A method for jointly drawing up the functional and design architectures of

complex systems during the preliminary system-definition phase

Éric Bonjour1*, Samuel Deniaud2, Jean-Pierre Micaëlli3

(1) Professor of Systems Engineering, Université de Lorraine, ENSGSI, Laboratoire ERPI,
EA 3767, 8 rue Bastien Lepage, BP647, F-54010 Nancy Cedex. [email protected]
(2) Associate Professor of Mechanical Engineering, IRTES-M3M, UTBM, F-90010 Belfort
Cedex. [email protected]
(3) Associate Professor of Industrial Management, Université de Lyon, INSA de Lyon, ITUS
Research Team, UMR CNRS 5600 EVS, Campus de La Doua LyonTech, 1 rue des
Humanités, F-69621 Villeurbanne Cedex. [email protected]
* Corresponding author

In systems engineering, determining a system’s architecture is a crucial part of the preliminary

system-definition phase because this architecture greatly impacts the quality of the system,
the way the subsequent phases of the design process are organized and the overall
performance of this process. Consequently, systems architects need methods for defining the
couplings of modules (or sub-systems) and for estimating the impact of allocation decisions
on candidate system architectures. This paper presents a method that helps simultaneously
define the functional and design architectures of complex systems. Starting from an allocation
matrix, or Domain Mapping Matrix (DMM), which represents the couplings between the
elements of two domains (functions vs. components), the method generates a Design
Structure Matrix (DSM) for each domain. A clustering algorithm is then applied to each DSM
in order to generate a rearranged architecture. This method can be used in the synthesis
process to provide architectural output that then forms input data for the systems analysis
process, where these data will help systems architects assess candidate architectures. We
illustrate the method via an industrial case study focusing on a new automobile engine
development project.

Keywords: systems engineering, systems architecture, dependency modelling, Domain

Mapping Matrix, DSM

1
1. Introduction

Complex systems are usually developed by decomposing them into sub-systems, which may
be further decomposed into smaller components if the sub-systems are still too complex
(Alexander, 1964; Pimmler and Eppinger, 1994; IEEE Std 1220™, 2005). This approach is
based on the assumption that most system architectures are “semi-decomposable”
(Simon, 1969). According to Ulrich (1995), architectures consist of (1) an arrangement of
functional elements, that is, the function structure; (2) mapping from functional elements to
components (or design elements); and (3) the specification of the interfaces between
interacting components. He defines architectures as “the scheme by which the function (…) is
allocated to components”. Thus, depending on the level of abstraction with which they are
considered, systems architectures may be viewed in terms of: (1) stakeholders’ expectations
and lifecycle process requirements (requirements view), (2) functions that are arranged within
the functional architecture (functional view), and (3) the sub-systems and components
included within the design architecture (component view) (IEEE Std 1220™, 2005).
Consequently, the systems engineering process is divided into four main sub-processes –
requirement analysis, functional analysis, synthesis and the systems analysis process – all of
which aim to build, analyse and assess candidate requirement baseline, functional
architectures or design architectures. These sub-processes concerning the generation of
candidate solutions should be performed in an iterative manner because systems architects
seek a satisfactory design solution (IEEE Std 1220™, 2005) in a “bounded” solution space
(Simon, 1969). Candidate solutions are characterized by specific allocation matrices that may
contain the same set of functions but different sets of components. The systems analysis
process requires systems architects to analyse and evaluate each candidate solution.
In this paper, we will focus on the synthesis process that mainly consists in grouping and
allocating functions to subsystems, and identifying design solution alternatives. System
architects have to group functions and subfunctions into logical functional elements in a
manner that permits their allocation to design elements (IEEE Std 1220™, 2005). Then they
generate the functional architecture and candidate design architectures (syn. arrangement of
system elements).
Candidate design architectures are produced, including the building of allocation matrices that
capture the mapping between functions and components in order to trace allocation decisions.
Allocation matrices (or traceability matrices), such as requirements/functions and
functions/components matrices, are highly important in systems engineering processes
(IEEE Std 1220™, 2005) because they ensure the cohesion between system requirements,
system functions and sub-systems or components. Such matrices are also called Domain
Mapping Matrices (DMM) (Danilovic and Browning, 2004; Danilovic and Browning, 2007).
The underlying allocation task is of great importance. It relies on best practices, rules and

2
systems architecting methods (Ulrich, 1995; Suh, 1990, 2001; Maier and Rechtin, 2002).
Allocation matrices with numerical values contain architectural data and are used to support
traceability. It should be noted that, in the preliminary stage of systems architecting,
components (or design elements) are not physical objects represented by geometric models.
All these components are not available: they are not commercial-off-the-shelf items (COTS).
Rather, they are abstract entities specified by the allocation of functions and by the
specification of flows, interfaces and events (Stone et al. 2000; Nagel et al. 2011). This
abstract approach helps systems architects explore the solution space without inhibiting
creativity through the predetermination of physical components. Although design methods
such as Axiomatic Design (Suh, 1990) and Quality Function Deployment (QFD) (Akao,
1990), and systems architecting principles (Sage and Lynch, 1998; Baldwin and Clark, 2000;
Ulrich and Eppinger, 2008) describe valuable rules and guidelines for allocating functions to
components, there is no formal method for handling architectural data from both functional
and component points of view. When the coupling between functions and components is not
formally addressed, synthesis is more difficult and may require numerous iterations.
The present paper describes a method that supports the joint generation of functional and
design architectures in the synthesis process. Starting from a rearranged DMM, which is used
to trace system architects’ decisions concerning the allocations of functions to components,
our method generates two Design Structure Matrices (DSM) (Steward and Donald, 1981)
corresponding to a design architecture (Component DSM) and an allocated functional
architecture (Function DSM). Our method helps them to jointly identify modules in the
Component DSM and Function DSM for each candidate solution. Systems architects can
judge the consistency of the obtained matrices by considering the working of the system. The
in-depth evaluation of each candidate solution shall be led later in the systems analysis
process.
This paper addresses two contributions: (1) a scientific contribution concerning the formal
integration of fuzzy representation and a matrix-based method, and (2) a methodological
contribution related to the use of a new tool in the preliminary systems definition. The result
of this method may facilitate further systems analysis, integration and verification processes
(ISO/CEI 15288, 2008). Moreover, the case study points out that by means of the proposed
method, industrial practices may be changed to better decompose the systems at different
layers.
In this paper, we begin by briefly reviewing research into systems architecture modelling and
by highlighting some of the expectations architects have for new methods. We then outline a
new matrix-based method for simultaneously generating the functional and design
architectures. In order to illustrate the value of our method, we applied it to the definition of
the architecture of a new diesel engine being designed by a French car manufacturer, and then
asked the project’s systems architects to interpret the results and analyse the consistency of

3
the DSM generated. A discussion is proposed to point out the interest of the proposed method
along with the limitations of its validation by a single moderately complex case study. Finally,
we conclude our work and propose future research directions.

2. Systems architecture modelling research

This section reviews research into systems architecture and matrix-based methods. A key
feature of a system’s architecture is the degree to which it is modular or integral (ESD, 2004).
In modular architectures, one product function maps onto one component (or design element).
In integral architectures, which represent the other extreme, several product functions map
onto one or a small number of components (Ulrich, 1995). In real design situations, designers
have to make a trade-off between modular and integral architectures; therefore, architectures
are hybrid, combining modular and integral elements (Sosa et al.2003). As a result, it is
difficult to group together a set of functions and to couple this set to a group of components.
Modules generally consist of groups of functions or groups of components. The degree of
interaction (syn. dependency) is maximized within each group (or module) and it is
minimized between groups (inter-modules) (Whitfield et al. 2002).
Matrix-based systems-modelling methods (DSM and DMM) represent system architectures
(system elements and their relationships) as matrices. They are being increasingly used in
systems engineering for product modularization (Ulrich and Eppinger, 2008; Stone et al.
2000; Alizon et al. 2007), for designing product platform architectures (Luh et al. 2011; Li et
al. 2012), for analysing technical interactions either within products (Whitfield et al. 2002;
Helmer et al. 2010; Bonjour et al. 2009) or within the project organization (Pimmler and
Eppinger, 1994; Sosa et al. 2003; Arundacahawat et al. 2011), for analysing change
propagation (Clarkson et al. 2004; Keller et al. 2009; Bonjour et al. 2010-a; Fei et al. 2011;
Cheng and Chu, 2012) and for representing the structure of collective design competencies
(Bonjour et al. 2010-b).
DSMs represent relationships between elements within the same domain, for example,
between components. They have become popular modelling and analysis tools, especially for
decomposition and integration (Pimmler and Eppinger, 1994; Browning 2001), because they
display the relationships between a system’s elements in a compact and visual format.
Pimmler and Eppinger (1994) proposed a typology of interactions (spatial, energy,
information, materials) within product architectures. Static DSMs are usually analysed using
clustering algorithms.
DMMs represent relationships between two domains (Danilovic and Browning, 2004;
Lindemann, 2007; Eichinger et al. 2006; Bartolomei et al. 2012). Other authors call these
matrices “incidence matrices” (Chen et al. 2005). In systems engineering, they are called
“traceability matrices” or “allocation matrices”. They can represent a set of design decisions
or the relationships between the what and the how. One example is the Axiomatic Design

4
Matrix (Suh, 1990; Suh, 2001), which captures the relationships between a function and the
design parameter that realizes that function. Another example is the QFD method (Akao,
1990), which uses several inter-domain matrices to convey customer requirements throughout
a development project. Inspired by QFD, Erixon (1998) used the concept of “module drivers”
to develop the Modular Function Deployment (MFD) method, whose aim is to draw up a
system architecture that takes into account key considerations of the system’s life cycle.
Strawbridge et al. (2002) used binary function-component matrices to obtain morphological
matrices. Fixson (2005) proposed creating a numerical “function-component allocation”
matrix, whose cells can be used to allocate “percentages of a function […] to components that
contribute to this function”.

3. From one DMM to two DSMs: A matrix-based method

As part of a research contract with a French car manufacturer, we analysed the activities of
systems architects involved in developing powertrains (engine block and gearbox). Our study
focused on the preliminary system-definition phase. This powertrain design department tends
to use DMMs (allocation matrices) as a representation tool, rather than DSMs, thus, the
architects expressed requirements for methods and tools capable of manipulating these
matrices:
Requirement 1: Compute the effects of allocation decisions (made during the synthesis
process, and formalized in the Function/Component DMM) on the functional and design
architectures,
Requirement 2: Generate the rearranged functional and design architectures in order to
provide input data for the systems analysis process.

Axioms and assumptions

This section of our paper explains the axioms on which our method is based, together with the
underlying assumptions. Although they are based on different axioms, the methods for
generating the Function DSM (F-DSM) and the Component DSM (C-DSM) are similar;
therefore, the following section only explains how the F-DSM is generated.
Axiom 1. If two functions Fi and Fk are allocated to a component Cu, then Fi and Fk are
coupled through that component (the corresponding component axiom is: “If two components
Cu and Cv contribute to the fulfilment of Fi, then Cu and Cv are coupled through that
function”).
Axiom 2. The intensity of the coupling between Fi and Fk (respectively Cu and Cv) is related
to the intensity of their coupling with component Cu (respectively Fi).
First, if the fulfilment of both Fi and Fk depends on component Cu, designers working on Cu
are likely to seek trade-offs between the performances of functions Fi and Fk. The coupling
between the two functions is clear. Negotiations need to be carried out to define the

5
contribution of Cu to the fulfilment of each function. Conversely, if there is no overlap
between Fi and Fk for the design of all the components, the two functions can be assumed to
be mutually independent (no coupling) of the component domain. Second, if a given function
Fi is partially solved by component Cu (i.e., the intensity of the coupling between Fi and Cu is
low in the F-C allocation matrix), the fulfilment of Fi can be assumed to be rather independent
of Cu, that is, the coupling between Fi and the other functions through the component Cu is
low.

A matrix-based method

Our proposed method is based on the five stages (Figure 1) outlined below. Because
intensities inside DMMs are quite imprecise and subjective, Fuzzy Logic (Dubois and Prade,
1980) provides a suitable approach for manipulating these values. In fact, systems architects
have to estimate the strength of the relation between each function and components. Functions
transform input flows into output flows, which may be material, energy or information flows.
Allocating a function to two components implies that these components exchange these types
of flows. Consequently, the proposed method generates one F-DSM for each component. The
method supposes that the allocation of functions onto that component is the source of
couplings between the functions. The rules introduced by the axioms are applied and
translated into a fuzzy inference system. The resulting DSM is then aggregated and the
aggregated DSM is filtered to delete meaningless values. A clustering algorithm is used to
identify modules and integrative elements. In this paper we refer to these results as candidate
architectures because displaying them will help the architect analyse the consequences of
allocation decisions on the rearranged F-DSM and C-DSM.

Figure 1. Overall structure of the matrix-based method

Building a numerical F-C DMM

The DMM has to be filled in by the systems architect and his/her team, who draw up a
function list. Hence, we asked the systems architect to fill in the functions-components DMM
(F-C DMM) by eliciting his/her decisions on the allocation of each function onto components.
Discussions were needed to determine whether or not the interactions really existed and to
estimate the intensity of each coupling. In a cell, coupling intensity is defined as the

6
“percentages of a function (that) can be allocated to components that contribute to this
function” Fixson (2005). Coupling intensity is assumed to be between 0 and 10.

Generate one F-DSM for each Component

The axioms presented above form the basis for generating one F-DSM per component, using
a Mamdani fuzzy inference system (Dubois and Prade, 1980; Zimmerman, 1991) to analyse
the inputs through a set of ‘if-then’ rules and a T-norm aggregation operator. If NC denotes the
number of components in the DMM, this step generates NC F-DSMs. The following step
describes an aggregating method that can be used to obtain a single DSM. Fuzzy processing,
implemented using Matlab® and its fuzzy toolbox, consists of three basic stages: (1)
fuzzification and input membership functions, (2) inferences and fuzzy rules, (3)
defuzzification and output membership functions. The fuzzification stage corresponds to the
transformation of a numerical value through fuzzy variables (input). Our choice of structure
for the membership functions characterizing the two inputs was made taking into account the
systems architect's opinion. We defined four linguistic variables, Null, Low, Medium and
High (see Figure 2), and used the most common membership function, that is, a trapezoidal
function.

Figure 2. Membership functions of fuzzy variables

The fuzzy ‘if-then’ rules were defined to relate the input variables to the output variables.
These rules represent an expert’s knowledge of the interactions between input variables and
their effects on the output. The inference system approximates the way systems architects
estimate the coupling between two Functions (Fi and Fk) through one component Cu:
1. IF (Fi-Cu is Null) OR (Fk-Cu is Null) THEN (Fi - Fk is Null)
2. IF (Fi-Cu is Low) AND (Fk-Cu is NOT Null) THEN (Fi - Fk is Weak)
3. IF (Fi-Cu is NOT Null) AND (Fk-Cu is Low) THEN (Fi - Fk is Weak)
4. IF (Fi-Cu is Medium) AND (Fk-Cu is Medium) THEN (Fi - Fk is Average)
5. IF (Fi-Cu is Medium) AND (Fk-Cu is High) THEN (Fi - Fk is Strong)
6. IF (Fi-Cu is High) AND (Fk-Cu is Medium) THEN (Fi - Fk is Strong)
7. IF (Fi-Cu is High) AND (Fk-Cu is High) THEN (Fi - Fk is Strong)

7
Because our goal was to generate an understandable inference system for systems architects,
we intentionally limited the number of rules and hence the number of linguistic variables.
The defuzzification stage involves using the output membership function to find a crisp value
for the coupling. The aggregated fuzzy output is defuzzified using the “centroid of area”
technique and the formula given in Dubois and Prade (1980). This is the most widely used
defuzzification method. Figure 3 summarizes the main features of the fuzzy inference system.
The choice of linguistic variables was designed to limit the influence of low inputs.

Figure 3. Features of the fuzzy inference system

Aggregating the NC DSM and filtering the aggregated DSM

In order to obtain the aggregated DSM, we used the average method: if DSM(i,j)k denotes the
coupling intensity between Fi and Fj in the DSM generated by component Ck, and DSMA
denotes the Aggregated DSM, the DSMA is calculated as follows (Equation 1):
DSMi, jk
NC
DSM A i, j  
k 1
NC
(1)
The F-DSM obtained via the average method may be dense because of the density of the
DMM. The density of the F-DSM is not a problem for interpreting the coupling between the
functions, but the clustering algorithm used to determine the “satisfactory” (Simon, 1969)
functional architecture is sensitive to the density of the couplings. Consequently, it may be
necessary to filter low values in the aggregated F-DSM, in order to reduce the influence of
meaningless coupling intensities by converting them to zero (Equation 2):
If DSM(i,j) ≤ X then DSM(i, j) = 0 (2)
where X is a parameter that is automatically fixed by the algorithm in order to obtain a DSM
density of less than 80%.

Use a clustering algorithm and display the architecture

A clustering algorithm is a method that generates modules in a systematic way by optimizing

an objective function. This algorithm is based on the use of DSM to map interactions.
Hartigan (1975) reviewed the basic clustering approach and discussed different applications
of clustering algorithms. The original goal of clustering was to find similarities between
elements and to group them together on the basis of a similarity threshold. Several algorithms

8
for optimizing modular design have been developed, including simulated annealing
(Fernandez, 1998; Thebeau, 2001) and genetic algorithms (Yu et al. 2003). The clustering
algorithm used in the present paper is based on an algorithm developed by Idicula (1995) and
improved by Fernandez (1998) and Thebeau (2001). Idicula's algorithm assumes an
underlying directed graph model for the development effort, and uses a depth-first-search
technique to solve the problem. It groups the “project tasks into clusters that are loosely
connected with each other, with each cluster consisting of densely connected inter-coupled
tasks” (Idicula, 1995). Both Fernandez (1998) and Thebeau (2001) added simulated annealing
procedures to better explore the solution space and significantly improve clustering results.
Due to the randomness of simulated annealing, several runs of the algorithm are required to
provide an optimal DSM clustering. It is also best to change the initial DSM arrangement and
then run the algorithm again, as this kind of algorithm may be dependent on the initial
configuration. Results can be displayed in a number of ways, and the clustered DSM can be
represented as either a numerical matrix or as a graphical matrix with diamonds. In the latter
case, the size of each diamond is proportional to the value in the corresponding cell of the
initial numerical DSM. As a graphical DSM gives a more visual representation than a
numerical DSM, we chose the former to display the results of our case study. The architecture
can be interpreted by direct inspection of the DSM (Sharman and Yassine, 2004).

Interpretation of results

The systems architect was asked to interpret and compare the rearranged functional and
design architectures.

Short discussion

From a mathematical perspective, it is worth noting that DMM size impacts the validity of the
proposed method. The mathematical operations involved are similar to a projection function.
Each domain has to be large enough and to have similar cardinality (matrix size greater than
15). However, when defining the architecture of a moderately complex system, the number of
elements within a domain is usually greater than 15 and a similar decomposition level should
be used for each domain.

4. An industrial case study: generating an engine architecture

The following section applies the proposed method to the definition of diesel engine
architectures. We briefly present the allocation matrix of a diesel engine and the results
provided by our method. In order to assess the value of the proposed method, we asked the
project’s systems architect to check the consistency of the two DSMs generated. Although a
formal functional lexicon has already been proposed (Stone et al. 2000; Strawbridge et al.

9
2002), another functional lexicon was adopted in this case study because of the design
vocabulary and rules used in that car manufacturer’s design office.

Numerical F-C DMM

We generated a F-C DMM by asking the systems architect to enter numerical values into the
DMM. Intensities could range from 0, for no coupling between the function (in the columns)
and the component (in the rows), to 10 for total coupling between the function and the
component.

Function F 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

Power Transmission

Functional Volumes
Secondary Energy
Functions /
Pressure-Torque
Fuel provision
Component C

Gas Cleaning
Air Provision

Combustion

Conversion.

Lubrication

Conversion

Ventilation
Components

Thermics
Vibration

Coupling
Friction

Control
DMM

1 EGR 9 9 9 9 7
2 Fuel System 8 8 7 5 9 8 9 8 6 8
3 Breech Block 8 9 6 7 8 9 5 9 5 9 8
4 Air Intake 7 5 9 6 8 7 9 8
5 Exhaust 8 8 4 8 8 7 8 7
6 Camshaft/Valve Train 9 8 8 4 9 5 6 8
7 Crankshaft 9 7 9 7 6 6 8 5 8 8 9
8 Casing 6 9 8 9 8 7 9 7 8 9 9
9 Lubrication and Blow-by 8 7 9 7 4 8 7
1 Accessory Drive 7 6 9 7 7 4 8
11 Synchronous Drive 8 5 7 8 7 6 7
12 Vacuum Circuit 8 4 9 6 8
13 Cooling Circuit 5 4 5 6 7 7
14 Secondary Energy 4 5 9 5 5 7
Generator
15 Sensors and Control 7 9 7 6 7 4 4

Table 1. Numerical F-C DMM

Table 1 shows a F-C DMM for an engine block. It is a numerical matrix with 15 elements in
rows, corresponding to the 15 components that make up the engine block, and 15 elements in
columns, corresponding to the 15 functions. In this case, the fact that the number of
components equals the number of functions is a pure coincidence. However, the allocation
matrix is not a simple transfer from functions to components, so this does not limit the
assessment of our method. For example, looking down the “Fuel provision” column reveals
that this function is allocated to three components: Fuel System, Breech Block and Sensors.
Similarly, reading along the “Fuel System” row shows that this component contributes to nine
functions: Fuel Provision, Combustion, Friction, Lubrication, Power Transmission, Vibration,
Control, Thermics and Functional Volumes.

10
Implementation

The proposed method uses a numerical F-C DMM as input and gives two DSMs (F-DSM and
C-DSM) as results. Hence, the clustering algorithm is applied to these two DSMs in order to
reveal underlying architectures. Because the F-DSM and C-DSM are the same size (both have
15 elements), the same parameter values can be used to cluster these two DSMs. Satisfactory
values are provided by Fernandez (1998) and Thebeau (2001). Furthermore, for the filtering
stage for this engine design, the algorithm put X=1.1 in order to reduce the density of each
DSM to less than 80% (i.e., 42 null elements out of the 225 elements in the F-DSM).

Function-DSM with clustering

The functional architecture shown in Figure 4 is as follows:

Functional module 1: Air provision, Fuel provision, Gas Cleaning and Combustion,
Functional module 2: Pressure-Torque Conversion, Friction, Lubrication, Secondary Energy
Conversion and Coupling,
Integrative elements: Functional Volumes, Thermics, Vibration, Ventilation and Power
Transmission.
The first functional module, which corresponds to the pre-combustion stage, groups all the
functions that lead to combustion, whereas the second functional module, which corresponds
to the post-combustion stage, groups all the functions that transform the combustion energy
into mechanical energy. The integrative elements act as functions that map together all the
other functions and give the system its overall cohesion.

Figure 4. F-DSM with clustering Figure 5. C-DSM with clustering

Component-DSM with clustering

Using the same parameter values for the clustering algorithm, we generated the clustered C-
DSM (Figure 5) at the same time as the clustered F-DSM. Given the complexity of the

11
engine, the development team decided not to try and optimize the development process by
identifying modular engine subsystems. The design architecture shown in Figure 5 is as
follows:
Component module 1: The first module consists of eight components that directly contribute
to the main external engine function, that is, the transformation of thermal energy into kinetic
energy,
Component module 2: The second module brings together components that contribute to
secondary engine functions, such as vacuum production and energy generation,
Integrative elements: The integrative elements are: casing, fuel system and camshaft.
The clusters with main and secondary external functions formalize the structure of the current
design process, which gives priority to designing the components responsible for the main
engine functions.

Validation by the systems architect

To validate the proposed method, we asked the systems architect to assess the consistency of
the DSM generated by our method. In this case study, the proposed method met the
requirements outlined in section 3.
Comparing the rearranged F-DSM and the C-DSM (Figures 4 and 5) showed that clustering
the two DSMs led to approximately the same architecture. In both cases, we obtained two
modules and five integrative elements. The functional architecture highlighted an important
separation between the pre- and post-combustion stages, whereas the design architecture
separated components that contribute to the main and secondary engine functions. In fact,
most of the components impacted by the functions involved in the pre-combustion stage
(functional module 1) belonged to the main module contributing to the main engine functions
(component module 1). Similarly, there was also a match between the post-combustion
module (functional module 2) and the bottom engine module (component module 2). Most of
the components in the bottom engine module contribute to the transfer of kinetic energy and
so fulfil post-combustion functions. The integrative components mainly contribute to post-
combustion functions. The casing is the component where most of the post-combustion
functions take place, and the crankshaft contributes to the transfer of the kinetic energy.
The clustered C-DSM initially received a mixed reaction from the systems architect, but when
he analysed it in detail, he judged that the rearranged architectures were consistent with the
energy, information, and material flows through the engine. The development team had not
identified the three integrative components in the C-DSM as integrative elements before we
presented them with our results. In fact, only the casing has a physical integrative role
because it connects the other engine components. The systems architect felt that it was
interesting to identify the fuel system and the camshaft as integrative elements, as they play a

12
central role in the operation of the engine. These components share many interfaces with other
components.
The results of the comparison between the functional and design architectures matched the
systems architects' expectations and objectively highlighted some of the modules already built
in past projects. Engine designers have long realized that there is a strong correlation between
the pre-combustion stage and the main engine functions, so the match between functional
module 2 and component module 1 was judged satisfactory.
Finally, according to the systems architect, the allocation matrix is satisfactory when the F-
DSM and the C-DSM show architectures that make sense for them. However, it is worth
noting that each DSM is interpreted differently.

5. Discussion

The proposed method generates F-DSMs and C-DSMs in similar ways. According to
Axiom 2, every relationship between the components in the C-DSM should be determined on
the basis of functional aspects. Of course, DSMs have to be enriched with information about
the types of coupling. Such enrichment can be based on the method described by Pimmler and
Eppinger (1994), which was validated with respect to an automotive climate control system.
Our method simultaneously provides function and component allocations; therefore it differs
from Pimmler and Eppinger’s method, which only focused on the C-DSM.
How will the proposed method affect systems architecting?
The industrial experiments we have led have showed us that system architects did not use to
respect a simple architecting recommendation: decompose each subsystem into 5 to 8
components only for better mastering of the complexity at each layer. In that moderately
complex case study, the system was decomposed into 15 components due to time constraints
in the project that force the designers to go faster to the solution. Then the resulting product
breakdown structure is broader and less deep. Such a structure does not facilitate the system
building in the following steps of the integration and verification processes (ISO/CEI 15288,
2008). The proposed method contributes to building an ex post product breakdown structure
and makes it possible to generate candidate design architectures according to the previous
architecting recommendation. As a result, the size of the DMM matrix may be larger, up to
about 30 elements in rows and columns. However, a larger number of elements would reveal
inappropriate industrial practices in the multi-layer decomposition of the system.
We are aware of the limitations of the presented case study for validating the proposed
method. First, one moderately complex case study is not sufficient to state that the method is
efficient. However, the step-by-step explanations of the results will help the practitioners to
understand the interest of this approach for aiding them in the synthesis process. It should be
noted that the proposed method and the simple architecting rule helped systems architects to
adopt a new design practice that they judged more appropriate for achieving the rest of the

13
design process. Second, the method should be tested with larger DMM (up to about 30
components). However, the number of expected subsystems would be the same as in the case
study: 5 to 8 subsystems (it would be possible by tuning the parameters of the clustering
algorithm). Third, it could be interesting to compare the design architecture obtained by the
proposed method with an already known architecture. This comparison could be possible if
we consider a detailed design of the architecture where all components are well defined.
However, this case study was taking place in the preliminary design phase. Hence such an
architecture was not known. Fourth, the consistency of the results was judged by only one
systems architect. Some biases may have been introduced in the validation of the method. To
reduce the risk of misinterpretation, we asked him to assess the value of this qualitative
method for objective explanations based on the working of the engine. Moreover, the results
were presented to the project manager, who gave the same explanations and validated the
proposed architectures.

6. Conclusions and perspectives

The method described in this paper transforms a given numerical F-C DMM into two DSMs
(F-DSM, C-DSM), and then generates rearranged functional and design architectures for
complex systems. Our method combines matrix-based models, a fuzzy inference system and a
clustering algorithm, in order to define architectures with modules and integrative elements.
The set of rearranged candidate architectures has to be assessed by taking into account other
constraints and requirements imposed by the life cycle of the system and the ex post product
breakdown structure provided by our method.
We have already used academic examples to check that the method is more efficient when the
DMM density is less than 60%. Recommendations inspired by the Axiomatic Design rules
may be useful for building DMMs. Moreover, two contradictory requirements could be added
to improve the proposed method:
(1) Be insensitive to the imprecision of the DMM values. These values are based on a scale
from 0 to 10. An imprecision of up to ±2 will not affect the architecture obtained;
(2) Be sensitive to changes requiring new allocation decisions. If the design manager wishes
to increase or decrease the weighting, a new allocation decision is obtained (change greater
than ± 3).
In order to evaluate these requirements, it will be necessary to analyse the sensitivity of the
method to the precision of the input value or to changes in the input value.
More generally, this method may also be used in other project domains (product domain, e.g.,
requirements, functions, components, and design organization, e.g., teams, actors, tasks) and
with other DMMs, such as requirement-function, component-task or task-team DMMs.
Currently, optimization of the DSM in the two domains (functions, components) is carried out
separately. One of our objectives is to adapt the clustering algorithm to jointly optimize both

14
DSMs and to enhance modularization. In the preliminary system definition phase, project
managers and systems architecture teams have to simultaneously design the preliminary
system architecture and the overall project structure. Our goal is to provide them with an
integrated tool that can be used to concurrently generate and structure the different project
domains.
Finally, in re-engineering situations, systems architects may re-use object knowledge
developed in past projects and included in different allocation matrices and DSMs. Thus,
depending on the modifications introduced by the re-engineering situation, our method should
help systems architects by providing them with a tool that tests the robustness of architectures
and propagates modifications and constraints through the product, process and organization
project domains.

Acknowledgements
This work was carried out as part of a research project funded by a car manufacturer’s
powertrain architecture department. We would like to thank the senior managers and
personnel of this department for their cooperation and assistance.

15
REFERENCES
Akao, Y., 1990. Quality Function Deployment, QFD – Integrating Customer Requirements
into Product Design, Portland: Productivity Press.
Alexander, C., 1964. Notes on the Synthesis of Form. Cambridge: Harvard University Press.
Alizon, F., Moon, S.K., Shooter, S.B. and Simpson, T.W., 2007. Three Dimensional Design
Structure Matrix with Cross-Module and Cross-Interface Analysis. Proc. of ASME 2007
International Design Engineering Technical Conferences and Computers and Information
in Engineering Conference (IDETC/CIE2007), September 4–7, Las Vegas, NV, 941-948,
Paper no. DETC2007-34510.
Arundacahawat, P., Roy R., and Al-Ashaab A. (2011) An analogy based estimation framework
for design rework efforts. Journal of Intelligent Manufacturing. DOI 10.1007/s10845-011-
0605-6. Published online: 21 December 2011.
Baldwin, C.Y. and Clark, K.B., 2000. Design Rules: The Power of Modularity Design.
Volume 1. Boston: The MIT Press.
Bartolomei, J., Hastings, D., de Neufville, R., and Rhodes D., 2012. Engineering Systems
Multiple-Domain Matrix: An Organizing Framework for Modeling Large-Scale Complex
Systems. Systems Engineering, 15(1), 41–61.
Bonjour, E., Deniaud, S., Dulmet, M., and Harmel, G., 2009. A fuzzy method for propagating
functional architecture constraints to physical architecture. Transactions of ASME,
Journal of Mechanical Design, 131(6), 061002-1–061002-11.
Bonjour, E., Harmel, G., Micaëlli, J-P., and Dulmet, M., 2010-a. Simulating change
propagation between product architecture and development organization. International
Journal of Mass Customization, 3(3), 288-310.
Bonjour, E. and Micaëlli, J-P., 2010-b. Design Core Competence Diagnosis: A Case from the
Automotive Industry’, IEEE Transaction on Engineering Management, 57(2), 323–337.
Browning, T.R., 2001. Applying the design structure matrix to system decomposition and
integration problems. IEEE Transactions on Engineering Management, 48(3), 292–306.
Chen, L., Ding, Z. and Li, S., 2005. A formal two-phase method for decomposition of
complex design problems. Journal of Mechanical Design, 127(2), 184–195.
Cheng H., and Chu, X., 2012. A network-based assessment approach for change impacts on
complex product. Journal of Intelligent Manufacturing, 23, 1419–1431.
Clarkson, P.J., Simons, C.S. and Eckert, C.M., 2004. Predicting change propagation in
complex design. Journal of Mechanical Design, 126 (5), 788–797.
Danilovic, M. and Browning, T., 2004. A formal approach for domain mapping matrices
(DMM) to complement design structuring matrices (DSM). Proc. of the 6th International
Design Structure Matrix workshop, Cambridge, UK, September 12–14.
Danilovic, M. and Browning, T.R., 2007. Managing complex product development projects
with design structures matrices and domain mapping matrices. International Journal of

16
 Project Management, 25, 300–314.
Dubois, D. and Prade, H., 1980. Fuzzy sets and systems: Theory and applications. New York:
Academic Press.
Eichinger, M., Maurer, M, Pulm, U. and Lindemann, U., 2006. Extending Design Structure
Matrices and Domain Mapping Matrices by Multiple Design Structure Matrices. Proc. of
the ASME 8th Biennial Conference on Engineering Systems Design and Analysis
(ESDA2006), July 4–7, Torino, Italy, 889–898, Paper no. ESDA2006-95266.
Erixon, G., 1998. Modular Function Deployment – A method for product modularisation.
Thesis (PhD), Royal Institute of Technology, Stockholm.
ESD Architecture Committee (Whitney, D.E., De Weck, O., Eppinger, S., Magee, C., Moses,
J., Seering, W., Schindall, J. and Wallace, D.), 2004. The influence of architecture in
engineering systems. MIT Engineering Systems Monograph, March 29-31.
Fei, G., Gao, J., Owodunni O. and Tang X. 2011. A method for engineering design change
analysis using system modelling and knowledge management techniques. International
Journal of Computer Integrated Manufacturing, 24(6), 535-551.
Fernandez, C., 1998. Integration analysis of product architecture to support effective team co-
location. Thesis (MSc), Massachusetts Institute of Technology.
Fixson, S.K., 2005. Product architecture assessment: a tool to link product, process and
supply chain design decisions. Journal of Operations Management, 23(3/4), 345–369.
Hartigan, J., 1975. Clustering Algorithms. New York: Wiley & Sons.
Helmer, R., Yassine A. and Meier, C., 2010. Systematic module and interface definition using
component design structure matrix. Journal of Engineering Design, 21(6), 647–675.
Idicula, J., 1995. Planning for concurrent engineering. Thesis draft, Nanyang Technology
University.
IEEE Standard 1220™, 2005. Standard for application and Management of the Systems
Engineering Process. IEEE Computer Society, Revision of IEEE Standard 1220-1998.
ISO/CEI 15288:2008, 2008. Systems and software engineering —System life cycle processes.
Copyright © 2008 ISO/IEC-IEEE.
Keller, R., Eckert, C.M. and Clarkson, P.J., 2009. Using an engineering change methodology
to support conceptual design. Journal of Engineering Design, 20(6), 571–587.
Kusiak, A. and Huang, C.C., 1996. Development of modular products. IEEE Transactions on
Components, Packaging, and Manufacturing Technology – part A, 19(4), 523–538.
Li Z., Cheng Z., Feng Y. and Yang, J., 2012. An integrated method for flexible platform
modular architecture design. Journal of Engineering Design, Version of record first
published: 09 Mar 2012. DOI:10.1080/09544828.2012.668614
Lindemann, U., 2007. A vision to overcome chaotic "design for X" processes in early phases.
Proc. of the International Conference on Engineering Design (ICED’07), Paris, France,
August 28-31.

17
Luh, D-B., Ko Y-T., and Ma C-H., 2011. A structural matrix-based modelling for designing
product variety, Journal of Engineering Design, 22(1), 1-29.
Maier, M.W. and Rechtin, E., 2002. The art of systems architecting. 2nd ed. New York: CRC
Press LLC.
Nagel, R., Hutcheson, R., McAdams D.A., and Stone R., 2011. Process and event modelling
for conceptual design. Journal of Engineering Design, 22(3), 145-164.
Pimmler, T.U. and Eppinger, S.D., 1994. Integration analysis of product decompositions.
Proc. of the ASME Design Theory and Methodology Conference (DTM’94), 343–351.
Sage, A.P. and Lynch, C.L., 1998. Systems Integration and Architecting: An Overview of
Principles, Practices, and Perspectives. Systems Engineering, 1(3), 176–227.
Sharman, D. and Yassine, A., 2004. Characterizing complex product architectures. Systems
Engineering, 7(1), 35–60.
Simon, H.A., 1969. The sciences of the artificial. Englewood Cliffs: Prentice Hall.
Sosa, M., Eppinger, S.D. and Rowles, C., 2003. Identifying modular and integrative systems
and their impact on design team interactions. Journal of Mechanical Design, 125, 240–
252.
Steward, R.P. and Donald, V., 1981. The design structure system: a method for managing the
design of complex systems. IEEE Transactions on Engineering Management, 28(3), 71–
74.
Stone, R., Wood, K. and Crawford, R., 2000. A heuristic method for identifying modules for
product architectures. Design Studies, 21(1), 5–31.
Strawbridge, Z., McAdams, D. and Stone, R., 2002. A computational approach to conceptual
design. Proc. of ASME 2002 International Design Engineering Technical Conferences and
Design Theory and Methodology Conference (IDETC/DTM2002), Montreal, Canada,
September 29-October 2.
Suh, N.P., 1990. Principles of design. New York: Oxford University Press.
Suh, N.P., 2001. Axiomatic design: advances and applications. New York: Oxford University
Press.
Thebeau, R., 2001. Knowledge management of system interfaces and interactions for product
development processes. Thesis (MSc), Massachusetts Institute of Technology.
Ulrich, K.T. and Eppinger, S.D., 2008. Product design and development. 4th ed. New York:
McGraw-Hill.
Ulrich, K.T., 1995. The role of product architecture in the manufacturing firm. Research
Policy, 24(3), 419–440.
Whitfield, R.I., Smith, J.S. and Duffy, A.H.B., 2002. Identifying component modules. Proc. of
the 7th International Conference on Artificial Intelligence in Design (AID’02),
Cambridge, UK, July 15-17, 571–592.
Yu, T., Yassine, A. and Goldberg, D., 2003. Genetic algorithm for developing modular product

18
 architectures. Proc. of ASME 2003 International Design Engineering Technical
Conferences and Design Theory and Methodology Conference (IDETC/DTM2003),
Chicago, September 2-6.
Zimmerman, H., 1991. Fuzzy set theory and its applications. 2nd ed. Boston: Kluwer.

19