoration tactic, the system architect adds a new restoration goal that prescribes restoration
mechanisms for situations when the obstacle impedes the goal. For the obstacle mitigation, the system
architect adds a new goal to attenuate the consequences of the obstacle actually occurring. The tactic
intends to achieve a weaker satisfaction of an obstructed goal. In the discussed scenario, the goal
g3.Achieve [Improved availability] is obstructed by the obstacle o6.Cloud transient fault as data
analytics platforms leveraging cloud might be temporarily unavailable due to reasons such as network
traffic or server workload. An implementation technique to mitigate this obstacle is a27.Retry
connection in the system architecture assuring a weaker version of the goal by specifying next
retrying to connect to the server when transient faults occur. Figure 6 shows the goal model after
introducing resolution tactics.
(vii) Do nothing accepts the risk of an obstacle occurrence.

10
 Figure 6. Decision alternatives for handling obstacles
The generated architectural decision alternatives either operationalise goals or tackle obstacles. They
form a solution space of different architectures that can be used to integrate ETL with data analytics
platforms. Selecting a suitable solution architecture is a challenging task which is systematically dealt
in Step 2 of the proposed approach.
As the last note for this step, we believe the scale of an architecture analysis scenario determines
whether adopting normative models, like the one presented in this research, necessitates. While an ad-
hoc approach for requirement analysis and possible solution architecture is applicable for small-scale
projects with a limited number of goals/risks and stakeholder participant, a systematic and
communicative presentation layers for specifying notations and model refinements is useful for large-
scale projects with multiple goals, potential obstacles, and possible resolution tactics.

3.2 Step 2. Exploring uncertainties in big data solution architecture

This step explores candidate solution architecture. The variables that are used in this step presented in
Table 3.
Table 3. Symbols used in exploring uncertainty for step 2
Symbols Definition
g A goal
G Set of goals
a An alternative to operationalise/implement an architectural decision
A Set of implementation alternatives to operationalise an architectural decision
d A architectural decision
D Set of architectural decisions
If an decision alternative is selected then
𝑥𝑎
𝑥𝑎 is 1, otherwise it is 0
𝑎𝑟𝑐ℎ A candidate solution architecture including its architectural decision alternatives
AS All possible solution architectures based on all architectural decision alternatives
𝑐𝑜𝑛
̃ 𝑔,𝑎 Contribution (positive/negative) of an alternative a on a goal g

11
 ̃
𝑆𝑔 (𝑎𝑟𝑐ℎ) The total value of a solution architecture with respect to a specific goal g
𝑃𝑔 The weight of the goal g from stakeholders’ point of view
𝑇ℎ𝑑𝑔 A threshold to goal g
𝑇ℎ𝑑𝑐 A constraint for the cost of solution architecture
𝐺𝑚𝑎𝑥 A goal g which is expected to be maximised
𝐺𝑚𝑖𝑛 A goal g which is expected to be minimised
̃𝑎
𝑐𝑜𝑠𝑡 Fuzzy cost of the alternative a
𝐶𝐻𝑘 (𝑆̃ (𝑎𝑟𝑐ℎ𝑖 )) Ranking index of solution architecture i with total value 𝑆̃
Step 2.1. Representing uncertain impact of decision alternatives on goals. We define variable D as
a set of architectural decisions. Recall from Step 3.1, the choice for the operationalisation of goal
g7.Achieve [Increased unstructured data storage capacity] through different big data stores is an
example of such a decision (Figure 6). Each architectural decision d ∈ D, itself, may have alternatives
for operationalisation, which is specified using set 𝐴𝑑 . For instance, the decision on big data storage
has five alternatives, namely: a17.MongoDB, a18.Accumulo, a19.HBase, a20.Cloudant, and
a21.BigTable (Figure 6). As mentioned in Step 3.1, an architectural decision d and its implementation
alternatives 𝐴𝑑 can be derived using resolution tactics. In the scenario, the architectural decision
prevent obstacle is used to handle the obstacle o8.Malicious attack by tenants. The system architect
assumes three different implementation alternatives a30.Redact data, a31.Obfuscate, and a32.Mask
data for this architectural decision. This set of implementation alternatives for the decision prevent
obstacle is defined as A = ∪𝑑∈𝐷 𝐴𝑑 . The solution space (SS), is a set of all possible alternatives of
architectural decisions and their associated implementation techniques. Thus, SA is represented as
follows:
𝑆𝑆 ≝ {𝑎𝑟𝑐ℎ ⊆ 𝐴|(∀ 𝑑 ∊ 𝐷 ∶ ∃ 𝑎 ∊ 𝐴𝑑 : 𝑎 ∊ 𝑎𝑟𝑐ℎ ) ^ (∀ 𝑎 ∊ 𝑎𝑟𝑐ℎ, 𝑎 ∊ 𝐴𝑑 ∶ ∄ 𝑏 ∊ 𝐴𝑑 ∶ 𝑏 ≠ 𝑎 ^ 𝑏 ∊ 𝑎𝑟𝑐ℎ)}

Therefore, with respect to Figure 6, the size of the solution space in the current scenario is
5*5*4*4*3*3*1*1*3 = 10800 potential alternative architectural solutions.
For an implementation alternative a ∈ A and goal g ∈ G, we define 𝑐𝑜𝑛 ̃ 𝑔,𝑎 which specifies the
contribution/impact of the alternative a on the goal g. The symbol ~ indicates that the contribution is
a fuzzy number. To represent a fuzzy impact, our approaches uses Triangular fuzzy numbers (TFNs)
(Pedrycz, 1994). TFNs are widely used to represent the approximate value range of linguistic
variables. A triangular fuzzy number is represented by A = (w, y, z) where the parameters w, y, and z,
respectively, show the smallest possible value, the most promising value, and the largest possible
value describing a fuzzy event. TFN linear membership function 𝜇A is defined by:
𝑥−𝑤
𝑤≤𝑥≤𝑦
y−w
𝜇𝐴 (𝑥) = z − 𝑥
𝑦≤𝑥≤𝑧
z−y
{ 0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

We divide the goal satisfaction into five levels: Very low (VL), Low (L), Medium (M), High (H), and
Very high (VH) as shown in Table 4. The numerical range for the goal value and the membership
function are respectively presented in Table 5 and Figure 7.

12
 Table 4 Linguistic variables used to show the impact of operationalisation alternatives on quality goals
Level Satisfaction value
Very low (VL) 0 and less than 1
Low (L) Between 0 and 2
Medium (M) Between 1 and 3
High (H) Between 2 and 4
Very high (VH) 3 and more than 3

Figure 7. Impact of an implementation alternative on a goal

Table 5. Triangular membership functions for the linguistic variables

𝑥−𝑤
3≤𝑥≤4 𝑥−𝑤
y−w 2≤𝑥≤3
𝑉𝐻𝐴 (𝑥) = { y−w
1 4≤𝑥
𝐻𝐴 (𝑥) = z − 𝑥
0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 3≤𝑥≤4
z−y
{ 0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
𝑥−𝑤 𝑥−𝑤
0≤𝑥≤1
1≤𝑥≤2 y−w
y−w z−𝑥
𝐿𝐴 (𝑥) = { 1≤𝑥≤2
𝑀𝐴 (𝑥) = 𝑧 − 𝑥 z−y
2≤𝑥≤3
z−y 0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
{ 0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
𝑥−𝑤
0≤𝑥≤1
y−w
𝑉𝐿𝐴 (𝑥) = {
0 𝑥≤0
0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

For example, given the impact of implementation alternative a8.Python NLTK, i.e. a8, on the goal
g2.Achieve [Processed sensor data under expected time], i.e. g2, is crisp number 3, the fuzzy
representation for the goal satisfaction is:
𝑐𝑜𝑛𝐴𝑐ℎ𝑖𝑒𝑣𝑒 [𝑃𝑟𝑜𝑐𝑒𝑠𝑠𝑒𝑑 𝑠𝑒𝑛𝑠𝑜𝑟 𝑑𝑎𝑡𝑎
̃ 𝑢𝑛𝑑𝑒𝑟 𝑒𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑡𝑖𝑚𝑒],𝑃𝑦𝑡ℎ𝑜𝑛 𝑁𝐿𝑇𝐾 = (0.1,0.25,0.25,0,0.4)

Step 2.2 Calculating solution architecture value. This is determined at two levels. Firstly, the fuzzy
aggregation of selected implementation alternatives’ contributions to a specific goal g in a given
architecture arch ∈ SS is defined through equation (i):
̃
𝑆𝑔 (𝑎𝑟𝑐ℎ) = ∑𝑎 ∈𝑎𝑟𝑐ℎ(𝑐𝑜𝑛
̃ 𝑔,𝑎 𝑥𝑎 ) (i)

Note that in (i), for each implementation alternative a ∈ A, we consider a binary decision variable 𝑥𝑎 ,
indicating whether an alternative is chosen, i.e. 𝑥𝑎 =1, or not chosen, i.e. 𝑥𝑎 =0. To calculate equation

13
(i), we utilise Mamdani fuzzy inference technique (Mamdani, 1974). For example, given the impact of
selected 32 operationalisation alternatives on g2.Achieve [Processed sensor data under expected
time], the value of a solution architecture in terms of g2 is computed using the fuzzy rules presented
in Table 6.

Table 6. An excerpt of fuzzy rules for determining the obtained value for a goal
g2.Achieve [Processed sensor data under expected time]in a given solution
architecture
a1 a2 a3 a4 … a32 𝑆̃
𝑔2

If VH and VH and VH and and … H then H

If VH and VH and VH and and … VH then VH
If VH and VH and VH and and … VH then VH
If VH and VH and VH and and … VH then VH
If VH and VH and VH and and … VH then VH
… … … … … … … … … … … … …
The same fuzzy rules are applied for other goals g2 to g7. Next, the total value of a given solution
architecture arch ∈ SS is the aggregation of attained fuzzy values for all goals. This is defined via
equation (ii):
𝑆̃ (𝑎𝑟𝑐ℎ)= ∑𝑔 ∈𝐺 (𝑃𝑔𝑆̃𝑔(𝑎𝑟𝑐ℎ) ) (ii)
Again fuzzy rules are defined to determine the total value of a solution architecture based on the fuzzy
values of goals as shown in Table 7.
Table 7. An excerpt of fuzzy rules for determining the value of a given solution architecture
g1 g2 g3 g4 g5 g6 g7 𝑆̃ (𝑎𝑟𝑐ℎ)
If VH and VH and VH and VH and VH and VH and H then H
If VH and VH and VH and VH and VH and VH and VH then VH
If VH and VH and VH and VH and VH and VH and VH then VH
If VH and VH and VH and VH and VH and VH and VH then VH
If VH and VH and VH and VH and VH and VH and VH then VH
… … … … … … … … … … … … … … … …
The same fuzzy rules are applied for other candidate solution architecture. Our aim is to find an arch
∈ SS with highest value of 𝑆̃(𝑎𝑟𝑐ℎ).
Note that in tables 6 and 7, the number of fuzzy rules can be dramatically increased if the there are
many goals, architectural decisions, and operationalisation alternatives. Given the fact that
stakeholders may have different emphasises on the quality goals, some fuzzy rules can be removed
from tables 6 and 7. In doing so, for each goal g ∈ G, we assign a numeric value between 𝑃𝑔 ∈ [1..10]
that shows the degree of priority of the goal in view of stakeholders. As such, for goals with low
priority, there is no need to write fuzzy rules.
Step 2.3. Specifying solution architecture constraints. A goal may have a certain constraint that has
to be satisfied, e.g. the constraint for g2.Achieve [Processed sensor data under expected time] is
expected to be less than 40 millisecond. A constraint for a goal g, represented by 𝐶𝑟𝑡𝑔 , defined
through equation (iii):
∀𝑔 ∈ 𝐺𝑚𝑎𝑥 ∶ 𝐶𝑟𝑡𝑔 ≤ 𝑆 ̃𝑔 (𝑎𝑟𝑐ℎ) (iii)
̃
∀𝑔 ∈ 𝐺𝑚𝑖𝑛 ∶ 𝑆𝑔 (𝑎𝑟𝑐ℎ) ≤ 𝐶𝑟𝑡𝑔
In accepting/rejecting a solution architecture, its cost is also an important constraint, which is shown
using 𝐶𝑟𝑡𝑐 and defined using equation (iv):
̃𝑎 𝑥𝑎 ) ≤ 𝐶𝑟𝑡𝑐 (iv)
∑𝑎 ∈𝑎𝑟𝑐ℎ(𝑐𝑜𝑠𝑡
̃𝑎 shows the fuzzy cost of the decision alternative a. Constraints on goals and cost are defined
𝑐𝑜𝑠𝑡
regarding project context in which they are applied.

14
Step 2.4. Comparing solution architectures. Given the obtained values for S̃ (arch), we can
compare solution architectures. Between two architectures arch1 and 𝑎𝑟𝑐ℎ2 ∈ SS , 𝑎𝑟𝑐ℎ1 is more
desirable if: 𝑆̃ (𝑎𝑟𝑐ℎ1 ) ≤ 𝑆̃ (𝑎𝑟𝑐ℎ2 ) .This is a fuzzy comparison of two fuzzy ranges of possible
values for two solution architectures resulting in the one with a better range. For this, we employ
Chen’s method (S.-H. Chen, 1985) where it defines the concepts of fuzzy maximising and minimising
sets expressed using equations (v) and (vi):
𝑆𝑚𝑎𝑥 (𝑥)= (𝑥 − 𝑥𝑚𝑖𝑛 /𝑥𝑚𝑎𝑥 − 𝑥𝑚𝑖𝑛 )𝑘 (v)
𝑆𝑚𝑖𝑛 (𝑥)= (𝑥𝑚𝑎𝑥 − 𝑥/𝑥𝑚𝑎𝑥 − 𝑥𝑚𝑖𝑛 )𝑘 (vi)
In equations (v) and (vi), 𝑆𝑚𝑎𝑥 = sup ⋃𝑛𝑖=1 𝑠𝑢𝑝 𝑆𝑖 and 𝑆𝑚𝑖𝑛 = inf ⋃𝑛𝑖=1 𝑠𝑢𝑝 𝑆𝑖 and k > 0 are real
numbers. Using these two sets, left and right utility of a fuzzy number 𝑆𝑖 , 𝑖 𝑡ℎ architecture solution, is
defined as:
L (𝑆̃ (𝑎𝑟𝑐ℎ𝑖 )= sup min (𝑆𝑚𝑖𝑛 (𝑥), 𝑆̃ (𝑎𝑟𝑐ℎ𝑖 )) x ∈ ℝ
R (𝑆̃ (𝑎𝑟𝑐ℎ𝑖 )) = sup min (𝑆𝑚𝑎𝑥 (𝑥), 𝑆̃ (𝑎𝑟𝑐ℎ𝑖 )) x ∈ ℝ
Given that, the ranking index for 𝑖 𝑡ℎ solution architecture is obtained using the equation (vii):
𝐶𝐻 𝑘 (𝑆̃ (𝑎𝑟𝑐ℎ𝑖 )) = ½ (R (𝑆̃ (𝑎𝑟𝑐ℎ𝑖 )) + 1 - L (𝑆̃ (𝑎𝑟𝑐ℎ𝑖 )) (vii)
Step 2.5. Finding the optimum solution architecture. Finding an architecture optimising quality
goals is a typical multi-objective optimisation problem under some constraints (H.-J. Zimmermann,
1978). A solution architecture is considered optimal if it maximises quality goals and satisfies
imposed constraints which is, in fact, defined as a linear programming problem through equation
(viii):

Maximize 𝑆̃ (𝑎𝑟𝑐ℎ) subject to the constraint equations (iii) and (iv) (viii)

(viii) maximises the cumulative value of architectures by selecting a combination of alternatives

resulting in the highest architecture value under equations (iii) and (iv) to avoid constraint violation.

4 Application exemplar
The scenario of integrating ETL with data analytics platforms (Section 2.1) is used to examine the
proposed approach. Tables 8 shows the goals for reengineering that are elaborated into architectural
decisions, operationalisation alternatives. They collectively form a space of possible solution
architectures for exploration.

15
 Table 8. Goals, architectural decisions, and operationalisation alternatives
Goal Architectural decision Operationalisation alternative
d0. Substitute goal Not applicable (the goal definition is refined)
d1. Social media data processing a8.Python NLTK
a6.Gate
a7.Lexalytics Sentiment Toolkit
g1. Achieve [Processed social media weekly under expected time] a5.AeroText
d2.Scheduler a1.Fair scheduler
a2.Capacity scheduler
a3.Delay scheduler
a4.Matchmaking scheduler
a9.SQLStream
d3. Real-time stream processing a10.Storm
a11.StreamCloud
d6. Reduce obstacle data analytic platform latency a22.Acquire more resources
g2. Achieve [Processed sensor data under expected time]
d7. Reduce obstacle Performance variability of data analytics a23.Refine network topology
platform
d8. Substitute data analytics platform Not applicable
d9. Reduce obstacle sensor data exceeds expected processing time a24.Prioritizing sensor data processing
d10. Restore goal for obstacle cloud transient fault a27.Retry connection
d11. Restore goal a29.Eventual Consistency
g3. Achieve [Improved availability]
a28.Weak Consistency
a30.Timeline Consistency
d12. Prevent obstacle a25.Adapt data
g4. Achieve [Maintained interoperability with other big data platforms]
d13.Prevent obstacle a26.Develop adaptor
a16.Google chart
a15.Tableau
g5. Achieve [Improved data visualisation] d4. Data visualisation a14.Data-driven
a13.document
a12.Fusion chart
g6. Achieve [Maintained data security on big data platform] d14.Prevent obstacle a30.Redact data
a32.Mask data
a31.Obfuscate data
a17.MongoDB
a18.Accumulo
g7. Achieve [Increased unstructured data storage capacity] d5. Big data storage a19.HBase
a20.Cloudant
a21.BigTable

16
Through collaboration with the stakeholders, the system architect specifies the impact of
operationalisation alternatives on the goals. She used linguistic variables shown in Table 4 and
triangular fuzzy membership functions defined in Step 2 (Section 3.2). The complexity of selecting
decision alternatives to generate a proper solution architecture in the goal model (Figure 6) is revealed
when the system architect is faced with a large number of goals and decision alternatives. The
scenario follows 7 goals that are expected to be satisfied. Table 9 shows 32 different
operationalisation alternatives in total where for each decision only one alternative can be selected.
The numbers in Table 9 are a part of the exemplar of stakeholders reflecting on the impact of decision
alternatives on the achievement of the goals. The exemplar is elaborated from (Lin et al., 2016). As
mentioned earlier, there are 10800 possible architectural solutions, each of which represents a trade-
off among the goals. The numbers are in essence subjective quality measures of the various
architectural decisions and are used to illustrate the overall process.
Predicting the impact of potential solution architectures on the goals and finding which portion of the
solution space is valid can be a challenging exercise, in particular at the early stage of the data
analytics enablement process where the real impact of decision alternatives on quality goals is
uncertain. Using the second step of the approach (Section 3.2), the system architect can explore the
solution space to find a new suitable architecture for ETL.

17
 Table 9. Impact of operationalisation alternatives on system quality goals (NA: Not applicable)
a) The fuzzy expression of decision b) Simple crisp expression of decision
alternatives’ impact on goals alternatives’ impact on goals (values
between 1 and 5)
Goal Goal
Operationalisation g1 g2 g3 g4 g5 g6 g1 g2 g3 g4 g5 g6
Decision
alternative
d0.Substiute goal NA NA NA NA NA NA NA NA NA NA NA NA NA
a8.Python NLTK L VL H L L H 2.7 1.8 4.3 2.4 2.2 4
a6.Gate M VL M M M VH 3.5 1.3 3.7 3.9 3.6 5
d1. Social media data processing a7.Lexalytics Sentiment
M M L M L L 3.3 3.8 2.8 3.2 2.8 2
Toolkit
a5.AeroText VH L H L H L 5.9 2.4 4.4 2.1 4.1 2
a1.Fair scheduler H M VH M VH L 4.2 3.9 5 3.6 5 2
d2.Scheduler a2.Capacity scheduler VL VL M M L L 1.3 1.2 3.2 3.4 2.8 2
a3.Delay scheduler M VL M M VH M 3.6 1.8 3.7 3.9 5 3
a4.Matchmaking scheduler M H H M H L 3.2 4.3 4.3 3.3 4.3 2
a9.SQLStream VL H M H L VH 1.8 4.6 3.9 4.1 2.1 5
d3.Real-time stream processing a10.Storm M VL M H L VL 3.7 1.3 3.7 4.3 2.5 1
a11.StreamCloud M M VH L M L 3.3 3.6 5 2.7 3.2 2
a16.Google chart L M H M VH VL 2.2 3.1 4.2 3.6 5.2 1.7
a15.Tableau L M H VH M VL 2.3 3.7 4.6 5 3.2 1.5
d4. Data visualisation a14.Data-driven M M M VH M VL 3.2 3.8 3.7 5 3.3 1.9
a13.Document H M M M H H 4.1 3.1 3.2 3.6 4.5 4.9
a12.Fusion chart VL H M M L VH 1.7 4.3 3.8 4.3 2.4 5
a17.MongoDB L VL M H VH VH 2.1 1.7 3.3 4.5 5.4 5
a18.Accumulo L L M M VH H 2.6 2.8 3.3 3.3 5 4.2
d5. Big data store a19.HBase H L H M VH VH 5 2.6 4.3 3.2 5 5
a20.Cloudant M H M VL M VL 3.9 4.6 3.7 1.6 3.2 1.9
a21.BigTable M M M H H VH 3.4 3.7 3.6 4.4 4.7 5
d6. Reduce obstacle big data analytic platform a22.Acquire more resources
M M M M VH H 3.8 3.8 3.2 3.1 5 4
latency
d7. Reduce obstacle performance variability of big a23.Refine network topology
VL H M M L VH 1.7 4.3 3.8 4.3 2.4 5
data platform
d8 – d10 NA NA NA NA NA NA NA NA NA NA NA NA NA
d11. Restore goal a29.Eventual consistency VH M L H VH H 5 3.3 1.7 4.7 5 4

18
 a28.Weak consistency VL M M M M M 1.3 3.6 3.9 3.9 3.2 3
a30.Timeline consistency M VH M M VH H 3.8 5 3.2 3.4 5 4
d12. Prevent obstacle a25.Adapt data VL M M H H M 1.4 3.9 3.5 4.5 4.8 3
d13. Prevent obstacle a26.Develop adaptor M L VL M VL M 3.6 2.2 1.6 3.6 1.4 3
a30.Redact data M VL L M M M 3.2 1.8 2.4 3.9 3.6 3
a32.Mask data M M M M VH H 3.8 3.8 3.2 3.1 5 4
d14. Prevent obstacle
a31.Obfuscate data VL H L M M H 1.1 4.2 2.8 3.9 3.6 4

19
Given that the goal weights are assumed to be equal by the stakeholders, the value of each solution
architecture is calculated using equation (ii) and fuzzy rules in Table 6. One of the constraints
imposed by the stakeholders was to keep the cost of a solution architecture implementation under
$30000, which is represented using equation (iv). This constraint ruled out 2531 solution architectures
out of 10800. Thus, 8269 solution architectures were left. Following further discussions with the
stakeholders, it was agreed to relax the cost constraint to $36000. Subsequently, this reduced the
number of rejected architectures to get a better chance for solution exploration. In other words, 564
solution architectures were added, i.e. 8833 solutions in total. As mentioned in Section 3.2, the second
constraint imposed by ETL users was to keep the data stream processing coming from sensors below
40 milliseconds. This allows the system architect to assess the system performance constraint on the
choice of solution architectures. Relaxing this constraint to 47 milliseconds yielded in increasing the
number of acceptable candidate architectures. With this change, 185 solution architectures were
further added to the solution space. That is, 9018 valid solutions remained for further analysis. Table
10 shows the top 10 solution architectures ranked using equation (vii) and the selected architectural
decision alternatives for all those. Recall from Step 2.4 (Section 3.2), among two solution
architectures the one is better if it has a greater fuzzy value. The best solution architecture ranked as
the first one has the best combination of decision alternatives in view of trade-off among goals
compared to the majority of candidates. That is, it has the best combination of fuzzy values.
Nevertheless, it is still likely that the architect may select a solution architecture that is slightly worse
than the optimal one due to some reasons (e.g. preference to a particular data analytics platform in the
marketplace. Figure 8 shows the final goal model based on the selected decision alternatives for the
first ranked solution architecture.

Figure 8. The first solution architecture including the best combination of implementation alternatives

20
 Table 10. Ranked solution architectures with respect to the decision alternatives based on fuzzy-logic and crisp approaches
(a) Fuzzy approach (b) Crisp approach
Decision alternative Decision alternative

d12. Prevent obstacle

d13. Prevent obstacle

d14. Prevent obstacle

d12. Prevent obstacle

d13. Prevent obstacle

d14. Prevent obstacle

d4.Data visualisation

d4.Data visualisation
d1.Social media data

d1.Social media data

d3.Real-time stream

d3.Real-time stream
d6. Reduce obstacle

d7. Reduce obstacle

d6. Reduce obstacle

d7. Reduce obstacle

d11. Restore goal

d11. Restore goal

d2. Scheduler
d5.Data store

d5.Data store
d2.Scheduler
processing

processing

processing

processing
d8-d10

d8-d10
Rank

Rank
Lexalytics Fair SQLStream Data- Mongo Acquire Refine NA Eventual Adapt Develop Redact Gate Fair Storm Googl Accumu Acquire Refine NA Timeline Adapt Develop Obfuscate data
1 Sentiment scheduler driven DB more network Consistency data adaptor data 1 scheduler e chart lo more network Consistency data adaptor
Toolkit resources topology resources topology
Gate Capacity StreamCloud Docume Clouda Acquire Refine NA Timeline Adapt Develop Mask data AeroText Capacity StreamC Docu BigTabl Acquire Refine NA Eventual Adapt Develop Redact data
2 scheduler nt nt more network Consistency data adaptor 2 scheduler loud ment e more network Consistency data adaptor
resources topology resources topology
Gate Capacity StreamCloud Fusion Clouda Acquire Refine NA Timeline Adapt Develop Obfuscate Python Matchmakin SQLStre Table Cloudan Acquire Refine NA Timeline Adapt Develop Mask data
3 scheduler chart nt more network Consistency data adaptor data 3 NLTK g scheduler am au t more network Consistency data adaptor
resources topology resources topology
Lexalytics Matchmak SQLStream Data- Accum Acquire Refine NA Weak Adapt Develop Redact Gate Fair StreamC Data- Mongo Acquire Refine NA Weak Adapt Develop Mask data
4 Sentiment ing driven ulo more network Consistency data adaptor data 4 scheduler loud driven DB more network Consistency data adaptor
Toolkit scheduler resources topology resources topology
Lexalytics Fair Storm Fusion Tableau Acquire Refine NA Eventual Adapt Develop Mask data Lexalytics Fair SQLStre Googl BigTabl Acquire Refine NA Eventual Adapt Develop Obfuscate data
5 Sentiment scheduler chart more network Consistency data adaptor 5 Sentiment scheduler am e chart e more network Consistency data adaptor
Toolkit resources topology Toolkit resources topology
Gate Delay Storm Data- Clouda Acquire Refine NA Timeline Adapt Develop Obfuscate Gate Capacity Storm Fusio Cloudan Acquire Refine NA Weak Adapt Develop Redact data
6 scheduler driven nt more network Consistency data adaptor data 6 scheduler n t more network Consistency data adaptor
resources topology chart resources topology
AeroText Matchmak StreamCloud Python HBase Acquire Refine NA Weak Adapt Develop Redact Gate Delay StreamC Docu HBase Acquire Refine NA Weak Adapt Develop Mask data
7 ing NLTK more network Consistency data adaptor data 7 scheduler loud ment more network Consistency data adaptor
scheduler resources topology resources topology
Lexalytics Capacity Storm Docume Google Acquire Refine NA Weak Adapt Develop Obfuscate AeroText Capacity SQLStre Table Cloudan Acquire Refine NA Timeline Adapt Develop Mask data
8 Sentiment scheduler nt chart more network Consistency data adaptor data 8 scheduler am au t more network Consistency data adaptor
Toolkit resources topology resources topology
Gate Fair SQLStream Fusion Mongo Acquire Refine NA Timeline Adapt Develop Redact Lexalytics Matchmakin StreamC Fusio Cloudan Acquire Refine NA Timeline Adapt Develop Obfuscate data
9 scheduler chart DB more network Consistency data adaptor data 9 Sentiment g scheduler loud n t more network Consistency data adaptor
resources topology Toolkit chart resources topology
10 AeroText Delay Storm Python BigTab Acquire Refine NA Eventual Adapt Develop Obfuscate 10 Lexalytics Delay SQLStre Data- Mongo Acquire Refine NA Eventual Adapt Develop Obfuscate data
scheduler NLTK le more network Consistency data adaptor data Sentiment scheduler am driven DB more network Consistency data adaptor
resources topology Toolkit resources topology

21
Interestingly, the system architect also compared the results generated through our approach and the
simple crisp approach (Table 9-b). Herein, the simple crisp approach refers to an approach that
ignores the uncertainty in the impact of decision alternatives on quality goals. Stakeholders used crisp
values to represent the impact of decision alternatives on the goals. This difference highlights the
contribution of our approach compared to the crisp one in the selection of a proper solution
architecture. The simple crisp approach for the calculation the value of a candidate solution
architecture would select 125th solution architecture as the optimal solution. This is in contrast to our
approach in which 125th approach is ranked as 46th suitable candidate architecture. In other words,
125th has a large negative consequence of uncertainty, which is ignored by the crisp approach. The
difference between 1th and 125th solution architectures can be also recognized through specific
decision alternatives selected for each architecture. Table 10 represents the selected decision
alternatives for the first top ten solution architectures based on two approaches. For example,
regarding the information in this right and left sides of the table to find selected decision alternatives,
it is observable that our approach chose a17.MongoDB for d5.Big data store for the first ranked
solution architecture whilst the crisp approach chose a18.Accumulo.

5 Related work
There is a paucity of research focus on the early goal-obstacle analysis and architecture decisions in
the scope of integrating manufacturing systems with data analytics platforms. The literature most
related is thus subsumed under three research streams: (i) traditional system re-engineering, (ii)
reengineering to cloud platforms, and (iii) reengineering to data analytics platforms. Hence, we
discuss how our approach presented is positioned in relation to notable research in each stream.

5.1 Traditional approaches for legacy system reengineering

Early decision making on selecting solution architectural under uncertainty have been already
discussed in the Introduction section. One of the earliest work is by Svahnberg et al. where they
provide a multi-criteria decision method using Analytic Hierarchy Process (AHP) supporting
comparison of different software architecture candidates for software quality attributes (Svahnberg,
Wohlin, Lundberg, & Mattsson, 2003). In its process, two sets of vectors of different candidate
architectures with respect to different quality attributes and vice versa are created and refined. The
variance of uncertainty is also calculated in each candidate architecture. The sets are used as input for
a consensus-based decision making process to identify underlying reasons for disagreements amongst
stakeholders. Our proposed equations in this work are inspired by GuideArch approach (Esfahani,
Malek, & Razavi, 2013). It presents a fuzzy-based exploration of the architectural solution space
under uncertainty aiding architecture in making architecture selection. GuideArch is later extended in
(Letier, Stefan, & Barr, 2014) where authors model uncertainty about parameters’ values as
probability distributions rather than fuzzy values and also assess to extent additional information
about uncertain parameters can reduce risks. All of these works including others e.g. (Al-Naeem,
Gorton, Babar, Rabhi, & Benatallah, 2005) do not provide a systematic support for top-down goal-
obstacle analysis towards generating possible decision architecture alternatives in view of system
quality goals. Our approach can be used as a complementary step to generate different alternatives to
be used as an input for this group of studies to identify suitable solution architecture.

5.2 Legacy systems and cloud computing

An impetus to look at this track of research is the popularity of hosting data solution architectures on
the cloud computing platforms (Agrawal et al., 2011). Khajeh‐Hosseini et al. (Khajeh‐Hosseini,
Greenwood, Smith, & Sommerville, 2012) define a cloud adoption conceptual framework to support
decision makers in identifying uncertainties. They focus particularly on the cost of deploying options
of legacy systems in cloud platforms, which may undergo network latency and service price. Coth
studies limit their view to the cost of legacy system reengineering. Similarly, Umar et al. (Umar &
Zordan, 2009) defines decision model for reengineering legacy systems to service-oriented
architecture to make trade-off between integration versus migration in terms of cost. On the contrary,
we do not confine our view to the reengineering cost; rather incorporate other system quality goals

22
that might be important for stakeholders along with elaborating them to potential obstacles and
operationalisation alternatives. The approach in (Zardari, Bahsoon, & Ekárt, 2014) uses goal-obstacle
analysis to represent risks encountered in using cloud services and mitigating strategies. We extended
Zardari’s goal-oriented approach by taking into account the risk of uncertainty impact of decision
alternatives on stakeholders’ goals using fuzzy math. Furthermore, previous cloud migration literature
(Alonso, Orue-Echevarria, Escalante, Gorronogoitia, & Presenza, 2013), (Menzel, Schönherr, & Tai,
2013), (Menzel & Ranjan, 2012), and (O. Zimmermann, 2017) suffer providing a meticulous process
for an early goal-obstacle exploration and architecture decisions with considering uncertainty issue at
the same time.

5.3 Legacy systems and big data

At the organisational level, some studies aimed at identifying and analysing business goals for data
analytics adoption. For example, Park emphasizes the importance of alignment between
organisational business processes and big data sides towards making better business decisions to
adopt data analytics platforms (Park, 2017). She defines a systematic process to ensure traceability
among high-level big data adoption goals and big data solutions in view of relevance (utility of a data
element), comprehensiveness (preventing omissions of potentially important data), and prioritization
(required effort in obtaining resources for the data). In another work, Supakkul’s approach discusses
insights gained from adopting big data to improve business goals (Supakkul, Zhao, & Chung, 2016).
Their approach generates two types of resulting insight through goal reasoning and decision-making:
(i) descriptive insights of current state of business e.g. the customer retention rate and (ii) predictive
insights e.g. customers who are likely to defect. GOBIA (Goal-Oriented Business Intelligence
Architecture) is a goal-oriented approach for transforming business goals into a customized big data
architecture (Fekete, 2016). GOBIA produces a layered-based conceptual solution architecture, which
can be realized by selecting an appropriate mix of data analytics platforms, though it leaves
technology selection to implementation phase. The main difference between our approach and the
above studies is that we narrow our focus on legacy systems as the unit of analysis and explore
solution architectures to integrate them with data analytics platforms.
On the other hand, some work deal with integrating existing legacy systems with data analytics
platforms. This genre of literature is deemed closest to our work. A key feature of existing works is
their motivations in making legacy systems big data enablement. Some studies develop intelligent
techniques such as clustering (Fahad et al., 2014), deep learning (Najafabadi et al., 2015), text mining
(Xiang, Schwartz, Gerdes, & Uysal, 2015), and machine learning algorithms (Scott et al., 2016) on
big data analytic platforms to mine hiding knowledge in given legacy system data. Once chosen, such
techniques can supply inputs to the second step of our approach as decision alternatives for the goal
operationalisation or obstacle resolution (see third column of Table 8 for example) where their impact
on quality goals is investigated for the optimum selection of solution architecture.
Jha et al. define both forward and backward reengineering activities through which legacy system
functionalities are reused, and their data can be accessed and processed by data analytics platforms (S.
Jha et al., 2014). They suggest a framework to construct an architectural view of big data solution
including business, data, and application architecture (M. Jha, Jha, & O'Brien, 2015). The need for
this is highlighted by (Varkhedi, Thati, Nanda, & Alper, 2014) discussing challenges of transferring
legacy system data, e.g. mainframes, to a platform configured for big data processing in the same or a
separate logical partition on the legacy systems. (Mathew & Pillai, 2015) suggest a three layer-based
architecture for handling heterogeneities between legacy systems and data analytics platforms.
Govindarajan’s work, as a part of Cloud Collaborative Manufacturing Networks (CCMN) project,
resolves integrating supply chain manufacturing system and logistic assets with cloud services by
developing data adapters for collecting and transforming data from heterogeneous sources to
appropriate format accepted by legacy systems, i.e. XML (Govindarajan et al., 2016). Similarly,
Givehchi et al. provide a cross-layer architecture enabling interoperability between legacy industrial
devices (e.g. I/O devices and sensors) and data analytics platforms (Givehchi, Landsdorf, Simoens, &
Colombo, 2017). They apply an information model in order to retain legacy device codes unchanged.
While above approaches acknowledge challenges in legacy system big data enablement scenarios,
they keep the description of their analysis process at a high-level that does not represent ‘actionable

23
intelligence’ towards big data enablement. Nor do they address the uncertainty issue. We have
prescribed a more detailed approach for analysing goals in moving manufacturing systems to data
analytics platforms, identifying, assessing, and generating resolution tactics in handling potential risks
(i.e. step 1 of the approach). Furthermore, we address selecting, prioritization, and ranking resolution
tactics in identifying proper solution big data solution architecture under uncertainty (i.e. step 2 of the
approach). We have not found other studies that outline a systematic approach on early requirements
and big data architecture decisions.
Giret et al. state that the main reason of complexity for developing service-oriented manufacturing
systems is the number of heterogeneous technologies and execution environments (Giret, Garcia, &
Botti, 2016). They combine multi-agent system design with service-oriented architectures for the
development of intelligent automation control and execution of manufacturing systems. Giret later
proposes a process model, named Go-green, including activities, guidelines, and tools to design and
develop sustainable manufacturing system architectures (Giret, Trentesaux, Salido, Garcia, & Adam,
2017). We believe that the second step of our approach can augment the design phase of Giret’ work
to fill its gap in addressing early architecture design of big data enabled manufacturing systems under
the uncertainty.

6 Conclusion, research limitations, and further work

Legacy manufacturing systems are expected to be able to utilize data analytics platforms for advanced
information analytics. A clear understanding of goals and risks against data analytics adoption and
how they relate to manufacturing systems is particularly crucial. As a business risk management
strategy, a systematic architecture design to enable existing manufacturing systems to use data
analytics platforms is an important contribution. Our goal-obstacle analysis which takes into account
imperfect information and unavoidable uncertainties is quite intuitive to follow. In particular, it
provides an early stage analysis, which is taken place before delving into technical aspects of
implementing a big data analytics architecture. To the best of our knowledge, such a harness is not
available in the literature.
Our approach applies goal reasoning and fuzzy-based logic for analysing suitability of big data
solution architecture for manufacturing systems. The approach starts with identifying high-level
architectural goals, architectural decision alternatives to realize these goals, generating probable
obstacles, and analysing uncertainties in selecting solution architectures. The output of the approach
gives the system architect a complete set of architectural requirements to be incorporated into the
implementation stage of data analytics architecture implementation to make appropriate trade-offs
based on, for instance, cost, security, or performance goals. The application of the approach was also
demonstrated the in a scenario of moving ETL to a set of data analytics platforms. Apart from
manufacturing and big data settings, due to the genericity of the approach, it can be used in other
scenarios of technology adoption when the system architect is interested in evaluating possible
solution architecture alternatives.
Our model describes the various options (resolutions) but it does not mandate any. The choice of the
resolutions is ultimately determined by whether on the values given to the goals. Hence, strictly
speaking, it is not normative per se. However, the scale of complexity involved in the architecture
analysis may well lead the architects to rely on the advice produced. Indeed, this is required as we
argue. In a lower complexity scenario, the proposed approach may be too intrusive and the system
architect may favour an ad-hoc approach for architectural requirements and possible solution
architecture. It is important to keep in mind that the work targets settings where systematicity is
desirable and actually sought by the system architects. Hence, whilst our approach is applicable for
small-scale projects with a limited number of goals/risks and stakeholders, it is more needed for large-
scale projects where there are multiple goals, potential obstacles, and possible resolution tactics.
Indeed, the approach targets settings where a systematic and communicative approach specifying
notations and representation and model refinement mechanisms is useful.
Although we have shown the applicability of our approach, further validation is required to account
for the variety in scenarios of integrating manufacturing systems with data analytics platforms. There

24
might be some other ways to satisfy goals or some hidden factors that hinder certain goal achievement
but are not defined in the approach’s steps. Another important way for the improvement of the
approach is to provide further automatic support. The size of the goal model in Step 1 and the number
of required computations in Step 2 can limit the usability of the approach without further tool support.
We plan to provide a tool support that facilitates using the approach when working with large solution
architecture space.

References
Agrawal, D., Das, S., & El Abbadi, A. (2011). Big data and cloud computing: current state and future
opportunities. Paper presented at the Proceedings of the 14th International Conference on
Extending Database Technology.
Al-Naeem, T., Gorton, I., Babar, M. A., Rabhi, F., & Benatallah, B. (2005). A quality-driven systematic
approach for architecting distributed software applications. Paper presented at the
Proceedings of the 27th international conference on Software engineering.
Alonso, J., Orue-Echevarria, L., Escalante, M., Gorronogoitia, J., & Presenza, D. (2013, 23-23 Sept.
2013). Cloud modernization assessment framework: Analyzing the impact of a potential
migration to Cloud. Paper presented at the Maintenance and Evolution of Service-Oriented
and Cloud-Based Systems (MESOCA), 2013 IEEE 7th International Symposium on the.
Bi, Z., & Cochran, D. (2014). Big data analytics with applications. Journal of Management Analytics,
1(4), 249-265.
Camarinha-Matos, L. M., Afsarmanesh, H., Galeano, N., & Molina, A. (2009). Collaborative
networked organizations–Concepts and practice in manufacturing enterprises. Computers &
Industrial Engineering, 57(1), 46-60.
Chen, C. P., & Zhang, C.-Y. (2014). Data-intensive applications, challenges, techniques and
technologies: A survey on Big Data. Information Sciences, 275, 314-347.
Chen, S.-H. (1985). Ranking fuzzy numbers with maximizing set and minimizing set. Fuzzy Sets and
Systems, 17(2), 113-129. doi:https://doi.org/10.1016/0165-0114(85)90050-8
Dubey, R., Gunasekaran, A., Childe, S. J., Wamba, S. F., & Papadopoulos, T. (2016). The impact of big
data on world-class sustainable manufacturing. The International Journal of Advanced
Manufacturing Technology, 84(1-4), 631-645.
Esfahani, N., Malek, S., & Razavi, K. (2013). GuideArch: guiding the exploration of architectural
solution space under uncertainty. Paper presented at the Software Engineering (ICSE), 2013
35th International Conference on.
Fahad, A., Alshatri, N., Tari, Z., Alamri, A., Khalil, I., Zomaya, A. Y., . . . Bouras, A. (2014). A survey of
clustering algorithms for big data: Taxonomy and empirical analysis. IEEE transactions on
emerging topics in computing, 2(3), 267-279.
Fahmideh, M., & Beydoun, G. (2018). Reusing empirical knowledge during cloud computing
adoption. Journal of Systems and Software, 138, 124-157.
Fekete, D. (2016). The GOBIA Method: Fusing Data Warehouses and Big Data in a Goal-Oriented BI
Architecture. Paper presented at the GvD.
Franklin, C. (1996). Lt. Gen (USAF) Commander ESC, January 1996, Memorandum for ESC Program
Managers, ESC/CC. Risk Management, Department of the Air Force, Headquarters ESC
(AFMC) Hanscom Air Force Base, MA.
Gartner. (2015). Gartner Says Business Intelligence and Analytics Leaders Must Focus on Mindsets
and Culture to Kick Start Advanced Analytics: availalbe at
http://www.gartner.com/newsroom/id/3130017.
Giret, A., Garcia, E., & Botti, V. (2016). An engineering framework for service-oriented intelligent
manufacturing systems. Computers in Industry, 81, 116-127.
Giret, A., Trentesaux, D., Salido, M. A., Garcia, E., & Adam, E. (2017). A holonic multi-agent
methodology to design sustainable intelligent manufacturing control systems. Journal of
Cleaner Production.

25
Givehchi, O., Landsdorf, K., Simoens, P., & Colombo, A. W. (2017). Interoperability for industrial
cyber-physical systems: an approach for legacy systems. IEEE Transactions on Industrial
Informatics.
Govindarajan, N., Ferrer, B. R., Xu, X., Nieto, A., & Lastra, J. L. M. (2016). An approach for integrating
legacy systems in the manufacturing industry. Paper presented at the Industrial Informatics
(INDIN), 2016 IEEE 14th International Conference on.
Jha, M., Jha, S., & O'Brien, L. (2015). Integrating big data solutions into enterprize architecture:
constructing the entire information landscape. Paper presented at the The International
Conference on Big Data, Internet of Things, and Zero-Size Intelligence (BIZ2015).
Jha, S., Jha, M., O'Brien, L., & Wells, M. (2014). Integrating legacy system into big data solutions:
Time to make the change. Paper presented at the Computer Science and Engineering (APWC
on CSE), 2014 Asia-Pacific World Congress on.
Khajeh‐Hosseini, A., Greenwood, D., Smith, J. W., & Sommerville, I. (2012). The cloud adoption
toolkit: supporting cloud adoption decisions in the enterprise. Software: Practice and
Experience, 42(4), 447-465.
Lee, J., Kao, H.-A., & Yang, S. (2014). Service innovation and smart analytics for industry 4.0 and big
data environment. Procedia Cirp, 16, 3-8.
Letier, E. (2001). Reasoning about agents in goal-oriented requirements engineering. PhD thesis,
Université catholique de Louvain.
Letier, E., Stefan, D., & Barr, E. T. (2014). Uncertainty, risk, and information value in software
requirements and architecture. Paper presented at the Proceedings of the 36th International
Conference on Software Engineering.
Letier, E., & Van Lamsweerde, A. (2004). Reasoning about partial goal satisfaction for requirements
and design engineering. Paper presented at the ACM SIGSOFT Software Engineering Notes.
Li, J., Tao, F., Cheng, Y., & Zhao, L. (2015). Big data in product lifecycle management. The
International Journal of Advanced Manufacturing Technology, 81(1-4), 667-684.
Lim, S. L., & Finkelstein, A. (2012). StakeRare: using social networks and collaborative filtering for
large-scale requirements elicitation. Software Engineering, IEEE Transactions on, 38(3), 707-
735.
Lin, H.-K., Harding, J. A., & Chen, C.-I. (2016). A Hyperconnected Manufacturing Collaboration System
Using the Semantic Web and Hadoop Ecosystem System. Procedia Cirp, 52, 18-23.
Lycett, M. (2013). ‘Datafication’: Making sense of (big) data in a complex world: Springer.
Mamdani, E. H. (1974). Application of fuzzy algorithms for control of simple dynamic plant. Paper
presented at the Proceedings of the institution of electrical engineers.
Mathew, P. S., & Pillai, A. S. (2015). Big Data solutions in Healthcare: Problems and perspectives.
Paper presented at the Innovations in Information, Embedded and Communication Systems
(ICIIECS), 2015 International Conference on.
McAfee, A., Brynjolfsson, E., & Davenport, T. H. (2012). Big data: the management revolution.
Harvard Business Review, 90(10), 60-68.
Menzel, M., & Ranjan, R. (2012). CloudGenius: decision support for web server cloud migration.
Paper presented at the Proceedings of the 21st international conference on World Wide
Web.
Menzel, M., Schönherr, M., & Tai, S. (2013). (MC2)2: criteria, requirements and a software prototype
for Cloud infrastructure decisions. Software: Practice and Experience, 43(11), 1283-1297.
doi:10.1002/spe.1110
Najafabadi, M. M., Villanustre, F., Khoshgoftaar, T. M., Seliya, N., Wald, R., & Muharemagic, E.
(2015). Deep learning applications and challenges in big data analytics. Journal of Big Data,
2(1), 1.
Pal, S. K., Meher, S. K., & Skowron, A. (2015). Data science, big data and granular mining. Pattern
Recognition Letters, 67, 109-112.
Park, E. (2017). IRIS: a goal-oriented big data business analytics framework.

26
Pedrycz, W. (1994). Why triangular membership functions? Fuzzy Sets and Systems, 64(1), 21-30.
Protiviti. (2017). Big Data Adoption in Manufacturing: Forging an Edge in a Challenging Environment:
Available at: https://www.protiviti.com/sites/default/files/united_states/insights/big-data-
adoption-in-manufacturing-protiviti.pdf.
Scott, S. L., Blocker, A. W., Bonassi, F. V., Chipman, H. A., George, E. I., & McCulloch, R. E. (2016).
Bayes and big data: The consensus Monte Carlo algorithm. International Journal of
Management Science and Engineering Management, 11(2), 78-88.
Stark, J. (2015). Product lifecycle management Product Lifecycle Management (Volume 1) (pp. 1-29):
Springer.
Supakkul, S., Zhao, L., & Chung, L. (2016). GOMA: Supporting Big Data Analytics with a Goal-Oriented
Approach. Paper presented at the Big Data (BigData Congress), 2016 IEEE International
Congress on.
Svahnberg, M., Wohlin, C., Lundberg, L., & Mattsson, M. (2003). A quality-driven decision-support
method for identifying software architecture candidates. International Journal of Software
Engineering and Knowledge Engineering, 13(05), 547-573.
Umar, A., & Zordan, A. (2009). Reengineering for service oriented architectures: A strategic decision
model for integration versus migration. Journal of Systems and Software, 82(3), 448-462.
Van Lamsweerde, A. (2009). Requirements engineering: from system goals to UML models to
software specifications.
van Lamsweerde, A., & Letier, E. (2000). Handling obstacles in goal-oriented requirements
engineering. Software Engineering, IEEE Transactions on, 26(10), 978-1005.
Varkhedi, A., Thati, V., Nanda, A., & Alper, M. (2014). System and method for extracting data from
legacy data systems to big data platforms: Google Patents.
Waller, M. A., & Fawcett, S. E. (2013). Data science, predictive analytics, and big data: a revolution
that will transform supply chain design and management. Journal of Business Logistics,
34(2), 77-84.
Wang, H., Xu, Z., Fujita, H., & Liu, S. (2016). Towards felicitous decision making: An overview on
challenges and trends of Big Data. Information Sciences, 367, 747-765.
Wegener, R., & Sinha, V. (2013). The value of big data: How analytics differentiates winners.
http://www.bain.com/publications/articles/the-value-of-big-data.aspx.
Xiang, Z., Schwartz, Z., Gerdes, J. H., & Uysal, M. (2015). What can big data and text analytics tell us
about hotel guest experience and satisfaction? International Journal of Hospitality
Management, 44, 120-130.
Yu, E. S. K., & Mylopoulos, J. (1994). Understanding “Why” in Software Process Modelling, Analysis,
and Design. Paper presented at the Proceedings of the 16th international conference on
Software engineering, Sorrento, Italy.
Zadeh, L. A., Fu, K.-S., & Tanaka, K. (2014). Fuzzy sets and their applications to cognitive and decision
processes: Proceedings of the us–japan seminar on fuzzy sets and their applications, held at
the university of california, berkeley, california, july 1-4, 1974: Academic press.
Zardari, S., Bahsoon, R., & Ekárt, A. (2014). Cloud Adoption: Prioritizing Obstacles and Obstacles
Resolution Tactics Using AHP.
Zimmermann, H.-J. (1978). Fuzzy programming and linear programming with several objective
functions. Fuzzy Sets and Systems, 1(1), 45-55.
Zimmermann, O. (2017). Architectural refactoring for the cloud: a decision-centric view on cloud
migration. Computing, 99(2), 129-145.

27
Dr. Mahdi Fahmideh received a PhD degree in Information Systems from the
University of New South Wales (UNSW), Sydney Australia. He also holds a
master degree in Software Engineering. He is currently a researcher at the
University of Technology Sydney. Mahdi is a design science researcher and
his vision in research focuses on developing new-to-the-world solutions that
help organisations in adopting IT initiatives. His research outputs can be in
the form of methodological approaches, frameworks, and conceptual models.
Mahdi's research interests lie in the areas of cloud computing, big data,
model-driven software development, and situational method engineering.
Prior to joining to the academia, Mahdi has worked as a programmer and
system analyst in development of software systems in different industry
sectors including accounting, insurance, defense, and publishing.
Professor Ghassan Beydoun received a degree in computer science and a
PhD degree in knowledge systems from the University of New South Wales.
He is currently a Professor of Information Systems at the University of
Technology Sydney. He has authored more than 100 papers international
journals and conferences. He is currently working on the metamodels for on
project sponsored by Australian Research Council and Australian companies
to investigate the endowing methodologies for distributed intelligent systems
and supply chains with intelligence. His other research interests include multi
agent systems applications, ontologies and their applications, and knowledge
acquisition.

28