A toolbox for the design, planning and operation of manufacturing networks in a mass customisation environment

G Model JMSY-302; No. of Pages 13

ARTICLE IN PRESS Journal of Manufacturing Systems xxx (2014) xxx–xxx

Contents lists available at ScienceDirect

Journal of Manufacturing Systems journal homepage: www.elsevier.com/locate/jmansys

Technical Paper

A toolbox for the design, planning and operation of manufacturing networks in a mass customisation environment Dimitris Mourtzis ∗ , Michalis Doukas, Foivos Psarommatis Laboratory for Manufacturing Systems and Automation, University of Patras, 26500, Greece

a r t i c l e

i n f o

Article history: Received 7 October 2013 Received in revised form 7 April 2014 Accepted 8 June 2014 Available online xxx Keywords: Mass customisation Design Planning Manufacturing networks Complexity Metaheuristics

a b s t r a c t The task of design, planning and operation of manufacturing networks is becoming more and more challenging for companies, as globalisation, mass customisation and the turbulent economic landscape create demand volatility, uncertainties and high complexity. In this context, this paper investigates the performance of decentralised manufacturing networks through a set of methods developed into a software framework in a toolbox approach. The Tabu Search and Simulated Annealing metaheuristic methods are used together with an Artificial Intelligence method, called Intelligent Search Algorithm. A multi-criteria decision making procedure is carried out for the evaluation of the quality of alternative manufacturing network configurations using multiple conflicting criteria including dynamic complexity, reliability, cost, time, quality and environmental footprint. A comparison of the performance of each method based on the quality of the solutions that it provided is carried out. The statistical design of experiments robust engineering technique is used for the calibration of the adjustable parameters of the methods. Moreover, the impact of demand fluctuation to the operational performance of the alternative networks, expressed thorough a dynamic complexity indicator, is investigated through simulation. The developed framework is validated through a real life case, with data coming from the CNC machine building industry. © 2014 The Society of Manufacturing Engineers. Published by Elsevier Ltd. All rights reserved.

1. Introduction and motivation The contemporary manufacturing industry is characterised by immense competition, divergent regional markets, high demand volatility and heterogeneity. Moreover distinctive characteristics include the impact of strict environmental regulations and an increasing need for customised products throughout the globe [1]. Thus, properly configured and easily adaptable manufacturing networks are needed, which are capable of handling the complexity and enormity of the supply chain structures. These qualities are critical for companies in order to maintain their viability. To support the strategic level decision making process, this research work describes a set of optimisation methods that have been developed and are integrated in a toolbox approach, for the design and operation of highly complex manufacturing networks. These networks operate under demand fluctuations, economic and environmental constraints [58,60]. The unpredictability of the operational performance of the networks is expressed through a dynamic complexity indicator [49]. The complexity of alternative network configurations is examined through

∗ Corresponding author. Tel.: +30 2610 997262; fax: +30 2610 997744. E-mail address: [email protected] (D. Mourtzis).

discrete event simulation, where the demand profile is modelled so as to represent different stochastic product demands. The interarrival times and volume of orders is affected by changes in the profile representing seasonal, volatile, rapid increases and fully unpredictable demand scenarios. The differentiated demands are modelled through statistical distributions from which the simulator samples the inter-arrival times of the ordered batches. The problem of the design of multi-stage, multi-product manufacturing networks under multiple pre and post-conditions and constraints is investigated. The problem is of NP-Hard computational complexity. To demonstrate that a problem is in the class of NP-Hard complexity, it is common practice to depict that it is at least as hard as another proven NP-hard problem [2]. To demonstrate the computational complexity of the inventory routing and scheduling problem, Liu and Chen, referred to the inventory routing problem, which is NP-Hard. By including in their method the scheduling of tasks, the problem becomes strongly NP-Hard [3]. Similarly, the fixed-charge capacitated network design problem was characterised in another study as NP-Hard as it extended, complexity-wise, the proven manufacturing network design problem [4]. The problem under investigation in the presented research work can be classified as a counterpart of the Simple Plant Location (SPL) multi-criteria assignment problem, which is NP-hard. In addition to the SPL problem, the proposed approach models a

http://dx.doi.org/10.1016/j.jmsy.2014.06.004 0278-6125/© 2014 The Society of Manufacturing Engineers. Published by Elsevier Ltd. All rights reserved.

Please cite this article in press as: Mourtzis D, et al. A toolbox for the design, planning and operation of manufacturing networks in a mass customisation environment. J Manuf Syst (2014), http://dx.doi.org/10.1016/j.jmsy.2014.06.004

G Model JMSY-302; No. of Pages 13

ARTICLE IN PRESS D. Mourtzis et al. / Journal of Manufacturing Systems xxx (2014) xxx–xxx

2

multiple tier multi-product supply chain and simultaneously considers multiple conflicting criteria in order to assess the quality alternatives. Therefore, it can be considered as NP-Hard [5]. Moreover, the identification of Pareto optimal solutions which is faced in the described approach is NP-Hard [5,6]. Indicatively, the feasible alternative solutions in the examined manufacturing network configuration decision-making problem are calculated at 4 × 1017 . It is evident that for this magnitude of search spaces, exhaustive enumerative techniques are rendered useless due to computational resource constraints. In addition, strategic level decision-making cannot be accurately performed based solely on the experience and past knowledge of a supply chain manager [7] due to the enormity and dynamic nature of the problem.

2. State of the art The manufacturing landscape is nowadays more complex and dynamic than ever, due to rapid globalisation [8] and recent economic recession [1] among other reasons. The design of the manufacturing network of interacting companies and its operation are key strategic decisions for companies trying to endure competition. The less push and more pull business model followed by many modern industries in order to address the product personalisation requirements calls for immediate decisions. This is necessary even in a supply chain level, in order for companies to respond effectively to the ever-changing market needs [9]. Stable cooperation structures based on rigid alliances between companies are no longer viable due to various disruptions that can affect a supply chain [10]. Likewise, recent environmental directives and fluctuating gas prices consist of additional constraints when designing and managing supply chains [11]. The implications for the implementation of mass customisation are examined in [50,51]. Specifically, on a product level, a key enabler towards achieving customisation in production is product modularity and postponement strategies. The division of a product into separate modules/components, provides the means to achieve high product variety at low costs [52]. Modularity, as a variety enabler, can be achieved during different stages of product realisation from design to production, assembly, as well as during sales and usage. Two forms of postponement strategies can be found in the literature, namely form and time postponement. During the former, differentiation is moved downstream in the value chain and is performed on a component level by the suppliers, whereas in the latter, the production of commonly used components is performed normally, while customer specific features and components are added at the final stages of the production [51]. In order to address the introduced complexity due to high variety, the Generic Bill of Material Structures (GBOM) have been proposed for the representation of product families and all the varieties [53]. Moreover, the coordination between the independent enterprises that form the manufacturing network must be achieved in order to align their objectives towards a common goal and maintain global performance. Two main approaches have been proposed over the years for realising this coordination: centralised and decentralised planning methods and tools [58]. The latter are the most prominent, as centralised practices have often been criticised for their limited flexibility and restricted coordination efficiency. Coordinated decentralised planning has been acknowledged as a resolution towards cooperative responsive manufacturing enterprises [60]. The channels for achieving the coordination are generally: supply chain contracts, information technology, information sharing and joint decision-making. A decentralised game theoretic framework is proposed in [57] using negotiation-based models and multi-agent systems. The results from a case study supported game theory as a promising

methodology for coordination in distributed production planning. Another study proposed flexible coalition strategies and investigated the advantages of using integrated production planning and negotiation in e-marketplace bringing real advantages for both customers and suppliers [59]. An agent-based negotiation framework has been proposed in [61]. Utilising a branch-and-bound heuristic, the method handled the negotiations for allocating numerous orders to multiple members for supply chain formation. Another study depicted the advantages of information sharing [62]. Though simulation-based comparison, the supply chain costs were calculated for a model with full information sharing policy and were compared against a model with traditional information sharing policy, resulting in a maximum difference of 12.1% lower costs in favour of the first model. Finally, the SCOR model helps in evaluating and improving enterprise wide supply chain performance and management by allowing companies to: evaluate their own processes effectively and compare their performance with the performance of other partners, pursue their competitive advantages, utilise best practice information to prioritise their activities, quantify the benefits of what-if scenarios and identify the most suitable IT support tools for their business [63,64]. The connection between product and process design with supply chain decisions was investigated and the dependencies between these activities are highlighted in [12]. The supply chain design and redesign problem with minimal costs and demand satisfaction at the same time is tackled in [13]. The method also allows the identification of unnecessary actors in order for them to be eliminated from the supply chain typology. A mixed integer linear programming formulation is presented in [14] for the design and planning of closed-loop supply chains that include the phases of production, distribution and reverse logistics. The demand profiles used in the study modelled diverse conditions of market requirements through optimistic, pessimistic and realistic scenarios. Gumus et al. [15] calculated the optimal product flow between the factories, warehouses and distributors of a globalised supply chain network. The uncertainty of cost and capacity in the supply chain was tackled through a neuro-fuzzy approximation that calculated these variables. Multi-echelon decentralised manufacturing network structures were modelled in [16] and were compared to traditional centralised networks typologies, depicting their superiority for satisfying cost, time and environmental footprint objectives optimisation. Another study focused on the redesign of an existing network [17]. The redesign decisions are supported by a Tabu Search algorithm and comprise of the relocation of existing facilities to new sites under budget and time restrictions and the flow of commodities through the network among others. Another recent approach dealing with the integration of the planning of the production and distribution activities presented a capacitated plant that produced multiple products based on a steady market demand [18]. The approach included two Tabu Search variants, one creating solutions and storing them in a short-term memory and the other integrating path relinking to the first using a longer-term memory. In Subramanian et al. [19], a closed loop supply chain design problem formulated using Integer Linear Programming, was addressed through a constructive heuristic for acquiring high quality solutions. Good initial solutions were created through the Vogel’s approximation method and were fed to a priority based Simulated Annealing heuristic for accelerating the algorithm’s convergence. A genetic algorithm with a novel encoding mechanism and a new crossover operator was used for addressing the problem of multi-stage, multi-product, multiple objective supply network design optimisation [65]. Moreover, hybrid approaches that utilise combinations of optimisation methods are common. A hybrid metaheuristic combining a genetic algorithm with a local improvement search and a shortest path augmenting path method were proposed in [20]. The objective was to

Please cite this article in press as: Mourtzis D, et al. A toolbox for the design, planning and operation of manufacturing networks in a mass customisation environment. J Manuf Syst (2014), http://dx.doi.org/10.1016/j.jmsy.2014.06.004

G Model

ARTICLE IN PRESS

JMSY-302; No. of Pages 13

D. Mourtzis et al. / Journal of Manufacturing Systems xxx (2014) xxx–xxx

design reliable networks that are disrupted by natural disasters, power outages and other factors that are difficult to predict. Further to that point, a large number of recent publications deals with the emerging aspects of increasing complexity of manufacturing activities and the dynamic nature of supply chains [21]. The importance of managing the complexity in supply chains is indicated in [22]. Moreover, a recent study [23] depicted that lower manufacturing network complexity is associated with reduced costs and overall network performance. Therefore, complexity, in a manufacturing network level, should be considered as a cost criterion, i.e. one that has to be minimised. In particular, dynamic complexity is related to the uncertainty of the system’s behaviour for a specific time period and deals with the probability of the system to be in control [24–27]. A comprehensive review on complexity measurement approaches for manufacturing was carried out in [28] and [29]. Dynamic complexity has been studied through approaches based on information theory [30], time series analysis [31] and axiomatic theory [32] among other. In this context, from an economic point of view, the important industrial sector of CNC machine building is facing challenges. As far as Europe is concerned, the CNC sector accounts for 158,000 jobs spread over 1474 companies with a worth of 17,512 billion D. Thus, it makes a key contribution to the economy and balance of payments [32]. For producing highly customised products, flexible and configurable machines are required for providing the capability to perform diversified manufacturing tasks in the concept of mass customisation. The proposed method aims to the timely identification of optimum or near optimum manufacturing network configurations for the production of heavily customised CNC machines. Both structural (resource cycle time, number and size of buffers, Mean Time Before Failure, Mean Time To Repair) and dynamic characteristics (flowtime timeseries evolution, machine utilisation, ratio of product demand to manufacturing capacity) of the system are captured, when examining the performance of different manufacturing network configurations under diversified market demands. A dynamic complexity measure is incorporated in the decision-making process, which is novel for manufacturing network design problems. Moreover, the network’s reliability has not been sufficiently used in quantitative terms in the literature as an indicator during the design phase, along with criteria of cost, time, quality and environmental footprint. Finally, the applicability of the approach is validated via a real life case study with data acquired from a CNC machine building industry.

3. Description of the proposed approach The input parameters of the method include the definition of the plants of the OEM, the suppliers and the dealers. Each plant is characterised by its location, the pool of resources that it includes and its manufacturing capabilities. Each resource has specific capabilities; it can produce one or multiple product components. Moreover, for each specific resource, operational parameters and values are defined (cycle time, cost and energy consumption). The product structure comprises of a number of standard components that are common to all the variants and a number of customisable options that are built and assembled based on the customer requirements. The Bill of Processes likewise includes operations that are standard and other, special tasks that are required by the customised features of the product. In contrast to the traditional centralised approach, in the presented research work, the customised components can be assembled on the product even by a supplier or a dealer whenever possible, in a decentralised manner [58]. This way, not all components have to be collected by the OEM to perform the final assembly. Finally, as input parameters, the weights of the criteria and the

3

optimisation method are required. The criteria weights represent the design objectives and include production and transportation cost, lead time, environmental impact, quality, reliability and dynamic complexity [49]. Afterwards, an optimisation method is selected, the output of which comprises of a high quality manufacturing network configuration together with the calculated criteria values that are visualised and stored for future use (Fig. 1). 4. Decision-making framework The three methods integrated in the design and planning of manufacturing networks toolbox are described hereafter. The methods used are the widely used Tabu Search (TS) and Simulated Annealing (SA) and the game theoretic artificial intelligence method called Intelligent Search Algorithm (ISA). 4.1. Optimisation algorithms 4.1.1. Intelligent Search Algorithm The Intelligent Search Algorithm (ISA) is an artificial intelligence algorithm which uses three adjustable control parameters. These are the Selected Number of Alternatives (SNA), the Decision Horizon (DH) and the Sampling Rate (SR). The SNA parameter dictates how many alternatives are considered from the pool of alternatives, the DH creates the layers of the search, whereas the SR determines the number of permutations (branches) to be randomly performed in order to fill the nodes after the layers of the tree are created by the DH. The algorithm is described extensively in [33–36]. 4.1.2. Tabu Search Tabu Search (TS) is a metaheuristic local search algorithm that can be used for solving combinatorial optimisation problems [37]. TS uses a local neighbourhood search procedure to iteratively move from one potential solution x to an improved solution x’ in the neighbourhood of x, until some stopping criterion has been satisfied. A tabu list L, of some maximum length l, of candidate solutions that have been searched so far is stored and constantly updated. Whenever a new candidate solution is adopted, it goes in the tabu list. If the number of rows of the tabu list exceeds l, the oldest candidate solution is removed and it is no longer a taboo to reconsider. The implemented TS algorithm is based on the pseudo-code provided in [39]. 4.1.3. Simulated Annealing Simulated Annealing (SA) is a generic probabilistic metaheuristic for the global optimisation problem of locating a good approximation to the global optimum of a given function in a large search space [38]. The SA starts with a randomly generated initial feasible solution S that is generated based on the manufacturing capabilities of the supply chain partners. This S is changed following a hill climbing notion. The performance of the alternative is represented by its utility value; the higher the utility, the better the alternative. However, the difference with simple hill climbing is when SA makes the decision on when to replace S with R, its newly tweaked offspring. More specifically, if R is better than S, SA will always replace S with R. But if R is worse than S, it may still replace S with a certain probability P(t,R,S): R(t, R, S) = e

Utility value(R)−Utility value(S) t

,

where t = 0

(1)

Thus, the algorithm sometimes goes downhill. The fraction in the exponent is negative because R has a lower utility values than S. First, if R is much worse than S, the fraction is larger, and so the probability is close to 0. If R is very close to S, the probability is close to 1. Thus if R is not much worse than S, SA will still select R with a reasonable probability. Second, the tuneable parameter

Please cite this article in press as: Mourtzis D, et al. A toolbox for the design, planning and operation of manufacturing networks in a mass customisation environment. J Manuf Syst (2014), http://dx.doi.org/10.1016/j.jmsy.2014.06.004

G Model JMSY-302; No. of Pages 13

ARTICLE IN PRESS D. Mourtzis et al. / Journal of Manufacturing Systems xxx (2014) xxx–xxx

4

Fig. 1. Overview of the framework.

t, if close to 0, the fraction is again a large number, and so the probability is close to 0. If t is a large number, the probability is close to 1. Initially t is set to a high number, which causes the algorithm to move to explore new neighbourhoods by adopting every newly created solution regardless of how good it is. Afterwards, t decreases slowly, eventually becoming 0, at which point the algorithm is doing nothing more than plain hill-climbing. The rate at which t is decreased is called the algorithm’s schedule. The longer t is stretched out the schedule, the longer the algorithm resembles a random walk and the more exploration it does. The implemented SA algorithm is based on the pseudo-code provided in [39]. 4.2. Decision making objectives The decision-making objectives that reflect the design goals are presented hereafter. These objectives represent the criteria of the decision-making process. They comprise of a number of classical indicators (cost, time and quality) and other that are becoming more and more important recently (environmental impact, reliability and dynamic complexity). The method is also able to consider additional criteria with minimal implementation effort on the software framework. 1. Production and transportation Cost (Eq. (2)). This criterion encapsulates the cost required for the production of the standard and customisable components as well as the costs required for the transportation of the parts, subassemblies with the supply chain until the final product reaches the customer [33,49]. 2. Lead time (Eq. (3)). The lead time criterion is calculated as the time duration between the points when an order is placed to the point that it is actually available for satisfying customer demand. It includes the processing time of work pieces, the waiting time in queues, the setup time of the machines and the time required for the transportations within the manufacturing network [40]. 3. Energy consumption (Eq. (4)). It is calculated as the sum of the energy consumed in transportation activities and by machines that process parts based on energy consumption (Watts)

4.

5.

6.

7.

specifications and processing time of each resource for specific tasks [41,42]. CO2 emissions (Eq. (5)). The CO2 emissions are calculated as the sum of the total distance covered by trucks that carry parts and assemblies within the network based on their CO2 emissions per kilometre (km) [41,42]. Quality (Eq. (4)). This indicator is the mean total quality of the manufacturing network partners that are selected in any given alternative. It takes values within the range [0–100]. It is calculated based on historical data provided by the OEM of the case study and encapsulates the observed quality of parts and services that the supplier/dealer provides and on their respect of due dates [33]. Reliability (Eqs. (7) and (8)). This indicator expresses the total network reliability, where s represents a serial and p a parallel resource [54]. Dynamic complexity. The dynamic complexity (CLZ ) expresses the unpredictability of the flowtime timeseries. CLZ is calculated through a Lempel–Ziv analysis of the timeseries obtained through simulation of each alternative manufacturing network design. Firstly, the average value of the flowtime values of the timeseries obtained through simulation, is calculated. The values of the flowtime that are equal or exceed the average value are encoded with the value of 1, and 0 otherwise. The string of 0s and 1s is used for the Lempel–Ziv analysis of the Kolmogorov complexity. The method for calculating dynamic complexity is presented in Section 6.2.3 [28,49].

4.2.1. Mathematical formulation of the problem The problem considered in this paper has been obtained from a company that builds CNC robotic laser cutting machines based on customer demand and preferences. The company offers a set of standard laser cutting machines that can be customised by the customers according to their specific requests. For instance, a mass producing customer can order the mass production setup that includes automated loading and unloading robots and automatic pallet changers, whereas a customer from the aerospace industry will select the configuration with a lateral entrance for a larger

Please cite this article in press as: Mourtzis D, et al. A toolbox for the design, planning and operation of manufacturing networks in a mass customisation environment. J Manuf Syst (2014), http://dx.doi.org/10.1016/j.jmsy.2014.06.004

G Model

ARTICLE IN PRESS

JMSY-302; No. of Pages 13

D. Mourtzis et al. / Journal of Manufacturing Systems xxx (2014) xxx–xxx

automatic pallet changer due to the large size of the components that need processing. The problem is a multi-stage manufacturing network design problem for single multi-component products. The objectives are the minimisation of: the production cost, lead time, CO2 emissions and energy consumption and the maximisation of: product quality and network reliability. The assumptions of the model are the following: (1) the number and the capabilities of suppliers, OEM plants and dealers are predetermined, (2) the production cost, lead time, quality and reliability of each supplier, OEM and dealer are fixed, (3) the customers and their locations are known, and (4) the product demand is predetermined. The indices used are the following: n is the set of tasks (n ∈ N) i is the set of alternative configurations (i ∈ I) p, m are the sets of partners’, where p represents the sources (p ∈ P) and m represents the depots (m ∈ M) j is the set of criteria (j ∈ J) s is the set of serial resources (s ∈ S) l is the set of parallel resources (l ∈ L)

min CO =

The decision variables are the following: Cpn is the Boolean flag of whether Partner p performs task n (1 if yes, 0 is no), and Cpm is the Boolean flag of whether Partner p transports parts to partner m (1 if yes, 0 is no). The Objective Functions for criteria calculation of the model are the following: minC (2) represents the minimisation of the total cost for the production and transportation of the customised product in the manufacturing network. minT (3) represents the minimisation of the lead time for the production and transportation of the customised product in the manufacturing network. minEC (4) and minCO (5) express the minimisation of the energy consumption and the CO2 emissions respectively. maxQ (6) represents the maximisation of the quality index of the product, maxRstot (7) and maxRltot (8) represent the maximisation of reliability. Since, objective functions (7) and (8) are non-linear, the problem is formulated into a Non-Linear Programming (NLP) model. min C =

N 

PCpn cpn +

M 

n

min T =

N 

LTpn cpn +

M 

n

min EC =

TCpm cpm

(2)

TTpm cpm

(3)

m

m

N  n

EPpn cpn +

M  m

ETpm cpm

(4)

COpm cpm

(5)

m

p P

max Q =

Qp

p

max Rstot =

(6)

cpn

S 

Rs ,

for serial resources

(7)

s

max Rltot = 1 −

L 

(1 − Rl ),

for parallel resources

(8)

l

The objective functions are conflicting; increasing the quality of the alternative leads to an increase of its cost. Therefore, a Pareto optimal solution is deriving from the algorithm. The overall performance of a manufacturing network alternative is represented by the utility value, which is the sum of products of the normalised objectives multiplied with the user-defined weight factor for each objective. The utility value is used for selecting the alternatives; the higher the utility value, the higher the performance of the network. Considering that the objective functions comprise the criteria of the decision-making process, the utility value is expressed by (9):

The problem parameters are the following: PCpn is the production cost for partner p to perform task n (Euros) TCpm is the transportation cost between partners p, m (Euros) LTpn is the lead time for partner p to perform task n (Hours) TTpm is the transportation time between partners p, m (Hours) COpm is the CO2 emissions for the transportation between p, m (gr of CO2 ) ETpm is the energy consumption for the transportation between p, m (Joules) EPpn is the energy required by partner p to perform task n (Joules) Qp is the quality of partner p Qitot is the total quality of alternative i Ui is the utility value of alternative i Wj assigned weight factor for criterion j

M 

5

max U =

J 

wj cij

(9)

j=1

The Constraints of the model are the following: 0 ≤ Qp,m ≤ 100 restriction of quality value of partner p and m 0 ≤Q itot ≤ 100 restriction for the total quality for the alternative i 0 ≤ Rp,m ≤ 100 is the restriction of the reliability value of partners p and m 0 ≤ Ritot ≤ 100 is the restriction for the total quality of alternative i



Cnp = 1∀n

(10)

Cmp = 1∀m

(11)

p

 p

Cpn = {0, 1}∀P, N

(12)

Cpm = {0, 1}∀P, N

(13)

Constraint (10) represents the unique assignment of a task to a partner, (11) represents the unique assignment of a transportation task between two partners, (12) and (13) impose the integrality restriction on the decision variables Cpn and Cpm . 4.2.2. Normalisation of the criteria values The problem under investigation is formally a Multi-Attribute Decision Making problem (MADM) [46]. Some of the considered criteria need to be minimised and other need to be maximised. Thus, a normalisation of the obtained criteria values is required to overcome their conflicting nature and their different units of measures [34,43,44]. Cost criteria such us production cost and lead time, have to be minimised and benefit criteria such us quality and reliability have to be maximised. Equation benefit is used for the normalisation of benefit criteria and equation cost is used for the normalisation of the cost criteria (Fig. 2), where Cij represents the consequence value of alternative i with respect to criterion j, and Cˆ j is the normalised value of Cj . The decision making steps are depicted in Fig. 2 below. The cardinal preference (utility value) is calculated using a sum of weighted criteria (Wj ) normalised to the sum of one based on the Simple Additive Weighting (SAW) method [45,46].

Please cite this article in press as: Mourtzis D, et al. A toolbox for the design, planning and operation of manufacturing networks in a mass customisation environment. J Manuf Syst (2014), http://dx.doi.org/10.1016/j.jmsy.2014.06.004

G Model JMSY-302; No. of Pages 13

ARTICLE IN PRESS D. Mourtzis et al. / Journal of Manufacturing Systems xxx (2014) xxx–xxx

6

Fig. 2. Decision-making steps.

4.3. Statistical design of experiments A four-phase statistical design of experiments (SDoE) based on the Taguchi methods [47,48] has been carried out for the calibration of the adjustable parameters of the three search methods (Fig. 3). The first step of conducting the SDoE is to determine the factors that are to be investigated and the number of levels that each factor has. Afterwards, the degrees of freedom have to be calculated in order to select an appropriate orthogonal array. The degrees of freedom of the problem define the minimum number of experiments that have to be conducted in order for the result of the SDoE to be valid. The selected orthogonal array should have more or equal rows than the number of degrees of freedom. The Analysis of Means (ANOM) is used for the determination of the optimum factor levels. The determination of the optimum factor levels is revealed from the selection as optimum level for each factor the level with the highest effect. After the determination of the optimum level for each

factor the influence of the observed value of each factor can be calculated through a statistics hypothesis testing using an Analysis of Variance (ANOVA). The final step is the validation of the results. The additive model is used to predict the value of U under the optimum conditions, denoted by (Uopt ) as per equation (14). (Uopt ) = m +

K 

(fkj − m)

(14)

k=1

where, m is the overall mean of the utility value, k is the observed factors and fkj is the effect of each factor under optimum conditions. If the sum of squares of some factors is small, the corresponding improvements in the prediction of (Uopt ) under optimum conditions is not included because these terms are included as errors. If the contribution from all factors is accounted, it can be shown that the predicted improvement in (Uopt ) exceeds the actual improvement and the prediction would be biased on the higher side. By

Fig. 3. Phases of the statistical design of experiments.

Please cite this article in press as: Mourtzis D, et al. A toolbox for the design, planning and operation of manufacturing networks in a mass customisation environment. J Manuf Syst (2014), http://dx.doi.org/10.1016/j.jmsy.2014.06.004

G Model

ARTICLE IN PRESS

JMSY-302; No. of Pages 13

D. Mourtzis et al. / Journal of Manufacturing Systems xxx (2014) xxx–xxx

7

Fig. 4. Screenshots of the developed DPMN© software toolbox.

ignoring the contribution from factors with small sum of squares, this bias is reduced. The next step is to determine the variance of the prediction error so that the closeness of the observed Uopt to the predicted (Uopt ) can be judged. The prediction error, which is the difference between the observed Uopt and the predicted (Uopt ) , has two independent components. The first one is the error in the prediction of (Uopt ) caused by the errors in the estimates of m and mSNA5 and the second is the repetition error of an experiment. Because these components are independent, the variance of the prediction error is the sum of their respective variances. The prediction error variance is given by (15): 2 pred =

1 n0

× e2 +

1 nr

× e2

(15)

where n0 is the equivalent sample size for the Uopt estimation and is given from Eq. (16): 1 1  = + n0 n K

k=1



1 1 − nkj n

(16)

where, n is the number of rows in the experiment matrix (selected orthogonal array) and nkj is the number of times level j of factor k was repeated in the experiment matrix. The second component of Eq. (16) is now considered. nr is the number of experiments conducted with the optimum conditions and  e is the estimation of error variance and its value (17): e2

sum of squares due to error = Error variance = degrees of freedom for error

(17)

The prediction error is the difference between the observed Uopt and the predicted (Uopt ) (18). Prediction error = (Uopt ) − Uopt

(18)

In order for the results of the statistical design to be valid the prediction error should be inside the limits of two standard deviations of the error variance.

5. Software tool development The proposed methodology is implemented into a web-based software toolbox for the Design and Planning of Manufacturing Networks, namely the DPMN©, where a digital simulator is also integrated (Fig. 4). The software tool is highly modular and flexible as it can be used as a standalone desktop based module, or over web, either as library or through web-services’ based communication. The algorithms for the generation and evaluation of the alternative manufacturing network configurations are programmed using the JAVATM framework, following the Software as a Service (SaaS) architectural pattern. The integrated simulator exchanges data through customised XML files generated on the fly. Moreover, the Google Maps API has been integrated in the software for the accurate calculation of the distances between the plants of the manufacturing network. In cases were transatlantic transportations are required, a coordinate-based distance calculation algorithm is utilised with appropriate correction factors based on the origin and destination locations. The data model is implemented using the MySQL relational database. The experiments were performed using a typical IntelTM i7 3.4 GHz powered processor on a workstation with 8GBs of RAM. 6. Industrial case study from the CNC machine building sector The customised product used in the pilot case is a robotic laser cutting machine that is produced over a global network of cooperating suppliers, dealers and OEM plants. The data for the product, processes and resources are acquired from a European machine building company. The robotic laser cutting machine can be configured by the customer through the addition of features (photoelectric protective cells, robotic loading arm, pallet changer, etc.). The machine can be customised in order to serve different production requirements, from configurations suited to mass production, to specific setups for aerospace industry constructions.

Please cite this article in press as: Mourtzis D, et al. A toolbox for the design, planning and operation of manufacturing networks in a mass customisation environment. J Manuf Syst (2014), http://dx.doi.org/10.1016/j.jmsy.2014.06.004

G Model

ARTICLE IN PRESS

JMSY-302; No. of Pages 13 8

D. Mourtzis et al. / Journal of Manufacturing Systems xxx (2014) xxx–xxx

Fig. 5. Bill of materials and bill of processes of the laser cutting machine.

However, the addition of features not included in the standard version of the machine leads to an increase in the number of alternatives. The selected configuration that is used for the experimentation is the so-called C3 variant. This configuration consists of nine basic components and four optional, customisable ones. The Bill of Materials (BoM) that includes the various components of the machine is depicted in Fig. 5. Moreover, the Bill of Processes (BoP) that describes the sequence of operations that must be performed for assembling the final CNC machine is included. The arrows depict an assembly task, i.e. addition of a component to a sub-assembly. 6.1. Experimenting methodology and results The procedure for performing the experiments is described henceforth. Initially, an SDoE for each method is conducted, where25 sets of experiments are performed, as mentioned in the previous chapter. Ten individual runs were conducted for each experiment set as defined by the selected orthogonal array. This is done to improve the accuracy of the results and to overcome the inherent randomness of creating alternatives in all three stochastic methods. In the context of SDoE, 750 experiments were conducted in total. The optimum set of parameters was identified that way for each method. Using the optimum parameter values, an additional set of ten individual runs were performed. Finally, the ten best network configurations identified by each method with the optimum

set of parameters (3 × 10 = 30 alternative configurations) were fed in a simulation software. The configurations were simulated under four different demand profiles. Thus, an additional number of120 simulation experiments was carried out. For the Intelligent Search Algorithm (ISA) there are three control parameters, namely the Selected Number of Alternatives (SNA), the Depth Horizon (DH) and the Sampling Rate (SR). Tabu Search (TS) has two control parameters, namely, the length of tabu list and number of iterations. Finally, Simulated Annealing (SA) has five control parameters, which are: the initial temperature, the cooling factor, the maximum number of accepts, the maximum number of rejects and the maximum number of iterations. Each of the factors above has five levels. Table 1 shows the factors and their levels for each method. The selection of the values here is performed empirically and based on the magnitude of the current problem modelling, i.e. number of components and available resources. Afterwards, the degrees of freedom for each method are calculated separately in order to select the appropriate orthogonal array, which will define the experiments that have to be conducted. The degrees of freedom for the ISA are 1 + 3 × (5 − 1) = 13, for the TS are 1 + 2 × (5 − 1) = 9 and for the SA are 1 + 5 × (5 − 1) = 21. Thus, the orthogonal array L25 is selected for all the cases due to the fact that it fits to all the cases. The results from the ten individual runs of each algorithm are presented in Table 2. Indicatively, for the ISA, the correlation of the SNA, DH and SR factors with the respective utility value is depicted in the

Please cite this article in press as: Mourtzis D, et al. A toolbox for the design, planning and operation of manufacturing networks in a mass customisation environment. J Manuf Syst (2014), http://dx.doi.org/10.1016/j.jmsy.2014.06.004

G Model

ARTICLE IN PRESS

JMSY-302; No. of Pages 13

D. Mourtzis et al. / Journal of Manufacturing Systems xxx (2014) xxx–xxx

9

Table 1 The selected factor values for the ISA, TS and SA methods. Method

Factors

Levels 1

2

3

4

5

1 1 1

10 3 10

100 5 50

1000 7 100

Tabu list length Max number of iterations

10 50

20 100

50 200

100 500

200 1000

Initial temperature Cooling factor Max number of accepts Max number of rejects Max number of iterations

1 0.8 20 50 50

2 0.85 40 100 100

5 0.9 60 150 250

10 0.95 80 250 500

20 0.99 100 500 1000

ISA

SNA DH SR

TS SA

10,000 10 1000

Fig. 6. Colour map of the utility value for different combinations of the SNA, DH and SR factors.

colour map of Fig. 6. The points of the surface are combinations of the 3-tuple (SNA, SR, DH) with values obtained from Table 1. Based on the results of experiments for the ISA, the ANOM and ANOVA tables can be formed, as well as the optimum set of parameters is identified. The optimum set of parameters for the ISA is SNA = 10,000, DH = 10 and SR = 50. In the same way, the optimum values for the TS method are: Tabu list length = 5 and max number of runs = 3 and for the SA: initial temperature = 2, cooling factor = 0.9, max number of accepts = 60, max number of rejects = 500 and max number of iterations = 50. After the analysis of the results of the conducted experiments for the three methods and the creation of the ANOM and ANOVA diagrams, the next step is to validate the SDoE results. The results’ validation is performed using a statistical method [30] as explained above. For the results of the SDoE to be valid the following condition (19) must be met: Prediction Error ≤ 2 × Prediction Error standard Deviation

(19)

Table 2 Statistical design of experiments – results validation. Method

Prediction error

Prediction error 2 ) variance (pred

2 ×  pred

ISA TS SA

0.03735 0.01421 0.08859

0.00308 0.00119 0.00332

0.1109 0.0691 0.1152

Based on Table 2 and Eq. (17) the validity of the SDoE results is proven. Thus, the results of the experiments for the optimum parameter set for each method can be used. The proposed toolbox includes the methods Tabu Search, Simulated Annealing and Intelligent Search Algorithm. These algorithms can provide a timely identification of a near optimum alternative manufacturing network configuration that satisfies the design and planning objectives of the system and the planner. These objectives are reflected by the appropriate adjustment of the variables and factors of the three search methods. 6.2. Comparison of the methods Through the SDoE, the optimum set of parameters for each search method were identified and were used for conducting ten individual runs of each method and calculating the average value for each criterion. The results are included in Table 3. To enable the comparison of the three methods, the criteria values are normalised using Eq. (20). Norm. val. (Cost crit.) =

min(xi ) xi

=

xi max(xi )

Norm. val. (Benefit crit.) (20)

The higher the normalised value is, the better the performance for both cost and benefit criteria. From Fig. 7, it is observed that TS

Please cite this article in press as: Mourtzis D, et al. A toolbox for the design, planning and operation of manufacturing networks in a mass customisation environment. J Manuf Syst (2014), http://dx.doi.org/10.1016/j.jmsy.2014.06.004

G Model

ARTICLE IN PRESS

JMSY-302; No. of Pages 13

D. Mourtzis et al. / Journal of Manufacturing Systems xxx (2014) xxx–xxx

10

Table 3 Criteria values for each method under optimum conditions. Criteria

Table 4 Objective function results.

Method

Set

ISA

TS

SA

Production cost (D) Lead time (minutes) CO2 emissions (grams) Energy consumption (Joules) Quality Reliability

347,379.94 59,321.90 2,823,760.00 852,330.24 84.37 0.68

259,968.06 42,195.81 2,807,680.00 745,924.32 70.92 0.44

433,956.64 141,139.93 69,642,520.00 2,517,440.48 76.19 0.53

Utility value Computation time (s)

0.85 187.56

0.71 1.64

0.25 13.66

Fig. 7. Performance of each method under optimum conditions.

outperforms the other two methods in four out of six criteria. The relative difference in the value of production cost is 25.16% lower than that of the ISA and 40.09% than that of the SA. Accordingly, the TS method yielded solutions with the smallest values for the lead time criterion, calculated at 42,195.81 min. This value is 28.86% lower than the value obtained with the ISA and 70.10% lower than the value provided by the SA. Analogous is the trend for the CO2 emission and the energy consumption criteria. The ISA identified solutions with better values for the quality and the reliability criteria. ISA outperformed SA by 9.69% and TS by 15.94%. For reliability, the solutions of ISA were 22.06% higher than the value of the SA and 35.47% higher than the value obtained by the TS method. Moreover, the diagram of Fig. 8 compares the utility values obtained by the three search methods and the computation time required to acquire that solutions. The utility value obtained by the ISA is the higher (0.8515). This value is 17.14% higher than the utility value provided by the TS and 70.75% better than the solution of the SA method. Further to that, the ISA may revealed the best utility value but also required the longest time to identify it. At 187.56 s it required 92.71% more time to provide the solution when compared to the SA method and 99.12% more time than TS, the latter requiring only 1.64 s to yield a high quality alternative.

1 2 3

Weights

Method

Wu

Wct

ISA

TS

SA

0.9 0.5 0.1

0.1 0.5 0.9

0.8710 0.4887 0.1064

0.7867 0.8813 0.9760

0.2073 0.5310 0.8548

assessment of the methods [48]. The objective function  is calculated by Eq. (21): Ui − Umin  = WU × + WCT × Umax − Umin



CTi − CTmin 1− CTmax − CTmin



(21)

where WU , the weight assigned to the utility value, Wct , the weight assigned to the computation time, Ui , the mean utility value of the experiment set i, Umin , the minimum utility value of all three methods, Umax , the maximum utility value of all three methods, CTi , the mean value of the computation times, CTmin , the minimum computation time of all three methods, CTmax , the maximum computation time of all three methods. The calculation of the objective function is performed for three sets of weight factors. Table 4 contains the results of the objective function calculation. The results are graphically depicted in Fig. 9. For the first set of weight factors (Set 1), the ISA performs better, providing a value for the objective function, which is 0.871, meaning 9.68% higher than the value of the TS and 76.19% higher than the value of SA. The second set of weight factors considers equally both the utility value and the computation time with weights defined at 0.5. TS has the best performance with this set, with a value 0.8813, which is 39.74% higher than the value from the SA, followed by the last ISA value with a relative difference of 47.64%. The final set of weighs emphasises on the required time to obtain the solution caring less on how good the solution is. For this set, TS revealed the best value, i.e. 0.9759, due to the fact that it required significantly less computation time. The relative difference from SA was 12.41%, followed by ISA, whose performance was 89.09% lower than the value of the Tabu Search. 6.2.2. Operational performance comparison In this section, the dynamic operational performance of the networks is investigated through their simulation over the demand profiles included in Fig. 10. The demand profiles represent four typical cases of demand fluctuation imposed to the manufacturing system by the market. Thus, the seasonal demand, modelled through a sinusoidal curve, represents a varying demand based on the seasons. The values for the volume of orders of the seasonal

6.2.1. Combined objective function In order to compare the three methods considering both the utility value and the computation time, the usage of the objective function is required, which quantitatively provides an all-around

Fig. 8. Comparison of the utility value and the computation time.

Fig. 9. Objective function results comparison.

Please cite this article in press as: Mourtzis D, et al. A toolbox for the design, planning and operation of manufacturing networks in a mass customisation environment. J Manuf Syst (2014), http://dx.doi.org/10.1016/j.jmsy.2014.06.004

G Model JMSY-302; No. of Pages 13

ARTICLE IN PRESS D. Mourtzis et al. / Journal of Manufacturing Systems xxx (2014) xxx–xxx

11

Fig. 10. The four demand profiles.

demand are calculated from the equation y = 4 + 3sin(0.5x + /3). Likewise, the rapid increase in demand, represents a situation where an initial low request for products is followed by an exponential increase in the ordered volumes and is modelled through the exponential function y = 0.01x2 . Additionally, the volatile demand, simulates the nowadays common, unpredictable market landscape. The ordered volume values depicted in the diagram below are sampled from a normal distribution with a mean  = 2.5 and a variance  = 1.5. Finally, the last demand profile against which the manufacturing networks are simulated, represents a fully unpredictable volume request from the environment and is used for benchmarking reasons. This purely stochastic demand derives from sampling the uniform distribution (0, 5). The simulation period is for 52 weeks, which translates to 6240 working hours based on three 8-hour shifts per day, 5 days per week. The experiments were performed for the 10 alternative network configurations identified by the ISA, TS and SA using the optimum set of parameters indicated by the SDoE. A simulation engine was integrated (as described in Section 5) and was used for modelling and executing discrete event simulation experiments. Each network was simulated four times, each time using a different demand profile as input. The results are exported in a text file, are stored locally and used for the timeseries analysis during the calculation of the dynamic complexity.

6.2.3. Dynamic complexity calculation The dynamic complexity (CLZ ) is expressed as the unpredictability of the timeseries of the flowtime indicator [55]. CLZ is calculated through a Lempel–Ziv algorithmic analysis of the Kolmogorov complexity of the timeseries [56]. The timeseries is obtained through simulation as described in Section 6.2.2. Firstly, the average value of the flowtime values of the timeseries obtained through simulation, is calculated. The values of the flowtime that are equal or exceed the average value are encoded with the value of 1, and 0 otherwise.

The string of 0s and 1s is used for the analysis of the complexity, where the algorithm identifies the patterns that exist in the string to estimate the complexity. A simple example that depicts the workflow of the algorithm is presented hereafter, examining the complexity of the simple string S = 0011. The following variables are considered: • S: the string of 0 and 1 (which is produced from the flow-time series) • c: the number of patterns (words) in the vocabulary • Q: the 0–1 string The algorithm starts with the first element of S: Q = S(1) = 0. The value 0 is not in the vocabulary, thus, it is added to the vocabulary and the number of words (c) is increased by 1, i.e. c = 1. Then the algorithm takes the second element Q = S(2) = 0. The value 0 has just been added in the vocabulary, therefore, the algorithm continues with the element Q = S(2)S(3) and checks if S(2)S(3) = 01 is in the vocabulary. The substring 01 is not in the vocabulary, thus, the number of words is increased by one, i.e. c = 2. After Q is renewed Q = S(4) and checks again if Q is in the vocabulary. The value 1 is not in the vocabulary. The entire string S has been processed resulting in a total complexity c = 3. Table 5 contains the complexity values of the experiments for the three methods. Based on the values of Table 5, the diagram of Fig. 11 is generated. It is observed that the complexity values of the seasonal and rapid increase demand profile are low. This can be attributed to the fact that manufacturing networks can handle more effectively predefined market demands. Complexity increases from 0.2 to 0.4 for deterministic demands (seasonal and rapid increase), however, to values of 1.0–1.2 for the volatile demands (Gauss and uniform distributions). The unpredictability of the input demand to the system creates high complexity that the system is not capable to handle.

Please cite this article in press as: Mourtzis D, et al. A toolbox for the design, planning and operation of manufacturing networks in a mass customisation environment. J Manuf Syst (2014), http://dx.doi.org/10.1016/j.jmsy.2014.06.004

G Model

ARTICLE IN PRESS

JMSY-302; No. of Pages 13

D. Mourtzis et al. / Journal of Manufacturing Systems xxx (2014) xxx–xxx

12

Table 5 Complexity values for the Intelligent Search Algorithm, Tabu Search and Simulated Annealing methods. Exp. set

Intelligent Search Algorithm

Tabu Search

Simulated Annealing

Seasonal demand

Rapid demand increase

Volatile demand

Benchmark

Seasonal demand

Rapid demand increase

Volatile demand

Benchmark

Seasonal demand

Rapid demand increase

Volatile demand

Benchmark

1 2 3 4 5 6 7 8 9 10

0.3197 0.3111 0.1823 0.1774 0.3289 0.2127 0.2836 0.1823 0.1823 0.1823

0.3386 0.3289 0.2699 0.1962 0.3602 0.3101 0.2304 0.2025 0.2025 0.2025

0.4385 0.3111 1.0480 1.0940 0.3386 0.4255 0.2127 0.6165 1.1258 1.0633

0.4385 0.5256 1.0790 0.9521 0.5644 0.7151 0.2836 0.8886 0.9812 1.1597

0.2058 0.2836 0.1849 0.1849 0.1774 0.2058 0.1798 0.2836 0.2398 0.1823

0.2283 0.3072 0.2058 0.2058 0.1993 0.2242 0.1993 0.2283 0.2658 0.2025

0.5578 0.3545 1.0007 1.0633 1.0940 0.3486 1.1548 0.3545 1.0633 1.0480

0.7791 0.5673 1.0314 1.0471 1.1424 0.6284 1.0161 0.5671 0.9812 1.0952

0.3491 0.3289 0.375 0.3111 0.2885 0.1876 0.2371 0.4922 0.3289 0.3750

0.3722 0.4654 0.4039 0.3386 0.2371 0.2092 0.2686 0.4148 0.3661 0.2991

0.3491 0.4385 0.4802 0.3197 0.2885 1.0308 0.5641 0.4922 0.3386 0.3750

0.5818 0.4385 0.2881 0.5329 0.5872 1.1295 0.3223 0.2953 0.4515 0.2813

Mean St. Dev.

0.2363 0.0659

0.2642 0.0652

0.6674 0.4404

0.7588 0.2962

0.2128 0.0417

0.2266 0.0349

0.8040 0.3515

0.8850 0.2269

0.3273 0.0828

0.3375 0.0827

0.4677 0.2164

0.4908 0.545

Fig. 11. Complexity values (CLZ ) for the ISA, TS and SA methods, for the four demand profiles.

The standard deviation is also significantly lower between the experiment sets for seasonal and rapid increases in the demand taking values between 0.03493 and 0.08276, and becomes considerable for the volatile and benchmarking profiles taking values between 0.21637 and 0.44043. The mean complexity value for the benchmark demand (uniform) for the networks identified by the TS was 80.31% higher than the corresponding value of the networks identified by SA and 16.63% higher than those of ISA. Finally, it is observed that a number of complexity values, especially for the volatile and benchmarking demands exceed the value 1, which is theoretically the upper limit for the Kolmogorov complexity. This behaviour is explained by the fact that the system expresses chaotic behaviour when the input demand is purely unpredictable. 7. Conclusions and future work This research work examined three different methods in a toolbox approach to support manufacturers on the design and planning of manufacturing networks for highly customised products. The obtained results from all three methods, namely the ISA, TS and SA, can support strategic level decisions related to the design of efficient manufacturing network configurations. The incorporation of criteria of cost, time, environmental impact and quality encapsulated some of the most significant objectives that manufacturing industries are striving to achieve nowadays. Moreover, the inclusion of dynamic complexity as a decision making criterion provided an insight to the operational characteristics of the networks, expressing the unpredictability of the flowtime performance indicator. The proposed methodology and toolbox can

comprise a valuable support aid for a decision maker in order to design and plan manufacturing networks that cope well with the market volatility imposed by customisation and even more by product personalisation demands. The results of all methods are of high quality whilst each method has its benefits and detriments. The Intelligent Search Algorithm revealed the best solution based on the utility value, whereas Simulated Annealing revealed low complexity values. Tabu Search on the other hand required the smallest computation time in order to solve the problem and yielded high quality solutions based on both the utility value and on complexity. Moreover, the difference in the values of complexity between the three methods is relatively small for the seasonal and rapid demand increase, and it becomes quite considerable for the stochastic demands, modelled through the uniform and normal distributions. Future work in this field will focus on investigating the structural characteristics of the manufacturing network. Thus, a static complexity index will be formulated that will encapsulate the relations between product, process and resource. Moreover, the correlation of the static and dynamic complexity will be investigated. The objective will be to examine whether the innate characteristics of a manufacturing network, such as the flexibility of different resources to process a variety of parts have an impact on its operational performance, which is expressed through the dynamic complexity. In case a correlation between static and dynamic complexity is identified, it can lead to the enhancement of the alternatives’ generation step of the decision-making procedure through a possible restriction of the search space. Acknowledgment The work reported in this paper has been partially supported by the EC FP7 funded project “e-CUSTOM-A web-based collaboration system for mass customisation” (Grant Agreement No: 260067). References [1] European Commission. Impact of the economic crisis on key sectors of the EU: the case of manufacturing and construction industries; 2010. [2] Garey M, Johnson D. Computers and intractability – a guide to the theory of NP-completeness. 1st ed. New York: W.H. Freeman & Co. Ltd.; 1990. [3] Liu SC, Chen AZ. Variable neighbourhood search for the inventory routing and scheduling problem in a supply chain. Expert Syst Appl 2012;39(4):4149–59. [4] Herrmann JW, Ioannou G, Minis I, Nagi R, Proth JM. Design of material flow networks in manufacturing facilities. J Manuf Syst 1995;14(4):277–89. [5] Krarup J, Pruzan PM. The simple plant location problem: survey and synthesis. Eur J Oper Res 1983;12(1):36–57. [6] Ehrgott M. Multicriteria optimisation. 2nd ed. Berlin Heidelberg: Springer; 2005. p. 253. [7] Shapiro J. Modelling the supply chain. Duxbury: Thomson Learning Inc.; 2001.

Please cite this article in press as: Mourtzis D, et al. A toolbox for the design, planning and operation of manufacturing networks in a mass customisation environment. J Manuf Syst (2014), http://dx.doi.org/10.1016/j.jmsy.2014.06.004

G Model JMSY-302; No. of Pages 13

ARTICLE IN PRESS D. Mourtzis et al. / Journal of Manufacturing Systems xxx (2014) xxx–xxx

[8] Ueda K, Takenaka T, Vancza J, Monostori L. Value creation and decision-making in sustainable society. CIRP Ann Manuf Technol 2009;58(2):681–700. [9] Simchi-Levi D. Operation rules. Cambridge, MA: MIT Press Ltd.; 2010. [10] Tang CS. Perspectives in supply chain risk management. Int J Prod Econ 2006;103(2):451–88. [11] Umeda Y, Takata S, Kimura F, Tomiyama T, Sutherland JW, Karaf S, et al. Toward integrated product and process life cycle planning – an environmental perspective. CIRP Ann Manuf Technol 2012;61(2):681–702. [12] Marsillac E, Roh JJ. Connecting product design, process and supply chain decisions to strengthen global supply chain capabilities,. Int J Prod Econ 2013 [In Press, available online 18.04.13]. [13] Nagurney A. Optimal supply chain network design and redesign at minimal total cost and with demand satisfaction. Int J Prod Econ 2010;128(1):200–8. [14] Cardoso SR, Barbosa-Póvoa AP, Relvas S. Design and planning of supply chains with integration of reverse logistics activities under demand uncertainty. Eur J Oper Res 2013;226(3):436–51. [15] Gumus AT, Guneri AF, Keles S. Supply chain network design using an integrated neuro-fuzzy and MILP approach: a comparative design study. Expert Syst Appl 2009;3610:12570–7. [16] Mourtzis D, Doukas M, Psarommatis F. Design and planning of decentralised production networks under high product variety demand. In: Procedia CIRP, 45th CIRP conference on manufacturing systems, vol. 3, no. 1. 2012. p. 330–8. [17] Melo MT, Nickel S, Saldanha-da-Gama F. A tabu search heuristic for redesigning a multi-echelon supply chain network over a planning horizon. Int J Prod Econ 2012;136(1):218–30. [18] Armentano VA, Shiguemoto AL, Loketangen A. Tabu search with path relinking for an integrated production distribution problem. Comput Oper Res 2011;38(8):1199–209. [19] Subramanian P, Ramkumar N, Narendran TT, Ganesh K. PRISM: priority based simulated annealing for a closed loop supply chain network design problem. Appl Soft Comput 2013;13(2):1121–35. [20] Peng P, Snyder LV, Lim A, Liu Z. Reliable logistics networks design with facility disruptions. Transp Res B Methodol 2011;45(8):1190–211. [21] Duffie NA, Roy D, Shi L. Dynamic modelling of production networks of autonomous work systems with local capacity control. CIRP Ann Manuf Technol 2008;57(1):463–6. [22] ElMaraghy W, ElMaraghy H, Tomiyama T, Monostori L. Complexity in engineering design and manufacturing. CIRP Ann Manuf Technol 2012;61(2):793–814. [23] Perona M, Miragliotta G. Complexity management and supply chain performance assessment. A field study and a conceptual framework. Int J Prod Econ 2004;90(1):103–15. [24] Suh NP. Complexity in Engineering. CIRP Ann Manuf Technol 2005;54(2):46–63. [25] Papakostas N, Efthymiou K, Mourtzis D, Chryssolouris G. Modelling the complexity of manufacturing systems using non-linear dynamics approaches. CIRP Ann Manuf Technol 2009;58(1):437–40. [26] Vrabik R, Butala P. Assessing operational complexity of manufacturing systems based on statistical complexity. Int J Prod Res 2012;50(14):3673–85. [27] Chryssolouris G, Efthymiou K, Papakostas N, Mourtzis D, Pagoropoulos A. Flexibility and Complexity: is it a trade-off? Int J Prod Res 2013 [Published online on 21.03.13]. [28] Efthymiou K, Pagoropoulos A, Papakostas N, Mourtzis D, Chryssolouris G. Manufacturing systems complexity review: challenges and outlook. In: Procedia CIRP, 45th CIRP Conf. on Manuf. Sys., vol. 3, no. 1. 2012. p. 644–9. [29] Smart J, Calinescu A, Huatuco H. Extending the information-theoretic measures of dynamic complexity of manufacturing systems. Int J Prod Res 2013;51(2):362–79. [30] Kumar V. Entropic measures of manufacturing complexity. Int J Prod Res 1987;25:957–66. [31] Sen AK. Complexity analysis of riverflow time series. Stoch Env Res Risk Asses 2009;23:361–6. [32] Suh NP. A theory of complexity periodicity and the design axioms. Res Eng Des 1999;11:116–31. [33] Mourtzis D, Doukas M, Psarommatis F. A multi-criteria evaluation of centralised and decentralised production networks in a highly customer-driven environment. CIRP Ann Manuf Technol 2012;61(1):427–30. [34] Michalos G, Makris S, Mourtzis D. An Intelligent Search Algorithm-based method to derive assembly line design alternatives. Int J Comp Integ Manuf 2012;25(3):211–29. [35] Chryssolouris G, Pierce J, Dicke K. A decision-making approach to the operation of flexible manufacturing systems. Int J Flex Manuf Syst 1992;4(3–4):309–30.

13

[36] Michalos G, Makris S, Mourtzis D. A web based tool for dynamic job rotation scheduling using multiple criteria. CIRP Ann Manuf Technol 2011;60(1): 453–6. [37] Glover F. Tabu search – part 1. ORSA J Comput 1989;1(2):190–206. [38] Kirkpatrick S, Gelatt CD, Vecchi MP. Optimisation by simulated annealing. Science 1983;220(4598):671–80. [39] Michalewicz Z, Fogel DB. How to solve it: modern heuristics. 2nd ed. Berlin Heidelberg: Springer Verlag; 2004. [40] Liao CJ, Shyu CH. An analytical determination of lead time with normal demand. Int J Oper Prod Manag 1991;11(9):72–8. [41] Mourtzis D, Doukas M, Psarommatis F. Environmental impact of centralised and decentralised production networks in the era of personalisation. In: Windt K, editor. Lecture notes in production engineering. Berlin Heidelberg: SpringerVerlag; 2013. p. 396–408, http://dx.doi.org/10.1007/978-3-642-30749-2. [42] EPA. www.epa.gov; 2010. [43] Milani AS, Shanian A, Madoliat R, Nemes JA. The effect of normalisation norms in multiple attribute decision making models: a case study in gear material selection. Struct Multidiscip Optim 2005;29(4):312–8. [44] Chryssolouris G. Manufacturing systems: theory and practice. 2nd ed. New York: Springer-Verlag; 2006. [45] Tseng GH, Huang JJ. Multiple attribute decision making: methods and applications. Boca Raton, FL: CRC Press/Taylor and Francis; 2011. [46] Yoon KP, Hwang CL. Multiple attribute decision making: an introduction. 1st ed. SAGE Publications, Inc.; 1995. ISBN: 9780803954861. [47] Phadke MS. Quality engineering using robust design. 1st ed. New Jersey, Englewood Cliffs: Prentice Hall; 1989. [48] Montgomery DC. Design and analysis of experiments. 7th ed. New York, NY: John Wiley & Sons; 2008. [49] Mourtzis D, Doukas M, Psarommatis F. Design and operation of manufacturing networks for mass customisation. CIRP Ann Manuf Technol 2013;63(1):467–70. [50] Helms MM, Ahmadi M, Jih WGK, Ettkin LP. Technologies in support of mass customization strategy: exploring the linkages between e-commerce and knowledge management. Comput Ind 2008;59(4):351–63. [51] Fogliatto FS, da Silveira GJC, Borenstein D. The mass customisation decade: an updated review of the literature. Int J Prod Econ 2012;138(1):14–25. [52] Koren Y. The global manufacturing revolution. 2nd ed. Hoboken, New Jersey: John Wiley & Sons, Inc.; 2009. [53] Jiao J, Tseng MM, Duffy VG, Lin F. Product family modelling for mass customisation. Comput Ind Eng 1998;35(3–4):495–8. [54] Evans JR. Applied production and operations management. 4th ed. St. Paul, MN, USA: West Publishing Company; 1993. [55] Kolmogorov AN. Three approaches to the quantitative definition of information. Probl Inf Transm 1965;1(1):1–7. [56] Kaspar F, Schuster HG. Easily calculable measure for the complexity of spatiotemporal patterns. Phys Rev A 1987;36(2):842–8. [57] Argoneto P, Bruccoleri M, Lo Nigro G, Perrone G, Noto La Diega S, Renna P, et al. High level planning of reconfigurable enterprises: a game theoretic approach. CIRP Ann Manuf Technol 2006;55(1):509–12. [58] Mourtzis D, Doukas M. The evolution of manufacturing systems: from craftsmanship to the era of customisation. In: Modrak V, Semanco P, editors. Design and management of lean production systems. IGI Global; 2014. ISBN13: 9781466650398. [59] Argoneto P, Renna P. Production planning, negotiation and coalition integration: a new tool for an innovative e-business model. Robot Comput Integr Manuf 2010;26:1–12. [60] Váncza J, Monostori L, Lutters D, Kumara SR, Tseng M, Valckenaers P, et al. Cooperative and responsive manufacturing enterprises. CIRP Ann Manuf Technol 2011;60(2):797–820. [61] Kim HS, Cho JH. Supply chain formation using agent negotiation. Decis Support Syst 2010;49(1):77–90. [62] Cachon GP, Fisher M. Supply chain inventory management and the value of shared information. Manage Sci 2000;46(8):1032–48. [63] Kanda AA, Deshmukh SG. Supply chain coordination: perspectives, empirical studies and research directions. Int J Prod Econ 2008;115(2):316–35. [64] Huan SH, Sheoran SK, Wang G. A review and analysis of supply chain operations reference (SCOR) model. Supply Chain Manag Int J 2004;9(1):23–9. [65] Altiparmak F, Gen M, Lin L, Paksoy T. A genetic algorithm approach for multi-objective optimization of supply chain networks. Comput Ind Eng 2006;51:196–215.

Please cite this article in press as: Mourtzis D, et al. A toolbox for the design, planning and operation of manufacturing networks in a mass customisation environment. J Manuf Syst (2014), http://dx.doi.org/10.1016/j.jmsy.2014.06.004