First published on: 19 February 2010

Transcription

1 This article was downloaded by: [Bibliotheek TU Delft] On: 6 September 2010 Access details: Access Details: [subscription number ] Publisher Taylor & Francis Informa Ltd Registered in England and Wales Registered Number: Registered office: Mortimer House, Mortimer Street, London W1T 3JH, UK Journal of Engineering Design Publication details, including instructions for authors and subscription information: A comparison of methodologies for designing for human variability Christopher J. Garneau a ; Matthew B. Parkinson ab a Department of Mechanical Engineering, Pennsylvania State University, University Park, PA, USA b Engineering Design Program, The Pennsylvania State University, University Park, PA, USA First published on: 19 February 2010 To cite this Article Garneau, Christopher J. and Parkinson, Matthew B.(2010) 'A comparison of methodologies for designing for human variability', Journal of Engineering Design,, First published on: 19 February 2010 (ifirst) To link to this Article: DOI: / URL: PLEASE SCROLL DOWN FOR ARTICLE Full terms and conditions of use: This article may be used for research, teaching and private study purposes. Any substantial or systematic reproduction, re-distribution, re-selling, loan or sub-licensing, systematic supply or distribution in any form to anyone is expressly forbidden. The publisher does not give any warranty express or implied or make any representation that the contents will be complete or accurate or up to date. The accuracy of any instructions, formulae and drug doses should be independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims, proceedings, demand or costs or damages whatsoever or howsoever caused arising directly or indirectly in connection with or arising out of the use of this material.

2 Journal of Engineering Design Vol. 00, No. 0, Month 2010, 1 17 A comparison of methodologies for designing for human variability Christopher J. Garneau a and Matthew B. Parkinson a,b * a Department of Mechanical Engineering, Pennsylvania State University, University Park, PA 16802, USA; b Engineering Design Program, The Pennsylvania State University, University Park, PA 16802, USA (Received 17 October 2009; final version received 4 December 2009 ) In the design of artefacts that interact with people, the spatial dimensions of the user population are often used to size and engineer the artefact. The variability in anthropometry indicates the fixed allocation of space, adjustability requirements, or how many sizes are needed to accommodate the intended user population. Various tools are used to achieve this goal, including boundary manikins, digital human models, prototypes and population models, and hybrid methods that combine the approaches. The present work explores each of these and their relative strengths and weaknesses. This is done in the context of a univariate case study involving the adjustability requirements of a stationary bicycle. An experiment involving 51 individuals was conducted to obtain the data necessary for utilising and evaluating the methods. Keywords: anthropometry; boundary manikins; population models; design methods; spatial analyses 1. Introduction When considering the human user of a product, designs are often assessed in terms of cost, fit, safety, and other performance metrics. An understanding of body dimensions, capabilities, and other characteristics of the population of potential users can assist engineers and designers in creating artefacts that meet these goals. Research into methods for quantifying the variability in a population has led to tools for predicting and assessing accommodation, the degree to which a design meets the needs of its users. Many recommended tools and practices are in common use today (Roe 1993, HFES 300 Committee 2004, SAE International 2009). These tools describe how to consider the variability in anthropometry, capability, and age, among other factors, in the target user population. The application of design automation tools facilitates the simultaneous consideration of these and other aspects of designs (Michalek et al. 2006, Osteras et al. 2006, van der Vegte and Horvath 2006, Zou and Mahadevan 2006, Parkinson et al. 2007). Although the ideal is a truly universal design that accommodates everyone equally well, practical limitations such as cost, development time, and conflicting user requirements make this impossible. Instead, the designer or ergonomist selects a design topology and dimensionally *Corresponding author. parkinson@psu.edu ISSN print/issn online 2010 Taylor & Francis DOI: /

3 2 C.J. Garneau and M.B. Parkinson optimises the product, task, or environment, with the objective of achieving some level of accommodation for its target users (Roe 1993). This is often achieved through adjustability, the creation of separate sizes, reconfigurability, or some combination of these attributes. For the purposes of this paper, this process will be called artefact design. There are various methods for designing artefacts and each has its strengths and weaknesses, and each is likely to produce differing solutions. These methods can be usefully partitioned into three categories: manikin, population model, and hybrid approaches. Each type is explored in this work to provide engineers with general guidance for making design decisions in the selection and application of an appropriate method Manikin-based approaches Manikins originated as 2D templates or kinematic linkages representing the spatial requirements of users (Diffrient et al. 1981). Their use has evolved to include 3D manikins, typically as digital human models (DHM). To facilitate analysis, the number of manikins used to evaluate a design is often only two or three. These are usually selected to lie at the boundary of the observed variability, so typical analysis might use a small female and a large male as boundary manikins. For a univariate problem, the sizes are selected so that the range of variability across the manikins represents the specified accommodation range (HFES 300 Committee 2004). For example, accommodation of 95% of the population might guide the designer to create manikins at the 2.5th- and 97.5th-percentiles of the measure of interest (e.g. stature). Typically the central (i.e th), lower (0 95th), or upper (5 100th) ranges are selected since they are most likely to reduce the material or adjustability requirements. It should be noted, however, that any pairing spanning the desired range could be used. There are many examples of manikin use in the literature. Among those, Das and Sengupta (1996) provide an outline for using boundary manikins and population-dependent anthropometry for workstation design. Schultz et al. (1998) use the same sort of boundary manikin approach to find the range of optimum viewing angles for retail touchscreen point-of-sale systems. Colombo and Cugini (2005) demonstrate the use of digital human modelling with a virtual prototype to evaluate vision constraints in a car and to optimise comfort for users of an industrial riveting workstation. The specifications (e.g. dimensions or range of adjustability) of the artefact are then determined through a virtual fitting trial in which the designer imposes some interaction between a prototype and the manikins. The dimensions of the artefact are prescribed by the spatial or other requirements of the boundary manikin during this interaction. The following manikin-based approaches were examined in the present work: Manikins, using proportionality constants. In this approach, the segment lengths of boundary manikins are determined using proportionality constants, which represent the average length of a particular body segment of interest as a proportion of stature. Manikins, using a database. This variation of the manikin approach uses relevant dimensions taken directly from an anthropometric database. Databases of measures collected from military personnel are commonly used because they contain a wide array of measures from a large population. Two recent surveys are ANSUR (Gordon et al. 1989) and CAESAR (Harrison and Robinette 2002). Manikins, using DHM. DHM translates the basic manikin approach to a virtual environment. Virtual manikins are created within a DHM package (e.g. Jack, SafeWork, etc.). These virtual manikins are then manipulated to fit a virtual prototype of the artefact being designed. Manikin-based approaches can be univariate (having the capability to consider only one body dimension at a time) or multivariate. Simultaneously considering multiple body dimensions

4 Journal of Engineering Design 3 introduces specific complexities that must be considered by an appropriate multivariate method (Haslegrave 1986, Bittner 2000). This is not explored here, since these methods are well established. Instead, this work investigates the application of basic manikin approaches to a simple univariate problem that illustrates the issues at hand Population model approaches Boundary manikin approaches simulate the interaction of a virtual user with the artefact being designed. In population model approaches, the interaction between users and a prototype or representative artefact is measured directly. Consequently, an attempt is made to match the experimental participants to the prospective target user population to improve validity of the results (Roe 1993). Models based on these results are created and used to prescribe the final design such that some percentage of the target user population is accommodated. This method requires both a sufficiently large representative sample population and a workable prototype. It forms the basis for the SAE International recommended practices (SAE International 2009), which are used for vehicle design. Other examples include Eksioglu (2004), which uses a sample of 12 users to determine the optimum grip span range for hand-held artefacts, such as portable power tools, and Meunier et al. (2000), which measures the standoff distance of 30 sample users to determine appropriate dimensions for various sizes of ballistic helmets. Since the use of a population model technique does not require assumptions about user anthropometry or behaviour, variability in the selections of a population that is not attributable to anthropometry may be captured. For instance, needs related to the culture or age of the users may become apparent by direct observation or measurement of user-device interaction (Siu 2005). The following population model approaches were examined in the present work: Population model, direct. A group of individuals representing the target-user population interact with a prototype or representation of the artefact and their preferred interaction is recorded. The spatial requirements are determined directly from the data. Population model, assumption of normality. Data are collected as above, but rather than using the data directly, a statistical analysis is performed. An underlying distribution (e.g. a normal distribution) is assumed and the limits of the design for a particular parameter may be determined by a desired number of standard deviations from the mean. Population models are often analysed in this way because sample sizes are typically small and, in such cases, direct population models would be impractical Hybrid approaches Hybrid approaches combine the reconfigurable nature of the manikin-based methods with the actual interaction between user and device of the population models. In these methods, statistical models relate the outcome measure of interest (measured in an experimental study) with user anthropometry. The primary advantage of this approach is that the model can be extended to new populations for which the anthropometry is known. For example, an experimental study involving preferred fore aft seat location in a vehicle could be conducted using a population of drivers and a representative vehicle. A regression model relating stature to preferred seat location could be created, and virtual fits using boundary manikins (as in the manikin-based approaches) conducted. Using known or prescribed stature distributions as inputs to the model, the minimum and maximum locations that the target percentage of any given population might prefer can be calculated. As an example of this methodology, Christiaans and Bremner (1998) evaluate the correlation between various preferred bicycle dimensions with relevant measures of anthropometry in

5 4 C.J. Garneau and M.B. Parkinson an experimental study using a highly configurable bicycle simulator. Regression analysis on these parameters is carried out to demonstrate a low correlation for all but one or two dimensions. Anthropometry-independent preference effects are shown to greatly influence most bicycle parameters, with the exception of saddle height, which is found to be strongly predicted by the crotch height. Although the regression analysis is not used for design, this study is interesting because it correlates user-selected parameters with measures of anthropometry. Additionally, the study is relevant to the bicycle study used in this paper and could guide future work there. Another example is Yakao et al. (1996), which investigates the relationship between hand length and handle diameter for cylindrical tool handles. Regression analysis is performed on the two measures to find the optimal handle diameter given the study population s mean hand length. These results could be extended to new populations provided the necessary anthropometric predictor (hand length) is available and the behaviour is assumed to be consistent across populations. The following hybrid approaches were examined in the present work: Hybrid, mean behaviour. In a hybrid of the manikin and population model approaches, the results of the population model and associated experiment are parameterised with respect to some anthropometry of the participants (e.g. stature). Regression analysis relates to the average selected artefact dimension, the dependent measure, against some measure of the user, the independent measure. Then, the relevant dimensions of the boundary manikins are entered into the regression equation to determine the limits of the design for the relevant parameter. For instance, in the fore aft seat location example discussed earlier, the regression analysis would model the selected seat position as a function of stature. Design limits of seat adjustment would be determined by entering 2.5th- and 97.5th-percentile stature into the regression equation. Hybrid, including residual variance. This extension of the mean-hybrid approach retains the residual variance from the regression. It is included via a stochastic component of behaviour prediction that accounts for the variability in behaviour of two users of the same size. A large virtual population is then created that mimics the shape and scatter of the original data points, but is reconfigurable to new populations. To determine design limits, the selections of a desired percentage of virtual users may be retained (e.g. keep 5th-percentile to 95th-percentile values). In the fore aft seat location example, the regression analysis would model selected seat position as a function of stature. A large number of statures, composing the virtual population, would be entered into the regression to determine the preferred seat location for each stature. The desired portion of seat locations (e.g. 95%) would then be retained to determine design limits. Each of these methods will be discussed in further detail throughout this paper. The case study considered is the design of an upright stationary bicycle with respect to the minimum height and range of adjustability of the seat. This simple problem was selected because the preferred bicycle seat height for an individual is widely assumed to be predicted well by a single anthropometric measure such as stature or leg length. This allows a relatively unfettered comparison of the methods, reducing the number of factors influencing the results. It is acknowledged that multivariate design scenarios introduce specific complications. These are not considered here since the complications that arise within a single method when extending from univariate cases to multivariate ones, particularly with respect to multivariate accommodation of anthropometric variability, are well-documented (Moroney and Smith 1972, Hudson et al. 1998). Instead the focus is on generalisable conclusions across methods. This case study is used only for the illustration of various design methods. The actual results are highly dependent on the product and task to be performed; general conclusions regarding the exercise cycle seat height may not be drawn from this study. The design dimension for this case study is the vertical height of the seat with respect to the ground, H ground, as shown in Figure 1. Another measure, termed inclined seat height, H inclined,is

6 Journal of Engineering Design 5 H inclined H ground (design variable) Figure 1. Illustration of relevant parameters: inclined seat height, H inclined, and seat height with respect to the ground, H ground. H ground is the design dimension in the analysis presented here. defined as the distance along the seat tube from the pedal axle to the top of the seat. The seat tube is inclined at an angle of 68 from the ground. Figure 1 indicates these dimensions. The target users are taken to be the male ANSUR population (Gordon et al. 1989), selected because of the extensive anthropometric measures available for the sample. In each scenario or application of a design method, the objective is to obtain an accommodation level of 95% of the target user population. This level of accommodation was chosen arbitrarily to illustrate the methods in this paper; some other level of accommodation (e.g. 90%) could have been specified with the result of different numerical results, but no impact on the implementation of the methods. However, in practice, the choice of the accommodation goal is largely specific to a given product or firm. It is a decision that is best left to experienced designers or managers with an understanding of the trade-off between adjustability and cost. 2. Case study experimental setup Several of the identified methods require experimental data. These data were obtained using a prototype and participants similar to those of the target population for whom the product is being designed. The stature and preferred stationary bicycle seat height were measured in a study conducted at The Pennsylvania State University. Although both males and females were sampled in the study, only the data from the males (51 individuals, aged years) were considered for this work. This simplified the presentation when considering the various design methods and should not impact the validity of the results or the methodology because the design target population is consistently defined to include only males. If the intent of the design were to include both males and females in the target population, the models would need to also consider experimental data from the females. The study received approval from an appropriately constituted internal review board. The bicycle seat was randomly set to its highest or lowest setting. Each participant was asked to make an initial adjustment to the seat, then get on the stationary bicycle and pedal a few revolutions as in a first fit when purchasing a bicycle or exercising at the gym. They were then free to adjust the seat height to a more comfortable setting. This process could be repeated as many times as necessary to achieve a desired position. After completing this task, the height of the seat top from

7 6 C.J. Garneau and M.B. Parkinson the ground was recorded for each participant. It should be noted that the seat post was unmodified, and the original pre-drilled stops were used, providing an adjustable range of 243 mm in discrete increments of 24.3 mm. No member of the sample noted a preference for a higher or lower setting than those available. 3. Design methods and results Each of the methods outlined in the introduction was used to make a design recommendation typical of its use. These methods are described below, first generally and then in the context of their use in this particular example. Additionally, the tools and information required to conduct the analysis as well as its results are provided Manikins derived from proportionality constants Stature and weight data are available in nearly all anthropometric databases (Centers for Disease Control and Prevention 2004). Therefore, distributions of those variables are often used to determine the sizes of the small and large virtual users. In the event that the true measure of interest is something besides stature, proportionality constants (Figure 2) are often used (Drillis and Contini 1966). These represent the average length of a particular body segment as a proportion of stature. In this case study, boundary manikins with dimensions belonging to the 2.5th- and 97.5th-percentile person are taken to represent the small and large extremes of 95% of the population. In each approach using manikins to design the stationary bicycle, the manner in which users will sit on the bicycle, and therefore their desired seat height, is entirely unknown. It is the Figure 2. Body lengths expressed as a proportion of stature [8]..039

8 Journal of Engineering Design 7 responsibility of the designer to select a posture that would most likely be representative of target users. Common methods for optimally sizing a bicycle saddle height reflect objectives of maximising power output or efficiency. To do this, relevant leg dimensions are multiplied by predetermined constants (Christiaans and Bremner 1998). One method makes use of the crotch height (i.e. inseam length) multiplied by a factor of 1.09 to determine the inclined seat height (Hamley and Thomas 1967). This method has the advantage that the crotch height is easy to measure when fitting a particular rider, but will usually result in slightly lower saddle heights than comparable methods. Another method makes use of trochanteric (hip) height multiplied by a factor of 1.05 (Nordeen-Snyder 1977) to determine the inclined seat height. This method has an advantage for the current design problem in that a proportionality constant for trochanteric height is given by Figure 2, and so it is this strategy that is used here. The first manikin approach for designing the stationary bicycle uses two boundary manikins with trochanteric height derived from the 2.5th- and 97.5th-percentile stature (for males) using the proportionality constants in Figure 2. The 2.5th- and 97.5th-percentile values of the male stature from ANSUR are 1625 and 1887 mm, respectively. Trochanteric height is found as a proportion of stature (T = 0.530S), and so the lower and upper values are 861 and 1000 mm. Equation (1) uses the method given above with a constant value of 1.05 to determine the inclined seat height, with 25 mm of shoe thickness added. Equation (2) converts the inclined seat height to the design dimension (seat height with respect to the ground) using the geometric relationship of the variables. H inclined = 1.05(0.53 S ANSUR ± ks ANSUR + 25) (1) H ground = sin 68 (H inclined 0.155) (2) In Equation (1), S ANSUR indicates the mean stature for men in the ANSUR database, and s ANSUR indicates the standard deviation of that stature. k is the k-value for the desired percentile (k = 1.96 for 95% accommodation). In Equation (2), 68 is the angle of inclination of the seat tube, mm is the length of the crank, and mm is the height of the crank axle off the ground. Performing these calculations for the given percentiles yields a lower limit of 862 mm and an upper limit of 997 mm. This approach is termed Manikin-k (boundary manikins-using proportionality constant k) in later comparisons Manikins derived from an anthropometric database The second manikin approach for designing the stationary bicycle uses two boundary manikins with a trochanteric height taken directly from the actual 2.5th- and 97.5th-percentile trochanteric height from ANSUR. The 2.5th- and 97.5th-percentile trochanteric height measures are 835 and 1022 mm. Equation (3) uses the method given above with a constant value of 1.05 to determine an inclined seat height, with the 25 mm shoe thickness added. Equation (2) converts an inclined seat height to the design dimension, as before. H inclined = 1.05(0.53 T ANSUR ± kt ANSUR + 25) (3) In Equation (3), T ANSUR indicates the mean trochanteric height for men in theansur database, and t ANSUR indicates the standard deviation of that trochanteric height. k is the k-value for the desired percentile (k = 1.96). Performing these calculations for the given percentiles yields a lower limit of 836 mm and an upper limit of 1018 mm. This approach is termed Manikin-ANSUR in later comparisons.

9 8 C.J. Garneau and M.B. Parkinson 3.3. Manikins implemented as DHMs Digital human modelling or DHM, expands upon simple manikin solutions. DHM packages use a CAD environment and 3D computerised manikins (e.g. Jack, UGS, Ramsis etc.). The required elements of a design may be created in the CAD environment and the manner in which a virtual user interacts with these elements may be simulated. Among the advantages of DHM is the immediate graphical feedback of the proposed solution. Also, they provide for multivariate assessment; once a DHM is postured, multiple geometric constraints can be simultaneously evaluated for a single manikin. Figure 3 shows two boundary manikins seated on a virtual stationary bicycle prototype in a DHM package. The bicycle used in the simulation is similar to the stationary bicycle used in subsequent methods, as it is a traditional upright bicycle with identical geometry relevant to the problem. Boundary manikins with a stature of 1625 and 1887 mm are used in the simulation, and are postured on the virtual bicycle in such a way that appears appropriate. For each manikin, a seat height is selected by the designer that allows for a slight bend in the knee in the more fully extended leg with the pedal cranks oriented vertically. The limits of adjustability are then determined by these designer-selected seat heights, which are 862 and 1012 mm for the small and large manikins, respectively. This approach is termed DHM in later comparisons. To use the DHM method, appropriate software is required. This includes both a DHM package (e.g. Jack) and a CAD program to create the artefact (e.g. Solidworks). Additionally, the sizes of the virtual boundary manikins must be determined. Typically designers select from built-in figures created to represent a spectrum of overall stature and mass combinations that might be observed in a population. Figure 3. Two male human figure models are virtually postured to select the upper and lower adjustability requirements.

10 3.4. Population model data used directly Journal of Engineering Design 9 To use any population model approach, a prototype of the artefact being designed and a group of sample users are required. This usually requires approval of an internal review board to use human subjects. Tools to assist with data analysis (e.g. R, Excel, SAS, etc.) are also helpful. To apply the direct population model to the bicycle seat design problem, the selections from the group of 51 individuals representing the target user population are recorded and ordered such that approximately 95% of the central selections are retained. This means that the selected seat height of the second person and 50th person define the limits for 95% accommodation, and this gives a lower limit of 867 mm and an upper limit of 1038 mm. These results are termed Population-direct in later comparisons Population model data used with an assumption of underlying normality To create the population model for the bicycle design, statistical analysis is performed directly on the experimental data described earlier. The data were assumed to be normally distributed and the 2.5th- and 97.5th-percentile values are determined using the mean and standard deviation. The mean of the seat height selections of the sample population is 951 mm and the standard deviation is 50 mm. Using a k-value of 1.96, the lower limit for 95% accommodation is 852 mm and the upper limit is 1050 mm. These results are termed the Population-normal in later comparisons Hybrid approach with a mean behaviour model In a hybrid of the manikin and population model approaches applied to the bicycle design, linear regression analysis is performed using the selected seat height and stature for the sample to create a seat height preference model. Stature is used as the predictor, instead of hip height, since stature data are widely available for a variety of populations. The resulting regression line is used to predict the selected seat height of two boundary manikins, characterised by 2.5th- and 97.5th-percentile stature. This approach is termed the Hybrid-mean in later comparisons. Figure 4 shows the plot of selected seat height versus the stature (with shoes) for the sample population. The equation for the resulting regression line is given by H ground = 0.476S (4) where H ground is the selected seat height position with respect to the ground and S is the stature. R 2 for the regression is 0.42 and the RMSE is 38.8 mm. Entering the 2.5th- and 97.5th-percentile statures (with shoe thickness added) into Equation (4) gives 885 and 1007 mm, respectively. These values define the limits of the required seat height adjustment for the Hybrid-mean method. As with the population model approach, a prototype of the artefact being designed and a group of sample users are required to use this hybrid approach and tools to assist with data analysis are helpful. Additionally, a tool to measure the relevant characteristics of the participants is required. If the characteristics to be determined are lengths, an anthropometer or similar measuring device is required. Also, anthropometry of the boundary manikins must be determined by referencing an anthropometric database Hybrid approach with a model considering preference Using a methodology described in Parkinson et al. (2007) and Parkinson and Reed (2006a), this approach for the bicycle design uses the regression parameters of the seat height preference model to generate a virtual population of 1000 users randomly sampled from theansur database.

11 10 C.J. Garneau and M.B. Parkinson 1050 seat height (mm) Figure 4. stature (mm) Selected seat height plotted against stature for the 51-member sample, with regression line. The preferred seat height includes a stochastic component that describes how the preferred seat height of the virtual user deviates from the mean observed height for someone of that stature. This accounts for the variability in the preferred height observed within individuals of the same stature. The magnitude of this stochastic term is determined by randomly sampling from a normal distribution with a mean of 0 and a standard deviation equal to the square root of the residual variance in the regression model (Flannagan et al. 1998). For this reason, this method is termed the Hybrid-resvar method in later comparisons. The virtual users were selected randomly from the ANSUR database without replacement. The mean of their stature distribution is 1756 mm, identical to the mean of all males in ANSUR. Figure 5(a) shows the result of plotting these statures using the regression results in Equation (4). Introducing an anthropometry-independent preference component by way of a term representing the residual variance to describe deviation from the mean gives H ground = 0.476S N(0, 38.8) (5) where N is a normal distribution with a mean of 0 and a standard deviation of 38.8 (RMSE of the regression). Figure 5(b) shows the 1000 seat positions predicted for the 1000 statures using seat height (mm) stature (mm) (a) stature (mm) (b) Figure 5. Results of the Hybrid-mean and Hybrid-resvar methods, respectively. (a) Seat height setting for 1000 random statures, located with regression equation with no anthropometry-independent preference component (Equation (4)). (b) Seat height setting of the 1000 member sample of the Hybrid-resvar method are plotted with anthropometry-independent preference included (Equation (5)). Adjustment limits defined by the central 950 users are denoted by the horizontal lines extending across the plot.

12 Journal of Engineering Design 11 seat height (mm) manikin ANSUR population direct hybrid mean manikin k DHM population normal hybrid resvar Figure 6. The relative adjustability ranges prescribed by the Manikin-k, Manikin-ANSUR, DHM, Population-direct, Population-normal, Hybrid-mean, and Hybrid-resvar methods. Each approach is intended to obtain the same accommodation level (the central 95% of users). Equation (5). An accommodation of 95% is achieved by selecting the central 950 users with their seat height selections placed in order. The lower adjustment limit is determined to be 850 mm and the upper limit is 1044 mm. To use this advanced hybrid method, a prototype of the artefact being designed and a group of sample users are required, and a tool (e.g. anthropometer) to measure their characteristics is also necessary. Additionally, anthropometry of the boundary manikins must be determined by referencing an anthropometric database. Statistical analysis software, such as Matlab, is required to efficiently analyse the large virtual population Combined results Figure 6, which depicts the adjustable range determined by using each method, shows the differences among the various solutions. The length of each bar indicates the amount of adjustability recommended by the application of the corresponding method. Plotting the results of all the methods in this way, on the same axes, facilitates comparison among them and also highlights differences. For instance, Figure 6 makes it clear that the Population-normal method prescribes the greatest amount of adjustability, and the Hybrid-mean method the least. 4. Discussion The various methods require differing amounts of time and effort. The methods that require an experiment introduce considerable cost and challenge. One might conclude that the method that requires the most time and effort would yield the best solution but that is not necessarily the case since a quickly obtained solution estimate may be sufficient. Designers might conduct some cost/benefit analysis to determine which solution strategy is most appropriate. Using the Hybrid-resvar method as a benchmark, the methods are compared to determine actual versus predicted (95% for each method) accommodation levels in Table 1. The range of adjustment prescribed by each method is also indicated. The results show a wide range of recommended designs which these various methods all of which are in common use provide to this extremely simple univariate design problem. A discussion of the strengths and weaknesses of each approach follows.

13 12 C.J. Garneau and M.B. Parkinson Table 1. Summary of the results for all methods: adjustable range and % accommodated (compared with the Hybrid-resvar method). Method Range % Accomodated Manikin-k Manikin-ANSUR DHM Population-direct Population-normal Hybrid-mean Hybrid-resvar Note: Dimensions in mm Manikin-based approaches Table 1 indicates that the Manikin-ANSUR and Population-normal methods yield, for this example, solutions that are very close to that of the Hybrid-resvar method. The difference in the adjustable range is only 12 mm for the Manikin-ANSUR method and 4 mm for the Population-normal method, compared with the Hybrid-resvar method. Since this is a small difference in a very simple problem, the Manikin-ANSUR method might be most appropriate, since it requires the least amount of effort for a reasonable solution. However, this similarity in solutions is likely attributable to the simple and univariate nature of this design problem. Although these small differences may not be significant for this problem, many other applications including those requiring multivariate accommodation or that have limited physical space for adjustability or high costs for added adjustability benefit greatly from optimised solutions. Using manikins with dimensions derived from an anthropometric database, as in the Manikin- ANSUR method, is often preferable to using manikins with proportionality constants, as in the Manikin-k method. This is because in a single individual, the proportional lengths vary widely from the mean. Neither is there a standard person with all dimensions belonging to the same percentile (Roebuck 1995). For example, when a manikin representing the body as a kinematic linkage is scaled so that an overall dimension, such as stature, meets some target percentile, the body dimensions that make up the aggregate dimension do not themselves define useful design limits. That is, a person who is 5th-percentile by stature has other body dimensions that vary widely from the 5th-percentile value for those measures (Fromuth and Parkinson 2008). Realistic posturing is required when using manikins in design (Parkinson et al. 2007). This posturing may be performed manually by the designer, as in the manikin methods presented in this work, but to improve accuracy and repeatability it is done increasingly algorithmically. One method for creating such an algorithm involves gathering data on a number of people performing similar tasks to the one in question (e.g. vehicle ingress/egress, interacting with an artefact, etc.). Statistical models of these data are then created to describe the mean anticipated behaviour as a function of anthropometric and task conditions (Faraway et al. 1999, Chaffin et al. 2000). This assumes that user interactions can be predicted by the anthropometry and geometric constraints of the artefact, task, or environment. Research in manikin-based approaches has shown that while they might be sufficient for univariate analyses, that particular care must be taken for multivariate problems. Since measures of interest such as stature, sitting height, and hip breadth are not perfectly correlated, it is unlikely that individuals disaccommodated on one measure will be exactly the same as those disaccommodated on the others. When spatial requirements are determined for multivariate problems using a series of univariate analyses, the actual accommodation is much less than predicted. Although this limitation has been recognised for decades (Moroney and Smith 1972), the method is still often misapplied. There are a number of multivariate approaches that properly consider anthropometric variability (Flannagan et al. 1998, Hudson et al. 1998, Friess 2005).

14 Journal of Engineering Design 13 Digital human modelling is often used to solve a variety of design problems (Quick et al. 2005, Reed et al. 2005). Although DHM appears to be superior to working out manikin dimensions by hand, the impressive visuals created by DHM packages in many ways obfuscate the task of prescribing only as much adjustment as necessary for a particular user group when designing a product. Since DHM often requires manual positioning by the designer, decisions are made in the positioning of the manikins that can drastically alter the design solution, and so DHM is very sensitive to the posturing decisions of the designer. This introduces errors in consistency and repeatability both within and across designers and ergonomists. An example of this sensitivity is apparent in Stefani et al. (2007). In that study, the use of an automobile head restraint as a safety device is investigated using a small sample of volunteers seated in an actual automobile. One measure obtained from the sample was backset, or the distance between the back of the head and the head restraint. When Jack, a popular DHM package, was used to predict backset for a group of manikins with similar dimensions to the actual sample, different results were obtained based upon the posture of the manikin. While the idea of modelling a large population with a very small set of manikins (manually or virtually) is an attractive one, the large amounts of variance that are not explained by body size or geometric constraints make it impractical. This finding is consistent with those in other studies involving more complex problems (Reed et al. 2000, Parkinson and Reed 2006b) Population model approaches Population models are an improvement on manikin-based approaches in some ways, since they specifically model the outcome measure of interest, e.g. reach, eye location, driver-selected seat position, rather than trying to predict the population distributions of those outcomes from boundary cases defined by anthropometry. These models have additional requirements beyond basic manikin approaches, however. They (1) require extensive human-subject data from a similar task scenario, parameterised using the design variables of interest and representing a large amount of variability on the population descriptors, e.g. body dimensions, (2) have not historically been parameterised for population attributes, although recent models (new J941 and J4004) overcome these limitations, and (3) they are essentially univariate, dealing with only a single outcome measure (e.g. preferred seat height) at one time. Since the Population and Hybrid-mean methods are based upon functional models in which the preferred seat heights were measured, one might anticipate those values to better represent the required adjustability. Additionally, since the behaviour is measured directly, no assumptions (e.g. riders will select a seat height such that their heels can reach the floor) are required. Due to the simplicity of the problem and the relative similarities between the sample and target user (ANSUR) populations, the Population-normal method predicted accommodation well. This is to be expected. Difficulties in the use of population models arise in two situations: (1) when designers try to use a model gathered on one population to predict accommodation for another; and (2) when designers consider explicitly univariate models simultaneously. For example, bicycle models predicting preferred handle bar location and preferred seat location might work well independently, but the results cannot be combined to accurately predict both (Parkinson et al. 2007) Hybrid approaches Using a hybrid model that only considers the mean behaviour, as in the Hybrid-mean method, results in a practice where artefacts are designed to meet the mean behaviour associated with a particular body size (i.e. two people with the same predictor value, such as stature, will have the same predicted performance). This ignores the residual variance in the experimental data. Reed et al.

15 14 C.J. Garneau and M.B. Parkinson (2000) discuss some limitations of using such a mean behaviour approach and provides motivation for incorporating into the model the variability that cannot be explained by the predictors. For example, the use of the Hybrid-resvar method produced a regression model relating stature to seat height. It has an R 2 -value of 0.42, so more than 50% of the variance in the seat height cannot be explained by the stature. The effects of anthropometry-independent preference are clearly visible when comparing Figure 5(a), which includes no residual variance in the model, with Figure 5(b), which does. It shows that when it is included, for any particular stature, a variety of seat positions might be chosen about a mean value. The estimated benefit of incorporating a stochastic component may be seen in the 18.1% increase in accommodation from the Hybridmean method to the Hybrid-resvar method, corresponding to 181 additional people that are now accommodated from the virtual 1000-member sample. Another way to think of the Hybrid-resvar method is the separation of the user behaviour into two quantifiable components: that related to anthropometry (measured by the slope of the preference model regression) and that attributable to unquantified differences in users (measured by RMSE). The interrelation of these two terms may be easily examined by the shape and scatter of the preference model. Sometimes, anthropometry is a primary consideration, and in such cases, the model will exhibit a steep slope with a high R 2 -value. In these situations, the manikin-based approaches might be sufficiently accurate. Other times, the anthropometry has little impact on the problem, and the primary consideration may be other factors. In these cases, the model will exhibit a very shallow slope with a low R 2 and lots of scatter. Both the population model and hybrid approaches, which involve user testing, might improve accuracy sufficiently to justify their additional cost in time and resources. For this example, the R 2 is 0.42, so anthropometry has a bearing on the problem but explains less than half of the observed variability in the user-selected seat height. Therefore, the result of plotting and analysing data from preference models forms a useful tool that aids the design engineer in selecting a solution method appropriate to the problem at hand. Although the Hybrid-resvar method illustrated here still requires experimental tests, one advantage of the method is that data from a relatively small sample of users evaluating physical prototypes can be used to make models that can be used for quantitative analysis. Sampling 1000 users for analysis in this study would be extremely impractical, whereas sampling roughly 50 people is much more manageable. In practice, the sample population would deliberately contain a diverse group of people, with care taken to oversample the tails of the distributions of relevant parameters (e.g. lots of short and tall people in this example). Since there is no expectation that all the sources of the model variance be identified, sample populations can be smaller than they might otherwise be. The Hybrid-resvar approach is the most theoretically advanced and acknowledges the complexity of the artefact design problem. In addition to providing a better estimate of adjustability, it also has some practical advantages. Since the selections of a population are parameterised as a function of population descriptors, the model is reconfigurable in that the results from one study may be be judiciously applied to different populations. The method also lends itself well to the application of optimisation methods. The method also requires the most time and expense in its implementation. For simple, well-understood problems, the application of this approach may not be justified, but the advantages of the method for more complex problems might warrant the added expense. 5. Conclusion Several methods for designing the optimal limits of an adjustable artefact for a desired level of accommodation of a population of users have been explored. The intent of this study is to

16 Journal of Engineering Design 15 better inform engineers about available methods for designing products that account for human variability. Comparing the various methods using a simple case study facilitates exploration of the relative advantages and disadvantages of each method, in terms of accuracy, cost, and complexity. It is the responsibility of the designer to select an appropriate method that produces reasonable accuracy with minimum cost and complexity. The results here indicate that the simple manikin and population model approaches sometimes produce accurate solutions, but it is important to recognise that more complex problems introduce additional considerations that may make such simple approaches inappropriate. Methods considering the behaviour of actual users tend to perform better than those requiring assumptions of user behaviour on the part of designers. In general, static assessments of theoretical humandevice interaction, as with the boundary manikin method, provide a solution with limited insight. Dynamic assessments of real people, as with the population and hybrid models, allow for the exploration of preference unattributable to body dimensions, which has been demonstrated to be a significant component of human variability. Even within the context of the specific bicycle example, it has been demonstrated that considering both individual user preference and user anthropometry offers a more accurate method of predicting bicycle configuration than anthropometry alone. Christiaans and Bremner (1998) support this conclusion and recommend the use of a highly adjustable bicycle simulator when fitting a bicycle to its rider, rather than relying upon rules of thumb that consider only anthropometry (although these guidelines may define a useful starting point in bicycle fitting). The sample data used in the bicycle example in this study have limitations. A larger, more diverse sample including females would be required to validate the model across an entire population. The inclusion of different types of bicycles would allow the model to be extrapolated more broadly. The mode of adjustability could also be altered, reconfiguring the bicycle in such a way so as to have continuous, rather than discrete, adjustment. The preferred seat height was selected using a quick sit methodology. This might correlate well with acceptability in a store or for a short ride but is likely to differ from what might be selected for longer duration rides. Additionally, the expertise of the rider was not considered; it is likely that more experienced riders will select relatively higher seat heights because they are more likely to realise the physical benefit of greater leg extension. Although these considerations would become important in the actual design of an stationary bicycle, the limitations do not impact the demonstration of the concepts discussed in this paper. A strictly univariate problem has been considered. One could imagine that a design with more than one adjustable dimension experiences a compounding of the unexplained variance seen in this problem, and examining these effects on higher dimensional problems offers an opportunity for future research. The notion of a just noticeable difference that is the degree to which a user could be disaccommodated without experiencing negative effects, and its impact on preference and accommodation is another aspect of the DfHV problem that has not been considered here but would offer an opportunity for future research. Rather than assuming a binary accommodation solution, wherein users are either strictly accommodated or disaccommodated, it may be insightful to understand how a certain degree of disaccommodation, which could lessen cost while not increasing discomfort appreciably, would impact the solution. References Bittner, A.C., A-CADRE: advanced family of manikins for workstation design. In: Proceedings of human factors and ergonomics society, Long Beach, CA: Human Factors and Ergonomics Society, Centers for Disease Control and Prevention, National health and nutrition examiniation survey , Hyattsville, MD: National Center for Health Statistics. Chaffin, D., et al., Stature, age, and gender effects on reach motion postures. Human Factors, 42 (3),