Construction of evolving models for the environmental evaluation of innovative sub-systems based on a hierarchical agglomerative clustering

Abstract

This paper deals with simplifying the environmental evaluation of an innovative sub-system related to the future complex system that will include it, by using evolving generic models built on a limited number of characteristics. For a complex system range, the evolving approach of the environmental modelling aims to generate a learning dynamics, to avoid the paralysing complexity induced in design by the valuation of many components according to many impact categories. Applied to the automotive sector, dendrograms are made with results of life cycle assessments (LCA) of 17 vehicles for 4 environmental indicators and on 3 life cycle steps. In an iterative process, a limit condition threshold on the resulting relative errors aims to cluster the vehicles. First, several calculation methods of dendrograms are tested. Second, the influence of the limit condition on the models is observed. Lastly, by simulating the vehicle population increase, the modelling capacity to evolve is tested. Five vehicle characteristics are sufficient to identify a model to be equivalent to the future vehicle. While the number of clusters is increased to simplify their identification with the system characteristics, the relative error variability increases too. The generic models are stable when adding LCA's results.

Keywords:

• life cycle assessment
• innovation
• complex system
• hierarchical agglomerative clustering
• evolving models
• automotive
• continuous improvement system

1. Background

1.1. Environmental evaluation of products

Companies are more and more encouraged by governments, non-governmental organizations and consumers associations, to evaluate and improve the environmental performances of their products with a life cycle and multi-criteria approach. The most widespread methodology for the environmental evaluation is the life cycle assessment (LCA) of products (Finnveden et al. 2009; Jacquemin, Pontalier, and Sablayrolles 2012). The international standards ISO 14040:2006 and 14044:2006 provide guidelines and a methodology framework of LCA. A lot of data and a very long time are needed to implement a LCA on a complex system; for instance, it may require more than 800 parameters during 3 months to make the LCA of an automobile. Often it is made by a dedicated team at the end of the product design and innovation process (DIP), when the product is entirely designed and manufactured. At this stage, it is too late for engineers to re-design their products. Therefore, in order to design more environmentally friendly products, companies must take into account the dimension ‘Environment’ earlier in the DIP.

The DIP is clearly complex (Tingström and Karlsson 2006): it is structured by complex organizations fed by complex information systems. Taking into account those structural and informational specificities is essential to integrate the environmental aspects (Johansson and Magnusson 2006; Le Pochat, Bertoluci, and Froelich 2007). Using a LCA tool during the DIP is a complicated task from a technical and organizational point of view (Millet et al. 2007). Although there is a research field to implement LCA based on the interoperability between design tools that facilitate data exchange (Rio, Reyes, and Roucoules 2013), such a strategy needs established environment skills within the company in order to analyse LCA results. Therefore, companies need eco-design tools that are developed to be integrated in the DIP, and to develop environmental skills within the company.

Integrating specific eco-design tools in the DIP of a company has been widely investigated. Companies must take into account simultaneously organizational and technical aspects (Johansson, Greif, and Fleischer 2007). Indeed, according to Millet et al. (2004), when a company would like to initiate an eco-design method, it has to deal with a ‘double paralysing complexity: how to introduce the fuzzy environmental dimension in the already hyper constrained universe of the product design process’. This paralysis leads companies to be reluctant to design a more environmentally friendly product. Millet et al. think that tools are the support of an organizational learning revealing new governing values. Many tools have been developed (Baumann, Boons, and Bragd 2002; Bovea and Pérez-Belis 2012) to help companies to evaluate and improve their products from an environmental point of view. A progressive approach for integrating these tools is required (Jacqueson, Millet, and Aoussat 2001): it is essential to gradually expand the Environment dimension from the regulation framework to a sustainable perspective. An eco-design method is generally undertaken in two phases: the diagnostic phase and the improvement phase. This article is focused on the environmental diagnostic phase and does not take into account economic or social concerns of sustainable development.

1.2. Automotive sector specificities

The experimental field in this paper is a car manufacturer whose goal is to initiate an eco-design method during the innovation phase. For a car manufacturer, innovations are most of the time not very intrusive, have an incremental nature and concern sub-systems or components (Oltra and Saint Jean 2009). A set of innovations is generally proposed by the innovation team to the routine design of the complex system, which Beaume, Maniak, and Midler (2009) represent as a flow of ‘innovative features’ from the innovation structure to the vehicle programme structure during the ‘contextualization phase’. Figure 1 adapted from Beaume et al. shows that both structures are completely separated. Such a company produces a range of automobiles that are complex products. With regard to the environmental evaluation of complex systems, the aggregation of LCA results of subsystem's components would give the LCA of the subsystem. In the same way the aggregation of LCA results of subsystems would give the LCA of the global product (Graedel et al. 2005). Thus, eco-designing a complex system may mainly be achieved at two levels: at the global system level, or at the sub-system or component level. The first case may lead to large scale actions as comparing several different choices of product architecture, or materials; it is nevertheless difficult to conduct this kind of actions because sufficient quantity of data is needed to make the evaluation, and the company strategy is engaged over the long term. On the opposite, the eco-design of a sub-system or a component is easier because a smaller quantity of data is needed; but the impact of the ensuing actions is lower.

Figure 1 Simplified development and innovation process of a car manufacturer adapted from Beaume et al. (2009).

In the automotive industry, car manufacturers are making LCA either on the complete vehicle (Finkbeiner and Hoffmann 2006; Schweimer and Levin 2000) or on subsystems (Keoleian et al. 1998; Schmidt and Beyer 1999); they are also developing eco-design methods, e.g. Ford's Product Sustainability Index (Schmidt 2006), Volvo's Environmental Priority Strategies (Karlsson, Steen, and Wendel 1997) and Environmental Effect Analysis (Tingström and Karlsson 2006) methods, or Toyota's Eco-Vehicle Assessment System (Nakano and Hirao 2011). With the study on the life cycle design of an air intake manifold, Keoleian and Kar (2003) conclude that the evaluation of the integration consequences of the innovative subsystem should be done at the future complex system level, but the environmental profiles of the complex systems are needed. As there are few data available during the innovation phase, the complex systems must be represented by models. Neither the majority of tools developed by car manufacturers, nor the majority of the available tools for eco-designing subsystems in the innovation phase (Bocken et al. 2011; Gasafi and Weil 2011; Poudelet et al. 2012) make it possible to use models for the simplification of innovative subsystems evaluation.

Figure 2 may be a representation of the research question. First, the innovation team (at the top left) makes innovative components (A₃), subsystems (A₂) or more rarely complete systems (A₁). Some innovative features are integrated into an existing system that is taken by the development team of vehicle programme (at the top right). At the end of the DIP, the system is generally evaluated by the environment team (at the bottom) that makes a LCA (B). The results of the LCA are saved in a database (C). The research question is: how to close the DIP cycle in order to evaluate the innovation features during the contextualization phase? The goal in this paper is to develop environmental vehicle models that will be used in a dedicated tool in order to make the contextual evaluation of an incremental innovative subsystem without evaluating the system entirely by developing vehicle models.

Figure 2 Schematic representation of the organizational context.

2. Issues

The goal of this paper is to present how the complex system that will include the innovation is modelled in a systematic way in order to avoid its complete environmental evaluation. Four key questions are related to this goal. The first one concerns the specificities of car manufacturer's product: they do not produce one kind of vehicles but a complete range. Each vehicle is different in terms of car classification (city, sedans, sports, etc.), motorization performances (diesel, gasoline, hybrid, etc.), bill of materials (BOM), etc. Q1: How to make an environmental modelling in order to represent the car manufacturer's entire range of vehicles?

The second question deals with the DIP cycle: the product design evolves after each design iteration in a continuous improvement system. Every project that begins may use capitalized data and knowledge from previous projects, and every project that ends may feed data knowledge bases. The modelling must take into account this evolution and have therefore the capacity to update. Q2: How to make an environmental modelling in order to support the continuous improvement of vehicles?

Lastly, in order to implement a learning strategy, a loop of organizational learning should be generated, fed by data feedbacks from the series of projects of the company, and oriented by an expansion strategy when taking the dimension ‘Environment’ into account, i.e. from a rough monocriterion eco-design approach to a detailed ‘Design for Environment’ approach. Q3: How to make an environmental modelling in order to generate a loop of learning around the dimension ‘Environment’?

In the environmental modelling literature, research has been focused on two options. The first one consists of building a product parameterized model from relevant physical parameters of the product, e.g. BOM, use consumption, etc. This method was used by Bohm et al. (2010), Ostad-Ahmad-Ghorabi and Collado-Ruiz (2011), Sousa and Wallace (2006), who identified, by using clustering methods, the design parameters that define the product and thus form the inputs of a simplified LCA. Collado-Ruiz and Ostad-Ahmad-Ghorabi (2010) developed a parameterization based on the detailed definition of functional units to categorize different products. Geyer (2008) implemented a parameterized model of LCA to simulate the impact of a material change in the automotive context on the greenhouse gases emissions; it has been implemented in the tool ‘Design Advisor’1 developed by WorldAutoSteel and dedicated to the automotive sector. The BOM of a vehicle is parameterized in such a way that a vehicle model can be defined with a set of parameters (mass, dimensions, shape, etc.). Then, the user may define old and new modules, and relate the global warming potential (GWP) impacts and the changing cost to the vehicle. The second option consists in building models directly with the environmental impacts resulting from a LCA. Some authors used clustering methods directly on the LCA results and thus defined environmental typologies (Andriankaja, Bertoluci, and Millet 2013) or classes (Curzons et al. 2007) of products, becoming the reference in a simplified comparative LCA. The models are environmental archetypes of products built on the environmental classes. The first option fits the requirements of the comparisons of several products that have different functionalities in order to categorize them. In the second option, the same functional unit is used in all LCA in order to compare the products. It is therefore adapted to the comparison of several ranges of the same product, e.g. a car or an airplane, and needs to have some LCA of previous products to create clusters.

None of the previous examples take into account both technical (key question Q1) and organizational (key questions Q2 and Q3) aspects at the same time for making a contextual evaluation of innovative subsystems, i.e. for relating the environmental impacts of an innovative feature to the complete system. There is an obvious lack of research in taking into account technical and organizational issues in the environmental modelling, as recommended by Johansson, Greif, and Fleischer (2007).

We have developed an approach to build evolving models for the environmental evaluation of complex systems that we named EMEECS. The EMEECS must be built to be used in a dedicated tool for evaluating the environmental impacts of innovative features on the vehicle. This tool is named tool for evaluating the environmental performances of innovations (TEEPI). The user makes the LCA of the initial and innovative features. He/she chooses the right EMEECS for his/her application by selecting the relevant vehicle characteristics that correspond to the vehicle programme into which the innovative solution will be integrated. He chooses the right assembly plant for which an environmental profile is available. Lastly, the initial feature values are subtracted to the EMEECS values, then the assembly plant values and the innovative feature values are added to the EMEECS values. The original EMEECS values and the transformed EMEECS values are compared to highlight the evolution of the environmental profile of the vehicle. EMEECS are tested at a French car manufacturer and they are supposed to satisfy the following (Figure 3):

• On the technical point, the contextual environmental evaluation of a sub-system, i.e. to relate the environmental impacts of a sub-system to the complex system, in the dedicated tool (named TEEPI in Figure 3).

  Figure 3 Schematic representation of the issue.

  

• On the organizational learning point, the expansion of taking into account the Environment dimension and the vehicle design evolution, represented by a loop generating building iterations of EMEECS that is defined by a succession of learning phases noted i and shown on Figure 3 by a grey dotted line loop.

3. Research methodology

The EMEECS model takes as its starting point the key idea that there are archetypes of environmental impacts calculated on several life cycle steps of complex system classes, i.e. the user of the proposed method will directly deal with the environmental impacts of the initial and innovative solutions and of the global system. The complex system classes are thus created by applying the hierarchical agglomerative classification (HAC) method (Hastie, Tibshirani, and Friedman 2001) on previous LCA's results. The classes from the HAC are therefore the EMEECS: the environmental indicator value of the EMEECS is equal to the average of the class, and the uncertainty of the EMEECS is equal to the standard deviation of the class.

Therefore, the EMEECS should first be identified with a restricted number of vehicles' characteristics; this number would directly be linked with the number of classes: the more the classes, the more the vehicles' characteristics are needed. A moderate level of number of classes created from a starting sample would be equal to at least one half of the number of the sample's individuals. Second, creating vehicle models would generate a modelling error that is intrinsic to the EMEECS. The following can be assumed: the more the classes, the more the errors. A compromise should be found between the number of classes and the errors. As it is acknowledged by Bala et al. (2010), 10% should be considered as a moderate level of discrepancy. Third, the EMEECS should be stable regarding the vehicle design evolution as the learning phases follow one another. Fourth, the EMEECS should be built in order to permit the expansion of the dimension ‘Environment’, i.e. adding life cycle stages and environmental indicators.

3.1. Conventions

In this paper, it is considered that, for the first learning phase i = 1, 17 LCA of cars (Table 1) on the full life cycle and for 10 environmental indicators such as global warming, eutrophication, acidification, metal depletion and so on are available. The functional unit that is common for all those LCA is ‘Transporting passengers and goods over 150,000 km during 10 years according to the new European driving cycle NEDC’. This functional unit is defined on an average mileage and a lifetime used by the majority of car manufacturers. Moreover, the emission measurements of pollutants during the use phase are only available on the NEDC, allowing comparing vehicles. Nevertheless, NEDC is not representative of the reality; this study will need to be supplemented by the new driving cycles worldwide harmonized light vehicles test procedures (WLTP) expected in 2014.

Table 1 List of available LCA results of cars.

Vehicle	Automotive segment	Fuel type	Gear box	Finish	Car's shape	Consumption (L/100 km)
V1	B	Diesel	Manual	Mid	Sedan	3.8
V2	B	Diesel	Manual	Mid	Sedan	3.8
V3	B	Gasoline	Electronically controlled	Mid	Sedan	4.5
V4	B	Gasoline	Manual	High	Sedan	4.3
V5	C	Gasoline	Manual	Mid	Estate	6.7
V6	C	Gasoline	Manual	Mid	Sedan	6.4
V7	C	Gasoline	Manual	High	Sedan	6.1
V8	C+	Diesel	Manual	High	Monospace (seven seats)	5.3
V9	Crossover	Gasoline	Manual	High	Sedan	6.7
V10	Crossover	Hybrid	Electronically controlled	High	Sedan	3.9
V11	D	Diesel	Electronically controlled	High	Sedan	4.4
V12	D	Diesel	Automatic	High	Sedan	5.7
V13	D	Hybrid	Electronically controlled	High	Sedan	3.6
V14	B	Gasoline	Manual	High	Convertible sedan	6
V15	C	Gasoline	Manual	Mid	Coupé Cabriolet	6.7
V16	C+	Diesel	Manual	Mid	Monospace (five seats)	5.2
V17	Crossover	Hybrid	Electronically controlled	High	Sedan	3.8

The production, use and maintenance, and end-of-life are considered. The painting and the assembly done during the production phase are directly related to the assembly plant; they are not considered in this paper because they are taken into account with creating typologies of assembly plants in future work: the environmental typologies of assembly plants would be added to the EMEECS's environmental indicator values.

3.2. Technical issues

By applying a HAC on the collected LCA (see Section 3.2.1), cars are grouped into classes; an iterative process based on a boundary condition (see Section 3.2.2) leads to identify optimal classes.

3.2.1. Hierarchical agglomerative classification

Given a set V_i of vehicles v_j with cardinality , a set X_i of environmental indicators x_k with cardinality and a set Y_i of life cycle steps y_l with cardinality . The LCA of the vehicle v_j gives the value of an environmental indicator x_k at a life cycle step y_l; this value is noted as w_jkl. The averages (respectively, ) and standard deviations (respectively, ) are calculated over the set V_i of vehicles v_j for the environmental indicator x_k and the life cycle step y_l (respectively, the full life cycle).

Because of the HAC treatment, vehicles that have similar LCA results are grouped into classes, by making dendrograms (e.g. see Figure 4 for the dendrogram on the manufacturing step and for the GWP). The elements to be classified are vehicles of which the original variables are environmental indicator values per life cycle step and expressed in different units; so, the standard score associated with the original variables is calculated for each life cycle step (Equation (1)) and for the full life cycle (Equation (2)) (see the Appendix Table 1).

(1)

Figure 4 Dendrogram obtained for the manufacturing step and on the GWP indicator.

(2)

The cluster dissimilarity metric δ – or distance between elements – is the Euclidean distance, and the linkage criteria Δ – or distance between classes – is the average linkage clustering. The vehicles or clusters of vehicles that have the minimum linkage criteria (Y-axis on Figure 4) are grouped together, and a tree turned upside down named ‘dendrogram’ and noted d is built (up arrow (1) – Figure 4): the ‘trunk’ is at the top and the ‘branches’ are at the bottom. Each branch is a cluster of vehicles.

In view of the nature of the original variables, there are four ways to calculate the cluster dissimilarity, and as many groups of dendrograms, with D the set of dendrograms d_kl calculated for the environmental indicator x_k and the life cycle step y_l:

• Group (D1): indicator and step accumulations, one dendrogram, (see the following Equation (3))

  (3)

• Group (D2): step accumulation and per indicator, N(X_i) dendrograms, (see the following Equation (4))

  (4)

• Group (D3): indicator accumulation and per step, N(Y_i) dendrograms, (see the following Equation (5))

  (5)

• Group (D4): per step and per indicator, N(X_i) * N(Y_i) dendrograms, (see the following Equation (6))

(6)

Given c_klm the mth class of the dendrogram d_kl. The linkage criteria between two classes c_klm and that, respectively, hold and vehicles, are calculated as follows in Equation (7):

(7)

3.2.2. Iterative process to determine optimal classes of cars

Once finalized, the dendrogram is used from top to bottom. Each branch and the trunk itself are groups of vehicles. Let be the set of classes c_klm in the dendrogram d_kl. For each class c_klm of the dendrogram d_kl, the average and the standard deviation are calculated over the values w_jkl of the vehicles included in the class c_klm. The average represents the value of the class and the standard deviation represents the uncertainty of the class. By using the dendrogram from top to down, the number of classes evolves from 1 (the trunk) to the number of vehicles in the dendrogram . Therefore, there is a number of classes that would be optimum according to the allowed maximum error between the value of the class and the values w_jkl of the vehicles in the class. The relative error ϵ_jklm between the value of the class and the values w_jkl of the vehicles in the class is calculated (8):

(8)

The optimal classes, with which the EMEECS are built, are determined by the following boundary condition (C), and for each dendrogram:

(C) The relative errors committed by the average of the class c_klm related to the values w_jkl of the environmental indicators of vehicles v_j included in this class must be inferior to a maximum limit ϵ_max in terms of absolute value, i.e. (Equation (9)):

(9)

The iterative process starts from the trunk, where the vehicles are grouped into one class (N (C_kl)  = 1) and go down towards the branches (down arrow (2) – Figure 4). If the condition (C) is satisfied on a branch, then the associated class is considered as optimal; else the branch is divided into two sub-branches (N(C_kl) = N(C_kl)+1) and the process is iterated on both sub-branches. Once the iterative process is over, the error relative mean and standard deviation are calculated on all the dendrograms of the group at which this iterative process is applied.

An extreme case would appear if the environmental profile of vehicles would be very different and the vehicles would not be grouped in classes; in this case, there would be as many classes as vehicles and dendrograms. To measure the reduction potential, the rate of class number reduction τ_r is defined in the following Equation (10):

(10)

We think that there is a compromise between a low level of errors and the high rate of class reduction in order to create classes that are identifiable with few vehicle's characteristics.

3.3. Organizational issues

LCA are made by the Environment team for all the life cycle steps (including upstream and downstream logistics) and with 10 environmental and flow indicators (including abiotic depletion, acidification, eutrophication, global warming and photochemical ozone creation). But, as it is stated in the issue, the integration of a new dimension within the product development and innovation process must be progressive. Integrating, at the same time, the evaluation of innovations on 10 environmental and flow indicators, and over all the life cycle steps, could create an opposition from the innovation leaders who are inexperienced within the dimension ‘Environment’. A learning strategy must therefore be implemented through the construction of a loop of learning while updating the EMEECS (Table 2). At each learning phase (i: = i+1), the Environment team may include another environmental indicator , another life cycle step , in order to expand the Environment dimension. We propose to initialize the process with four environmental indicators that car manufacturers are communicating (Table 3) or many papers in the automotive sector are using (Table 4). The four environmental indicators that will be considered are calculated by the CML 2001 method:

• Global warming potential, GWP, (kg CO₂-equivalent);

  Table 2 Example of an expansion strategy in taking the Environment dimension into account.

  Parameters of the learning strategy	Learning phase i
  	1	2	3	…
  Number of cars N(Vi)	17	20	24	…
  Number of environmental indicators N(Xi)	4	4	5	…
  Number of life cycle steps N(Yi)	3	3	4	…
  Driving cycles	NEDC	NEDC	WLTP	…

  Table 3 Environmental indicators communicated by car manufacturers.

  Manufacturer	Sources	Car	Indicators
  Nissan	Sustainability reports	LEAF	CO2
  Volkswagen	Environmental commendations	Golf	GWP, POCP, AP, ODP, EP
  		Passat	GWP, POCP
  		Polo	GWP, POCP
  Mercedes	Environmental certificates	A-Class	GWP, AP, EP, ADP, POCP
  		B-Class	GWP, AP, EP, ADP, POCP
  	Finkbeiner and Hoffmann (2006)	S-Class	GWP, AP, EP, ADP, POCP
  Hyundai Motor	Sustainability reports	Tucson ix, i10 et i20	CO2
  Honda	Sustainability reports	Civic, CR-V, freed hybrid, N BOX, fit shuttle hybrid	CO2
  Ford	Schmidt (2006)	S-Max et Galaxy	GWP, POCP

  Table 4 Environmental indicators and flows used in six publications from the automotive sector.

  Indicators	Schmidt et al. (2004)	Muñoz et al. (2005)	Alonso et al. (2006)	Ribeiro, Ferreira, and Partidário (2006)	Puri, Compston, and Pantano (2009)	Dlamini et al. (2011)
  ADP	X	X	X	X	X
  AP	X	X	X	X	X	X
  EC		X			X	X
  EP			X	X
  GWP	X	X	X	X	X	X
  HTP		X	X	X
  ODP	X			X	X
  POCP	X	X	X	X	X	X
  Toxicité		X		X		X
  Waste	X	X			X	X
  Water pollution					X
  WC		X

  ADP, abiotic depletion potential (kg Sb-equiv.); AP, acidification potential (kg SO₂-equiv.); EC, energy consumption (MJ); EP, eutrophication potential (kg phosphate-equiv.); GWP, global warming potential (kg CO₂-equiv.); HTP, human toxicity potential (kg DCB-equiv.); ODP, ozone depletion potential (kg R11-equiv.); POCP, photochemical ozone creation potential (kg ethene-equiv.); WC, water consumption (L).

• Eutrophication potential, EP, (kg phosphate-equivalent);

• Photochemical ozone creation potential, POCP, (kg ethene-equivalent);

• Abiotic depletion potential, ADP, (kg Sb-equivalent).

When adding environmental indicators, the EMEECS building process must be stable. We think that the group (D4) is the most stable regarding the number of classes and the uncertainty because the independency of all life cycle stages and environmental indicators is assured.

At each learning phase, the environment team may have made the LCA of a new vehicle . The integration of the LCA results into the EMEECS building process should also be stable regarding the number of classes and the uncertainty. We think that the stability is assured by the HAC treatment.

4. Results

In the previous part, the hypotheses and inputs used in this paper were presented. In this part, the experimental design is first described. Second, the construction of the optimal classes is tested; the influences of the environmental indicators and life cycle steps are also shown. Third, the influence of the boundary condition (C) threshold is tested. Fourth, the influence of the different starting sample's populations is tested.

4.1. Experimental design

The EMEECS are tested with three experiments. For each one, the following elements are shown:

• The mean of errors and the error variability, represented by the standard deviation ;

• The rate of class number reduction τ_r;

• The groups of vehicles are made with the characteristics of the vehicles that are described in Table 1.

The first experiment tests the hypothesis that the HAC treatment of previous LCA's results may create optimal classes that are identified with a restricted number of vehicle characteristics to build the EMEECS. As there are different solutions to build the dendrograms, the results of the four dendrogram groups (D1)–(D4) are established over all the 17 vehicles and compared. The graphs of influence of the environmental indicators are shown, and the table, in which the classes are synthetized, is made.

In the second test, the influence of the threshold ϵ_max on the mean and the standard deviation of the errors, and rate of class number reduction is experimented over the 17 vehicles, with 33 values of ϵ_max between 0.1% and 40%. It deals with having the best compromise between the highest rate of class number reduction and the smallest variability of errors represented by the standard deviation .

In the third test, the influence of the loop of learning is evaluated by comparing the results obtained over 10 populations of 13 vehicles, with the results obtained over all the 17 vehicles. The goal is to simulate the addition of LCA results and the update of the EMEECS. The groups must be similar and the number of optimal classes equal, leading to a rise of the rate of class reduction.

4.2. Comparing the groups of dendrograms

First, the four groups of dendrograms are compared; the results are summarized in Table 5. The rate of class number reduction is clearly different from one group to another. It is below 50% for groups (D1) and (D3), respectively, 29% and 45%. In both groups, the environmental indicators are accumulated in the calculation of the cluster dissimilarity. The rate of class number reduction of groups (D2) and (D4) is higher than 50%, respectively, 72% and 75%. In the first group, the environmental indicator values are calculated over the full life cycle, while the environmental indicators are discretized for each life cycle step. For each group of dendrograms, the mean of the relative errors is clearly very low, and below 0.20%. The standard deviation is clearly lower for group (D1) than for groups (D2) and (D4); the standard deviation of group (D3) has an intermediate value. It can be observed that subdividing delivers higher reductions but also higher errors; this effect is stronger when subdividing life cycle steps than impact categories. As it has been stated in the research methodology, the rate of class reduction should be at least 50%; the groups (D1) and (D3) do not fit the requirements and a choice has to be made between (D2) and (D4) where the environmental indicators are not aggregated.

Table 5 Results for the four groups of dendrograms.

	(D1)	(D2)	(D3)	(D4)
N(Vi) × N(Xi) × N(Yi)	17	68	51	204
	12	19	28	52
τr	29%	72%	45%	75%
μ(ϵjklm)	0.06%	0.18%	0.11%	0.19%
σ(ϵjklm)	2.56%	4.28%	3.35%	4.30%

Second, the influences of the environmental impacts' values on the groups of dendrograms (D2) and (D4) are shown in Figure 5. Group (D2) where the life cycle steps are aggregated is represented with diamond-shaped areas; group (D4), where neither the life cycle steps, nor the environmental indicators are aggregated, is represented with squares. The environmental indicators of both groups form polygons. It can be observed that the area of the (D2)-polygon is higher than the area of the (D4)-polygon. It means that group (D2) is more sensitive to the environmental indicators than the dendrograms of group (D4); the variability of the errors and of the number of class is higher in group (D2) than in group (D4). Moreover, adding an environmental indicator and/or a life cycle step during the succession of learning phases will not have any influence on the already implemented ones for group (D4). Therefore, the method used to create the dendrograms of group (D4) is more advantageous than the one of group (D2). The dendrogram group (D4) is used in the subsequent parts of the paper.

Figure 5 Representation of the rate of class number reduction as a function of the standard deviation of the relative errors of the environmental indicators taken separately of dendrogram groups (D2) and (D4).

The results shown in Table 6 are focused on the manufacturing step and concern the clusters of cars obtained with the group of dendrogram (D4); the complete results are available in the Appendix Table 2. Here are comments for the four environmental impacts:

• GWP: Hybrids are grouped together because of the high density of precious metals in the battery. The other groups are made by car classification: {B}, {C} and {C+; D; crossover};

  Table 6 Description of the vehicles' clusters for the manufacturing step.

  Cluster number	GWP	EP	POCP	ADP
  1	V10V13V17	V8	V10V13V17	V10V13V17
  2	V1V2V3V4V14	V10	V8V9V11V12V15V16	V8V9
  3	V5V6V7	V9V13V17	V1V2V3V4V14	V11V12
  4	V8V9V11V12V15V16	V11V12	V5V6V7	V7V15V16
  5	–	V4V14V16	–	V1V2
  6	–	V1V2	–	V3V4V5V6V14
  7	–	V3V5V6V7V15	–	–

• EP: The presence of leather in segment C or upper vehicles is the prime separation criterion; it can be assimilated to a high quality finish. The hybrid crossover {V10}, the heaviest car in the collection, is separated from the other hybrids that are grouped with the crossover that has the most leather {V9; V13; V17}. The seven-seat monospace {V8} is also separated. Mid quality finish segment B vehicles are grouped {V1; V2}, such as high quality finish segment D vehicles {V11; V12}. The high quality finish segment B vehicles are grouped with the mid finish quality monospace {V4; V14; V16}. All the segment C vehicles are grouped with the high quality finish segment B that has less leather {V3; V5; V6; V7; V15};

• POCP: Same groups as GWP – manufacturing;

• ADP: Hybrids are grouped together. Then groups are made per segment and per finish level. The high quality finish segment B vehicles are grouped with the mid quality finish segment C vehicles {V3; V4; V5; V6; V14}. The mid quality finish segment B vehicles are grouped together {V1; V2}, such as the high quality finish segment D {V11; V12}. The high quality finish segment C vehicle is grouped with the coupé-cabriolet and with the five-seat mid quality finish monospace {V7; V15; V16}. Lastly, the seven-seat monospace is grouped with the ICE crossover {V8; V9}.

The environmental indicators EP and POCP during the use step are depending on the emissions of the cars that are strongly and randomly varying from one car to another. The vehicle characteristics are insufficient to describe the groups of cars obtained for POCP during the use step. The classes obtained with (D4) may be described with five vehicle characteristics: segment, fuel type, finish, car's shape and consumption.

4.3. Influence of the thresholdϵ_max on the boundary condition (C)

In this part, the evolution of the mean and the variability of the relative errors, and of the rate of class number reduction, as functions of the threshold ϵ_max, is tested. Those three functions are, respectively, represented in Figure 6(a)–(c), and their nature is, respectively, polynomial, linear and logarithmic. We can therefore assume that changing the maximum value of errors is equivalent to changing the standard deviation.

Figure 6 (a) Evolution of the mean of the relative errors as a function of the boundary condition (C) threshold of which value is varying between 0.1% and 40%. (b) Evolution of the standard deviation of the relative errors as a function of the boundary condition (C) threshold of which value is varying between 0.1% and 40%. (c) Evolution of the rate of class number reduction as a function of the boundary condition (C) threshold of which value is varying between 0.1% and 40%.

In Figure 7(a), the evolution of the rate of class number reduction is shown as a function of the standard deviation of the relative errors; its nature is logarithmic. The fewer the classes, the less dispersed and the closer to 0 the relative errors are. Figure 7(b) is a zoom of Figure 7(a) between ϵ_max = 7% and ϵ_max = 13%, with a constant step of 1%. The distance on the Y-axis between the points ϵ_max = 9% and ϵ_max = 10% is significant; it means that decreasing the threshold value to decrease the error variability leads to a higher number of classes. In the same way, the distance on the X-axis between the points ϵ_max = 10% and ϵ_max = 11% is significant; it means that increasing the threshold value to increase the rate of class number reduction leads to a higher variability of errors.

Figure 7 (a) Evolution of the rate of class number reduction as a function of the standard deviation of the relative errors for different values of the boundary condition (C) threshold varying between 0.1% and 40%. (b) Evolution of the rate of class number reduction as a function of the standard deviation of the relative errors for different values of the boundary condition (C) threshold varying between 7% and 13%.

The following can be noticed:

• A relevant benefit in terms of class number reduction is not made by allowing maximum relative errors higher than 10%.

• The number of optimal classes increases from 52 to 55 while the threshold is only reduced by 1%, and the number of characteristics to identify the classes would increase.

ϵ_max = 10% seems, therefore, to be a good compromise between a high rate of class number reduction and a low dispersion of relative errors.

4.4. Comparing results between different populations of the starting sample

In this third test, 10 simulations are made with different populations of 13 vehicles among the original 17. The 13 vehicles in the 10 simulations are the same as the original 17, but 4 vehicles are omitted for each simulation (1)–(10). These four omitted vehicles are given in Table 7.

Table 7 List of the omitted vehicles for the 10 simulations.

(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)
V14	V1	V6	V8	V5	V9	V1	V2	V3	V1
V15	V3	V9	V10	V6	V10	V9	V4	V6	V2
V16	V7	V12	V13	V7	V12	V14	V5	V10	V3
V17	V8	V13	V17	V15	V17	V16	V11	V11	V4

Results of the 10 simulations are represented on the graph of Figure 8. For nine simulations, the rate of class reduction is clearly lower and different from the original sample of 17 vehicles noted (0). It means that the number of classes is changing little and is between 46 and 50. In contrast, the standard deviations of the relative errors are dispersed between 4.02% and 4.79%. A rate of class reduction of 72%, the highest rate reached by the 10 simulations, is achieved by the simulation (4) for which all hybrids and the seven-seat monospace are excluded, bringing the number of optimal classes to 44.

Figure 8 Representation of the rate of class reduction as a function of the standard deviation of the relative errors of 10 simulations with 13 vehicles compared with the results of the group (D4) with 17 vehicles.

The groups of vehicles obtained with simulation (4) are similar to those obtained with (D4), by omitting vehicles V8, V10, V13 and V17. Nevertheless, for the manufacturing step, vehicle V9 is grouped with V11 and V12 on the indicator EP while they were separated with 17 vehicles (see Table 8).

Table 8 Description of the vehicles' clusters for the manufacturing step of the simulation (4).

Cluster number	GWP	EP	POCP	ADP
1	V1V2V3V4V14	V9V11V12	V9V11V12V15V16	V11V12
2	V5V6V7	V4V14V16	V1V2V3V4V14	V9
3	V9V11V12V15V16	V1V2	V5V6V7	V7V15V16
4	–	V3V5V6V7V15	–	V1V2
5	–	–	–	V3V4V5V6V14

The groups of vehicles are stable and describable according to the vehicle characteristics of Table 1.

5. Discussion

The first test concerns the key question Q1. Classes obtained by the HAC and iterative process may be described with five vehicle characteristics: segment, fuel type, finish, car's shape and consumption. The car manufacturer's entire range of vehicles may be modelled. The boundary condition (C) ϵ_max = 10% is a good compromise between a satisfying number of classes and a variability of the low relative errors, as shown in the third test.

The classes deduced on the indicator POCP for the use phase are nevertheless less easy to identify because all vehicles used in this paper answer the same depollution standard Euro V. It is due to the difficulty to find a clear relationship between the pollutant emissions and a vehicle characteristics, as shown in Figure 9, where POCP values are calculated on the tank to wheels (TtW) step for 66 diesel vehicles and 96 gasoline vehicles (segments B, C, D and crossovers). The emissions were collected from the ADEME car comparison website for different versions of vehicles used in this paper. Two relatively separated clouds of points appear, but none of them have significant differences in terms of POCP values.

Figure 9 Value of the indicator POCP on the TtW step of 162 vehicles as a function of their fuel consumption and of the type of fuel (gasoline or diesel).

For each phase of the loop of learning, environmental indicators, life cycle steps and/or vehicles are added. On the one hand (key question Q3), it has been shown that for the first two cases, their influence on the rate of class reduction and on the variability of the relative errors is indeed strong. However, the optimal classes obtained for the existing life cycle steps and environmental indicators will not be modified. On the other hand (key question Q2), adding vehicles has little influence on the identification of optimal classes. The variation of the number of optimal classes will be small, leading to an increase of the rate of class number reduction. Nevertheless, the continuous improvement system is based on making LCA systematically whose results are integrated in the EMEECS building process for updating the environmental indicators scores of the models. If the company's strategy concerning LCA changes, the continuous improvement process is not assured.

The tool ‘Design Advisor’ developed by WorldAutoSteel has a strategy similar to TEEPI's strategy. The BOM of a vehicle is parameterized in such a way that a vehicle can be defined with a set of parameters (mass, dimensions, shape, etc.). Then, the user may define old and new modules, and relate the GWP impacts and the changing cost of the vehicle. Therefore, the philosophy of WorldAutoSteel is identical to that of EMEECS in TEEPI tool; but the strategies are diverging on the method of creating vehicle models: ‘Design Advisor’ creates a model from physical assumption, while EMEECS are directly built on previous LCA results. In the first tool, the range may nevertheless be larger than that represented by EMEECS and the user may define more precisely the vehicle. Lastly, ‘Design Advisor’ is dedicated to the GWP impacts while TEEPI deals with four environmental impacts, if not more.

6. Conclusions

The method of constructing EMEECS presented in this paper has two benefits: first, the contextual evaluation of an innovative sub-system is simplified by avoiding the evaluation of the entire future complex system, and second, in a long-term perspective, a loop of organizational learning is generated between the project, innovation and environment teams on the thematic eco-design, leading to a continuous improvement of the DIP. Engineers may make environmental evaluation of their innovation earlier in the DIP without having a complete description of the vehicle programme. An eco-design baseline integrated into the innovation quality process has been developed by the environment team in order to guarantee the use of TEEPI. All the innovations that may a priori have an environmental risk according to qualitative criteria defined in the eco-design baseline should be evaluated with TEEPI. Another publication concerning the use of TEEPI and its application to case studies is already planned.

We identify areas of further research that we intend to conduct. First, increasing the granularity level for representing models: these could be constructed over the sub-systems (body-in-white, engine, etc.) instead of the entire vehicle. A better accuracy could be reached without excessively increasing the number of characteristics needed to select the optimal classes. To do that, the vehicle functional scopes would be treated with HAC and the iterative process. Furthermore, it will be interesting to observe the influence of introducing vehicles of different generations

Second, only the environmental diagnostic phase of an eco-design method has been treated in this paper, and more particularly the creation of an environmental model. It would be necessary to propose improvement of tools to engineers: linking the environmental vehicle models to an improvement method would be really helpful for engineers in order to improve their innovations. For instance, TEEPI and the EMEECS would be linked with the method developed by Tchertchian, Yvars, and Millet (2013) on using the constrained satisfaction problem in resolving an eco-design problem.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

^1.http://www.worldautosteel.org/projects/design-advisor/, accessed on 30 May 2013.

Appendices

Appendix 1 Standard scores per life cycle step and for the full life cycle.

Life cycle steps	Manufacturing				Use				End of life				Life cycle
Environmental indicators	GWP	EP	POCP	ADP	GWP	EP	POCP	ADP	GWP	EP	POCP	ADP	GWP	EP	POCP	ADP
V1	− 1.222	− 1.093	− 0.981	− 0.640	− 1.131	0.865	0.487	− 0.998	− 0.948	− 0.969	− 0.991	− 0.936	− 1.524	0.129	− 0.429	− 0.649
V2	− 1.126	− 1.097	− 0.973	− 0.596	− 1.163	0.671	0.863	− 0.304	− 1.554	− 1.344	− 1.388	− 1.587	− 1.540	− 0.014	0.161	− 0.597
V3	− 1.171	− 0.786	− 0.935	− 0.551	− 0.591	− 1.217	1.657	2.027	− 1.198	− 0.721	− 0.843	− 1.448	− 0.977	− 1.216	1.445	− 0.527
V4	− 1.164	− 0.350	− 0.902	− 0.535	− 0.778	− 1.240	0.216	− 0.375	− 0.979	− 0.775	− 0.855	− 1.116	− 1.156	− 1.042	− 0.753	− 0.537
V5	− 0.194	− 0.778	− 0.427	− 0.527	1.351	− 0.935	0.028	− 0.918	− 0.469	− 0.616	− 0.568	− 0.325	1.271	− 1.011	− 0.473	− 0.536
V6	− 0.380	− 0.837	− 0.559	− 0.540	1.053	− 0.877	0.590	− 0.929	− 0.648	− 0.808	− 0.771	− 0.515	0.913	− 0.997	0.239	− 0.548
V7	− 0.498	− 0.809	− 0.510	− 0.466	0.792	− 0.913	− 0.207	− 0.932	− 0.309	− 0.142	− 0.206	− 0.445	0.625	− 1.004	− 0.935	− 0.475
V8	0.158	2.221	0.276	− 0.299	0.194	1.248	− 0.426	0.617	1.517	1.888	1.798	1.198	0.277	1.875	− 0.315	− 0.291
V9	0.225	0.791	0.047	− 0.335	1.366	− 1.025	1.449	− 0.953	0.811	1.037	0.989	0.638	1.446	− 0.374	2.316	− 0.344
V10	1.856	1.472	1.994	2.064	− 1.010	0.518	− 1.459	0.661	1.578	1.435	1.477	1.595	− 0.391	1.025	0.140	2.065
V11	0.666	0.522	0.308	− 0.119	− 0.593	0.870	− 0.352	1.697	0.849	0.498	0.562	0.982	− 0.363	0.852	− 0.171	− 0.099
V12	0.189	0.398	0.057	− 0.091	0.613	1.091	− 1.051	0.867	0.368	0.089	0.172	0.556	0.676	0.951	− 1.562	− 0.081
V13	1.605	0.902	1.894	2.220	− 1.255	0.806	− 1.979	0.583	0.424	0.114	0.228	0.676	− 0.739	0.969	− 0.800	2.219
V14	− 1.044	− 0.451	− 0.892	− 0.567	0.725	− 0.825	− 0.049	− 0.953	− 1.073	− 1.215	− 1.220	− 1.005	0.371	− 0.795	− 1.157	− 0.576
V15	0.387	− 0.636	− 0.086	− 0.471	1.351	− 1.031	0.645	− 0.918	− 0.268	− 0.603	− 0.476	0.051	1.456	− 1.017	0.893	− 0.480
V16	0.222	− 0.505	− 0.071	− 0.468	0.189	1.220	0.736	0.799	0.803	1.115	1.042	0.567	0.275	0.658	1.067	− 0.457
V17	1.490	1.034	1.761	1.920	− 1.113	0.774	− 1.150	0.028	1.098	1.017	1.049	1.113	− 0.618	1.011	0.335	1.914

Appendix 2 Description of the clusters of vehicles.

Life cycle steps	Environmental indicators	1	2	3	4	5	6	7	Comments
Manufacturing	GWP	V10V13V17	V1V2V3V4V14	V5V6V7	V8V9V11V12V15V16	–	–	–	Hybrids are grouped together because of the high density of precious metals in the battery. The other groups are made by car classification: {B}, {C}, and {C+; D; crossover}.
	EP	V8	V10	V9V13V17	V11V12	V4V14V16	V1V2	V3V5V6V7V15	The presence of leather in segment C or upper vehicles is the prime separation criterion; it can be assimilated to a high quality finish. The hybrid crossover {V10}, the heaviest car in the collection, is separated from the other hybrids that are grouped with the crossover that has the most leather {V9; V13; V17}. The seven-seat monospace {V8} is also separated. Mid quality finish segment B vehicles are grouped {V1; V2}, such as high quality finish segment D vehicles {V11; V12}. The high quality finish segment B vehicles are grouped with the mid finish quality monospace {V4; V14; V16}. All the segment C vehicles are grouped with the high quality finish segment B that has less leather {V3; V5; V6; V7; V15}.
POCP	V10V13V17	V8V9V11V12V15V16	V1V2V3V4V14	V5V6V7	–	–	–	Same groups as GWP – Manufacturing.
ADP	V10V13V17	V8V9	V11V12	V7V15V16	V1V2	V3V4V5V6V14	–	Hybrids are grouped together. Then groups are made per segment and per finish level. The high quality finish segment B vehicles are grouped with the mid quality finish segment C vehicles {V3; V4; V5; V6; V14}. The mid quality finish segment B vehicles are grouped together {V1; V2}, such as the high quality finish segment D {V11; V12}. The high quality finish segment C vehicle is grouped with the coupé-cabriolet and with the five-seat mid quality finish monospace {V7; V15; V16}. Lastly, the seven-seat monospace is grouped with the ICE crossover {V8; V9}.
Use	GWP	V1V2V3V4V10V11V13V17	V5V6V9V15	V7V8V12V14V16	–	–	–	–	Vehicle groups are made according to the fuel consumption to which the GWP value is linearly linked on the well to tank and TtW steps. Vehicles whose consumption is lower than 4.5 L / 100 km are grouped together {V1; V2; V3; V4; V10; V11; V13; V17}, such as the vehicles whose consumption is between 5.2 and 6.1 L / 100 km, and as the vehicles whose consumption is higher than 6.4 L / 100 km.
	EP	V1V2V10V11V13V17	V8V12V16	V3V4	V5V6V7V9V14V15	–	–	–	The prime separation criterion is the nature of the fuel (Diesel or gasoline} that clearly impacts the well to wheels step. The second criterion is the NOx emission quantity on the TtW step. The vehicle groups are therefore made according to the fuel consumption and the nature of the fuel: Diesel <  4.4 L / 100 km {V1; V2; V10; V11; V13; V17}, Diesel > 5.2 L / 100 km {V8; V12; V16}, gasoline <  4.5 L / 100 km {V3; V4}, and gasoline > 6 L / 100 km {V5; V6; V7; V9; V14; V15}
POCP		V3V9	V1V2V6V15V16	V4V5V7V8V11V14	V10V12V17	V13	–	–	The value of the POCP on the TtW is a function of CO, HC and NOx emissions. No group can be identified according to the vehicle characteristics of Table 1.
ADP		V1V2V4V5V6V7V9V14V15V17	V3V8V10V11V12V13V16	–	–	–	–	–	The environmental indicator is influenced by the maintenance phase, especially by the change of the lead battery. No group can be identified according to the vehicle characteristics of Table 1.
End of life	GWP	V8V10	V9V11V12V13V16V17	V5V6V7V15	V1V3V4V14	V2	–	–	Vehicle groups are made per segment: B {V1; V3; V4; V14}, C {V5; V6; V7; V15}, D with crossovers and the five-seat monospace {V9; V11; V12; V13; V16; V17}. One segment B vehicle is alone {V2}, and the seven-seat monospace is grouped with the heaviest vehicle {V8; V10}.
	EP	V8V9V10V16V17	V7V11V12V13	V1V3V4V5V6V15	V2V14	–	–	–	Vehicle groups are made per segment: B and C {V1; V3; V4; V5; V6; V15}, two from segment B {V2; V14}, crossovers and monospaces {V8; V9; V10; V16; V17}, and segment D with one segment C {V7; V11; V12; V13}.
	POCP	V8V9V10V16V17	V11V12V13	V5V7V15	V1V3V4V6	V2V14	–	–	Vehicle groups are made per segment: B and one C {V1; V3; V4; V6}, two from segment B {V2; V14}, C {V5; V7; V15}, D {V11; V12; V13} and monospaces with crossovers {V8; V9; V10; V16; V17}.
	ADP	V8V9V10V11V12V13V16V17	V1V2V3V4V14	V5V6V7V15	–	–	–	–	Vehicle groups are made per segment: B{V1; V2; V3; V4; V14}, C {V5; V6; V7; V15} and D with monospaces and crossovers {V8; V9; V11; V12; V13; V16; V17}.