Quantifying the Relevance of Product Feature Classification in Product Family Design

Abstract

The methodology proposed in this chapter aims to address the link between the evolution of product feature relevance and the implications to product platform and product family design. By quantifying relevant/irrelevant product features to be included in next-generation product platform design, designers can identify the stand-alone or platform sharing components required to achieve desired product functionality. A data mining algorithm is introduced that uses time series data (consisting of product features) to determine the standard, nonstandard, and obsolete product features in the design of next-generation products. Product features are then mapped to engineering components/modules by employing data mining Natural Language Processing techniques that quantify the functionality requirements that are needed for a given set of product features. The goal of this work is to demonstrate the value of incorporating evolving product feature trends in the market space directly into product platform and product family sharing decisions.

These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

1 Introduction

The increase in globalization and the prevalence of low-cost communication infrastructure present ever-increasing challenges for enterprise decision makers aiming to satisfy customer needs. In recent years, companies have had to consider the impacts of a socially connected digital age in shaping customer preferences and expectations in the market space (Tucker and Kim 2011a). The evolution of product preferences in the market space can be highly dynamic and difficult to capture using traditional customer-driven frameworks employed by design engineers. Mass customization has been proposed as a viable approach to accommodate the diverse product preferences in the market space. From an engineering perspective, however, mass customization presents the added challenge of establishing design and manufacturing processes to meet the needs of mass customization. Product family design is an enterprise-driven strategy aimed at mitigating the added costs that arise due to product customization. The two design strategies that have been proposed in the product family design literature are the Bottom-Up approach and the Top-Down approach (Simpson et al. 2001). Commonality indices proposed in the literature investigate component/module sharing strategies for existing products within a product family and are well suited for Bottom-Up product family design. In Top-Down product family design, a product family emerges from an existing market-driven need. The data mining component classification framework in this work will enable designers to identify components that are well suited for sharing in the product family design process through the use of large-scale market-driven product feature preference data.

2 Related Work

This section presents work relevant to the three main aspects of this research: (1) Data mining-driven product design, (2) Translating customer needs into engineering targets, and (3) Product platform and sharing decisions.

2.1 Data Mining-Driven Product Design

Data Mining-Driven Product Design is an emerging field of research aimed at incorporating large-scale data in the design of next-generation products (Braha 2001). Agard and Kusiak employ data mining association rules to cluster product functions in the design of product families (Agard and Kusiak 2004). Tucker and Kim (2008) employ Naive Bayes Classification techniques that enable designers to identify novel product feature combinations in a high dimensional product feature space. Moon et al. (2006) employ data mining Fuzzy c clustering techniques as a platform identification strategy in product family design. Data Mining techniques have been employed by Tucker and Kim to determine the optimal product feature combination for product family optimization (Tucker and Kim 2009; Tucker et al. 2010). Moon proposes a Data Mining framework for extracting design knowledge for product platform and variant design (Wang 2008).

While the aforementioned data mining techniques proposed in the literature aim to address product family design problems, they are static in nature and primarily consider large-scale data at an instant in time, hereby omitting the changes in product feature preferences that may occur in the market space over time. In order to accommodate evolving product trends in the market space, Tucker and Kim propose a temporal product feature classification algorithm that classifies product features as Standard, Nonstandard, or Obsolete, based on their time series predictive power (Tucker and Kim 2011b). The classification of product components will enable design engineers to determine when to retire certain components (classified as obsolete in the metric), include in the design of a product platform (classified as standard in the metric), or aid in the creation of modules for product variants (classified as nonstandard in the metric).

2.2 Translating Customer Needs into Engineering Targets

Quality function deployment (QFD) is a well-established approach employed in the design community for translating customer preference requirements into engineering design targets/functional specifications (Pullmana et al. 2002). A house of quality (HOQ) would be designed, mapping the customer requirements into tangible engineering design targets (Bouchereau and Rowlands 2000). Customer preferences towards certain product features can be weighted through feature rankings acquired through surveys or focus groups (Kwong and Bai 2003). As a result, a QFD matrix can be used to depict the interdependence between customer requirements and the engineering metrics (EM).

The QFD model is highly dependent on the domain expert (engineers) translating the customer wants into engineering metrics. As the complexity of modern technology increases, so does the availability of product features and customization options. The increased product feature space (high dimensional feature space) and the highly dynamic nature of many consumer markets today make traditional translation of customer preferences into engineering metrics cumbersome. Furthermore, the expertise of these techniques is limited to the domain expert(s), making the process highly dependent on a subset of the product development team. By employing a data mining-driven approach to customer preference modeling and then translating the knowledge gained into tangible engineering metrics, design engineers will be able to incorporate market-driven trends during the translation of customer wants into engineering specifications. Instead of relying on survey or focus group feedback in an effort to quantify the evolution of product preferences in the market space, designers can employ the data mining methodology proposed in this chapter as a means of generating predictive models about evolving product feature preferences that can then be used for product family optimization.

2.3 Product Platform and Sharing Decisions

A product platform can be defined as a set of parameters/features or components that are shared across products within a product family (Simpson et al. 2001). Meyer and Lehnerd provide guidelines for product platform development and encourage companies to design products around a shared platform, rather than subsequent independent designs (Meyer and Lehnerd 1997). Commonality refers to the level of sharing of components/subassemblies, processes, etc. across different products within a family of products (Boas 2008). Commonality therefore has the ability of reducing manufacturing and design costs (by sharing the same component across different products), while concurrently providing the level of product diversity expected within the market space. The trade-off between product commonality and product diversity has been studied extensively in the literature and is discussed in Simpson et al. (2001). de Weck highlights the challenges that exist in determining the extent of product platforming in product family design (Simpson et al. 2006).

Several commonality metrics have been proposed in the literature in an effort to quantify the effects of platform sharing decisions on product family design. For example, Collier proposed the degree of commonality index (DCI) as a way to measure the ratio of common components existing among products within a product family to the total number of components (Collier 1981). A modified version of the DCI called the total constant commonality index (TCCI) has absolute bounds (0–1), hereby making commonality comparisons within and between product families more quantifiable (Wacker and Trelevan 1986). The commonality index (CI) proposed by Martin and Ishii measures the ratio of unique components in a product family and the total components in a product family (Martin and Ishii 1996, 1997). The Percent Commonality Index (%C) measures product commonality within a shared product platform, rather than across product families using a weighted sum of multiple variables for a total commonality scale ranging from 0 (no commonality) to 100 (complete commonality) (Siddique et al. 1998). Another extension of the DCI called the component part commonality index (CI^(C)) takes into account factors such as the cost of each component, product volume, and quantity per operation in determining the effects of component sharing decisions on a product family (Jiao and Tseng 2000). The product line commonality index (PCI) is a departure from traditional commonality indices that penalize broad product variation and instead, penalizes products with nonunique components within a product family (Kota et al. 2000). The generational variety index (GVI) proposed by Martin and Ishii measures the level of redesign work needed for future iterations of a product and helps designers determine which components may change over time (Martin and Ishii 2002). The comprehensive metric for commonality (CMC) is a data intensive approach to product commonality based on the components’, size, geometry, material, manufacturing process, assembly, cost, and allowed diversity in the family (Thevenot and Simpson 2006). Alizon et al. (2009) propose a commonality diversity index (CDI) that compares components relating to a specific function(s) and investigates the trade-off between commonality and diversity based on the product family’s functional requirements. With a plethora of commonality metrics proposed in the literature, Simpson et al. (2012) approach the product platforming problem by proposing an integrative approach that incorporates a market segmentation grid, the GVI, design structure matrix (DSM), commonality indices, mathematical modeling and optimization, along with multidimensional data visualization tools.

The methodology proposed in this chapter aims to address the link between the evolution of product feature relevance and the implications to product platform and product family design. Specifically, this work aims to:

• Translate a product feature classification from the market-driven domain to the detailed engineering domain.

• Determine the optimal product platform sharing decisions based on the market-driven evolution of product features and customer preferences.

3 Methodology

The methodology proposed in this work (Fig. 6.1) aims to guide product family design by linking Temporal Market-Driven Responses relating to product feature trends with Engineering Design Optimization objectives such as product platforming and commonality decisions. As presented in Sect. 6.2.3, there are well-established metrics for evaluating commonality decisions in product family design. However, temporal, market-driven forces are typically not included in these models.

Fig. 6.1

Linking market-driven response with engineering design optimization

Market-Driven Responses can be a critical design input to product family design by quantifying the evolution of product features in the market space and identifying product features that are Standard, Nonstandard, or Obsolete. In the proposed methodology, the Product Domain in the Market-Driven Response step refers to the results of a data mining-driven approach to modeling the relevant/irrelevant product features across a wide array of products existing in that domain.

3.1 Level 1: Temporal Market-Driven Preferences

Level 1 of the proposed methodology is based on a knowledge discovery in databases (KDD) framework. KDD is the umbrella term used to describe the sequential steps of Data Acquisition → Data Selection and Cleaning → Data Transformation → Data Mining/Pattern Discovery → finally leading to the Interpretation and Evaluation of the resulting model. This data-driven approach to modeling will enable designers to understand the temporal changes in the market space relating to product preferences and use this knowledge in the design of next-generation product families. The sequence of the KDD steps will now be expounded upon:

3.1.1 KDD Step 1: Data Acquisition

The data in the proposed methodology represents structured, time series data that exists within a company’s database or acquired online through publicly available customer product preference websites using automated data acquisition techniques (Tucker and Kim 2011a). The two types of data employed in the proposed methodology are structured and unstructured data.

Structured data typically refers to data that can be conceptualized using an Entity-Relationship structure and easily stored in a Database Management System (Chen 1976). The Entity-Relationship Model is an example of the format of structured data where the entity (e.g., product domain) is related to certain features (e.g., product features).

Figure 6.2 is an example of time series structured data suitable for the proposed methodology, where each column (1, P) at time t _i is defined by a unique product feature (j). The last column containing the Class variable represents the dependent/output variable which is influenced by the levels/values of the product features. Examples of a class variable could be a market price segment (>$199, $99–$199, $0–$99) or a purchasing decision (purchased, not purchased), etc. The proposed methodology assumes that the product features (j) can be categorical or numeric in nature, while the class variable is considered categorical for the subsequent data mining algorithm. In the proposed methodology, structured data will be used to quantify the relevance of product features in the market space over time (Level 1: Temporal Market-Driven Preferences), which will then help guide Product Family decisions in Level 2 (Engineering Design Objective).

Fig. 6.2

Time series product data containing product features and class

Unstructured data on the other hand refers to data that is not well suited for DBMS due to a lack of a well-formed entity-relation model (Buneman et al. 1996). Unstructured data primarily includes text data found in documents, web pages, numeric values, etc. An example of unstructured data would be a product review containing both textual and numeric information.

As can be seen in Fig. 6.3, the information contained in textual data does not have a well-defined feature/class relation found in Fig. 6.2, hereby making traditional data mining classification algorithms ill-suited for such data. However, unstructured data contains extremely valuable information regarding the domain of investigation and can be mined to quantify patterns using Natural Language Processing techniques that will be presented in the Data Mining step in the KDD process. In the proposed methodology, Natural Language Processing will be employed to understand the relation between product feature and component function in Level 2 (Engineering Design Objectives).

Fig. 6.3

Unstructured data of a cell phone product review

3.1.2 KDD Step 2: Data Selection and Cleaning

The second step in the KDD process aims to minimize noise in the data set that may arise due to missing data values, erroneous/ambiguous features, etc. Data selection and cleaning techniques should be employed for each data type used in the proposed methodology. For categorical features/class found in the structured data in Level 1, missing/erroneous values can be addressed by either replacing them with global constant values or the most probable values (based on the frequency of occurrence of a particular feature/class value) (Han et al. 2011). For unstructured data, used in Level 2 of the methodology, data selection and cleaning techniques may include text grammatical correction processing for nonword error detection, isolated-word error correction, and context-dependent word correction (Kukich 1992). For example, word corrections could be as straightforward as correcting “cel phne → cell phone” or more complex in trying to determine context and correct “real time whether updates → real time weather updates.”

3.1.3 KDD Step 3: Data Transformation

Step 3 of the KDD process is where the data is transformed into acceptable forms for the subsequent Data Mining/Pattern Discovery (Step 4) process. For structured data for example, binning techniques can help smooth data values of a feature by first sorting and placing feature values in predefined bin categories, where each bin category can be represented by the mean of the feature values in the specific bin (Han et al. 2011). For unstructured data, data transformation techniques may include Stemming; a process that aims to reduce morphological variants of words to their root form so that word variants can be mapped together (Paice 1994). For example, the words charger and charging both have the same root word charge, which could refer to a phone charger in product design. Data transformation techniques will reduce the noise caused by redundant feature values or words in a large data set.

3.1.4 KDD Step 4: Data Mining/Pattern Discovery

The Data Mining/Pattern Discovery step in the KDD process is where statistical/machine learning algorithms are employed to the transformed data (from Step 3) in order to discover novel, previously unknown knowledge about the domain of interest. The methodology begins with Phase 1, the iterative evaluation of the relevance of product features to the final class variable. Phase 2 presents the classification of the product features deemed irrelevant by the data mining predictive model that are then classified as Standard, Nonstandard, or Obsolete. The subsequent product family optimization step is guided by the predictive data mining results that are based on the temporal market-driven product preference data.

3.1.4.1 Phase 1: Iterative Feature Evaluation

The feature classification metric is modeled based on a time series decision tree induction algorithm that captures the emerging product feature trends over time (Tucker and Kim 2011b). Phase 1 in Fig. 6.4 sequentially tests each product feature’s entropy (at each iteration of the algorithm) using n time-stamped data sets. The calculation of the entropy values are used to rank each product feature’s relevance to the class variable and also used as the test statistic to classify irrelevant product features in phase 2 of the methodology. In this work, the term relevance is defined as a product feature’s relationship to the class/output variable.

Fig. 6.4

Data Mining model generation based on time series product feature data

Given n time intervals, t ₁ to t _n , each time interval t _i contains a training data set T. For training data set T at time t _i , each of the feature is tested in order to determine that feature’s ability to reduce the uncertainty of the class variable (please see Fig. 6.2). There are several metrics proposed in the literature for evaluating a feature’s relation to a class variable including the Gini Index, Gain Ratio, Likelihood-Ratio Chi-Squared Statistics, DKM Criterion, Twoing Criterion, etc. (Maimon and Rokach 2005). The methodology proposed in this work employs the Gain Ratio metric, although the algorithm is not limited to this metric.

The Gain Ratio is a well-established feature evaluation metric for determining the best split of the data set at each iteration. The assumption is that both the class variable and product features have values that are mutually exclusive of one another. Also, it is assumed that the variables are categorical or if continuous, can be discretized using existing statistical discretization techniques (Dougherty et al. 1995). The goal of the feature classification algorithm is to iteratively test each product feature for its ability to reduce the uncertainty/randomness of the class variable, generate a decision tree model, and then classify the features that do not show up in the resulting decision tree model as Standard, Nonstandard, or Obsolete.

Given a training data T set at time t _i , each with n features (continuous or discrete) and a class variable c _i , the Gain Ratio is defined as (Quinlan 1992):

$$ GainRatio(X)=\frac{{Entropy(T)-Entrop{y_X}(T)}}{Split(T) } $$

(6.1)

where:

$$ Entropy(T)=-\sum\nolimits_{i=1}^q {p({c_i})*lo{g_2}p({c_i})} $$

(6.2)

p (c_i) represents the probability (relative frequency) of a class variable c _i in the training data set T.

q: represents the number of mutually exclusive class values within the data set.

$$ Entrop{y_X}(T)=-\sum\nolimits_{j=1}^j {\frac{{{T_j}}}{T}*Entropy({T_j})} $$

(6.3)

T _j: represents a subset of the training data T that contains one of the mutually exclusive outcomes of a product feature. For example, if product feature X is wireless connectivity containing 3 mutually exclusive outcomes (WiFi, Bluetooth, NFC), then T _j represents all the instances in T that contain one of those outcomes.

J: represents the number of mutually exclusive outcomes for a given feature.

The denominator of the Gain Ratio metric, Split (T) normalizes the numerator, hereby reducing the bias of the metric towards features with a large number of mutually exclusive outcomes (T _j ).

$$ Split(T)=-\sum\nolimits_{j=1}^j {\frac{Tj }{T}*lo{g_2}\frac{Tj }{T}} $$

(6.4)

From time periods t ₁ to t _n, the Gain Ratio values for each product feature are computed and stored. A time series predictive model is then use to predict which product feature will have the maximum Gain Ratio values at future time periods t_n+k, where k represents the length of time before the next generation of products are to be launched (Tucker and Kim 2011b). Therefore feature F _i, appearing at the top of the Data Mining Predictive model in Fig. 6.4, represents the product feature with the highest predicted Gain Ratio, given a history of stored Gain Ratio statistics from time periods t ₁ to t _n. Product feature F _p at iteration 2 represents the product feature with the highest predicted Gain Ratio, given a history stored Gain Ratio statistics from time periods t ₁ to t _n. The algorithm continues to partition the original data time series data sets until a homogeneous class distribution exists for each leaf of the Data Mining Predictive Model, as seen in Fig. 6.4. The resulting model is represented as a decision tree, which can be read as a sequence of decision rules by traversing down each unique path of the tree. The resulting model will help design teams determine the specific product feature combinations that yield a particular outcome in the market space (e.g., price).

3.1.4.2 Phase 2: Model Generation and Irrelevant Feature Classification

Once the Data Mining Predictive Model has been generated from Phase 1, Phase 2 of the methodology (Fig. 6.5) introduces a technique to classify irrelevant product features based on the evolution of their importance to future product launches. A challenge in traditional engineering decision support models has been the understanding of the relationship between product features with low model relevance and the effects on product family design decisions. Phase 2 in Fig. 6.5 overcomes these challenges by utilizing the time history entropy values (calculated and stored at each iteration in Phase 1) to determine the best course of action for irrelevant product features. Product feature irrelevance is defined simply as product features that do not show up in the resulting predictive model in Fig. 6.5. These product features are classified as either a Standard Feature, Nonstandard Feature, or an Obsolete Feature, with the pseudocode for the algorithm provide below (Tucker and Kim 2011b).

Fig. 6.5

Phase 2: Model generation and irrelevant feature classification

Start: Iteration j = 1

• 1. If predicted Gain Ratio of Feature F _i is not the highest, Feature F _i is considered irrelevant

• 2. Employ Mann–Kendall (MK) trend test for Feature F _i

  • a. If MK τ is negative (with p-value < alpha), irrelevant classification = Standard

  • b. Else If MK τ is positive (with p-value < alpha), irrelevant classification = Obsolete

  • c. Else If MK τ is positive/negative (with p-value > alpha), irrelevant classification = Nonstandard

• 3. While data set/subset does not contain a homogeneous class

  • a. Split the data set into subsets based on the number of mutually exclusive values of the feature with the highest Gain Ratio from Step 2

  • b. j = j + 1 and revert to Step 2 for each data subset

• 4. End Tree, Classify Irrelevant Feature F _i based on highest variable value (SF _t=1,…,n ; NF _t=1,…,n ; OF _t=1,…,n )

In order to classify product features, the emerging predictive power of each product feature must be quantified over time (i.e., each product feature’s relevance to the class variable over time as seen in Fig. 6.5). This is achieved by employing the nonparametric Mann–Kendall trend test, mathematically represented as (Kendall and Gibbons 1990):

$$ \tau =\frac{S}{{\frac{1}{2}n(n-1)}} $$

(6.5)

where

$$ S=\sum\nolimits_{i=1}^{n-1 } {\sum\nolimits_{j=j+1}^n {sgn({x_j}-{x_i})} } $$

(6.6)

• n: represents the total number of time series data points

• x_j: represents the data point one time step ahead

• x_i: represents the current data point

  $$ sgn=\left\{ {\begin{array}{llll} {1if({x_j}-{x_i})>0} \\{0if({x_j}-{x_i})=0} \\{-1if({x_j}-{x_i})<0} \\\end{array}} \right. $$

  (6.7)

The Mann–Kendall begins with a null hypothesis of no trend and rejects or does not reject the null hypothesis based on the resulting p-value and level of significance (α).

The product feature classification framework relies on the results of the Mann–Kendal trend test to quantify the magnitude of the relationship between a given product feature and the output (class) variable. The 3 product feature classification categories are provided below with an application example in Fig. 6.6, illustrating the how the product feature classification could be used to guide enterprise level product family design decisions.

Fig. 6.6

Examples of product feature classification in consumer electronics

3.1.4.2.1 Standard Feature (SF)

A feature F _s is defined as standard if it does not show up in the final decision tree model (as seen in Fig. 6.5) and subsequent tests of the times series Entropy statistics using data from t ₁,…t _n (acquired at each iteration of the model generation process) reveal a monotonically decreasing trend. The Mann–Kendall trend detection test is used as the statistical measure to detect trends. If a monotonically decreasing trend is detected by the Mann–Kendall trend test, this means that despite Feature F _s’s absence from the decision tree model in Fig. 6.5, it is consistently gaining relevance over time and should therefore be considered as a candidate to be included in the product platform decision in Level 2 of the methodology (Engineering Design Optimization level). The Mann–Kendall would return a negative τ and a p-value below the significance level (α).

A binary variable (SF) is defined for the Standard Feature classification that represents the results of the Mann–Kendall trend test at each iteration j. That is, if the Mann–Kendall trend test determines that product feature F _s has a monotonically decreasing entropy trend at iteration j, the binary variable (SF) assumes a value of 1, otherwise 0. Each iteration of the Standard Feature classification SF _j is weighted based on the number of supporting instances in the data set (T _j/T)

$$ SF(t=1,\ldots,n)=\sum\nolimits_{j=1}^j {\left. {S{F_j}\cdot \frac{{{T_j}}}{T}} \right)} $$

(6.8)

A product feature with a Standard classification could be considered for the product platform integration during the product family design process. The engineering components providing the functionality for this product feature could be shared across multiple products within the product family. For example, Fig. 6.6 shows 2 product features that were integrated into the 1st-generation Xbox platform: Internal Hard Drive and a DVD player. Other video game manufacturers such as Sega opted not to include DVD player functionality as a standard product feature in their video game platform (Sega Dreamcast) which contributed to the failure of the system, and ultimately the company as a whole (Aoyama and Izushi 2003). Understanding when to make product features standard (part of the product platform) or nonstandard (modular design that can be replaced/removed) is extremely critical to market success as will be seen in the following classification definitions.

3.1.4.2.2 Nonstandard Feature (NF)

A feature F _n is defined as Nonstandard if it does not show up in the final decision tree model (as seen in Fig. 6.5) and subsequent tests of the times series Entropy statistics using data from t ₁,…t _n (acquired at each iteration of the model generation process) reveal no discernible trend pattern. The Mann–Kendall trend detection test is used as the statistical measure to detect trends. If no discernible trend is detected by the Mann–Kendall trend test, this means that despite Feature F _n’s absence from the decision tree model in Fig. 6.5, Feature F _n exhibits inconsistent relevance patterns through time and should therefore be investigated during the detailed Engineering Design process (Level 2 of the methodology). The Mann–Kendall trend test would return a p-value above the significance level (α) (which would mean that we do not reject the null hypothesis of no trend). A binary variable (NF) is defined for the Nonstandard Feature classification that represents the results of the Mann–Kendall trend test at each iteration j. That is, if the Mann–Kendall trend test determines that product feature F_i has no discernible entropy trend at iteration j, the binary variable (NF) assumes a value of 1, otherwise 0. Each iteration of the Nonstandard Feature classification NF _j is weighted based on the number of supporting instances in the data set (T _j/T).

$$ NF(t=1,\ldots,n)=\sum\nolimits_{j=1}^j {\left. {N{F_j}\cdot \frac{{{T_j}}}{T}} \right)} $$

(6.9)

As opposed to having product variants share the same component addressing a given Nonstandard product feature, designers should avoid component sharing decisions within a product family, and instead develop unique components for each product variant in the product family. The engineering components providing the functionality for this product feature could therefore subsequently be replaced, upgraded, or removed altogether if the product feature eventually becomes obsolete in the market space. Figure 6.6 shows 2 product features of the 2nd-generation Xbox (Xbox 360) that had a modular design that was not shared between product variants within a product family: HD-DVD player and Removable External Hard Drive. During the video game console wars in the mid-2000s, two competing media formats were in direct competition with one another: the Blu-ray and HD-DVD (Brookey 2007). With the uncertainty of a clear winner in the market space, Microsoft opted for a modular add-on HD-DVD device (see Fig. 6.6) that could seamlessly integrate with the Xbox 360 product variants in the market space if High Definition media consumption was desired by consumers. The add-on HD-DVD was discontinued soon after it was clear that Sony’s Blu-ray format had won the next-generation media platform wars (Daidj et al. 2010), making the HD-DVD modular device for the Xbox 360, Obsolete. If Microsoft had made the decision early on in the Xbox 360 product design process to integrate the HD-DVD player into the product platform, shared across multiple product variants, an entire redesign of the Xbox 360 system may have resulted after the HD-DVD product feature failed in the market space. The Removable External Hard Drive also allowed different variants of the Xbox 360 to have different storage capacities (20 GB, 60 GB, 120 GB, etc.), allowing greater customization in the market space, while keeping the core Xbox 360 relatively unchanged (Microsoft Inc. 2007; Farkas 2009).

3.1.4.2.3 Obsolete Feature (OF)

A feature F _o is defined as obsolete if it does not show up in the final decision tree model (as seen in Fig. 6.5) and subsequent tests of the times series Entropy statistics using data from t ₁,…t _n (acquired at each iteration of the model generation process) reveal a monotonically increasing trend. The Mann–Kendall trend detection test is used as the statistical measure to detect trends. If a monotonically increasing trend is revealed by the Mann–Kendall trend test, this means that despite Feature F_o’s absence from the decision tree model in Fig. 6.5, it is consistently losing relevance over time and should therefore be investigated during the detailed Engineering Design process in Step 3. The Mann–Kendall trend test would return a positive τ and a p-value below the significance level (α). A binary variable (OF) is defined for the Obsolete Feature classification that represents the value results of the Mann–Kendall trend test at each iteration j. That is, if the Mann–Kendall trend test determines that product feature F _o has a monotonically increasing entropy trend at iteration j, the binary variable (OF) assumes a value of 1, otherwise 0. Each iteration of the Obsolete Feature classification OF _j is weighted based on the number of supporting instances in the data set (T _j/T).

$$ OF(t=1,\ldots,n)=\sum\nolimits_{j=1}^j {\left. {O{F_j}\cdot \frac{{{T_j}}}{T}} \right)} $$

(6.10)

The final classification of a feature (that does not show up in the decision tree model in Phase 2 of Fig. 6.5) is achieved by summing across all iterations of each of the feature classification variables (Standard, Nonstandard, and Obsolete) and selecting the variable with the highest value.

A product feature with an Obsolete classification indicates that it has little market-driven significance over time. The engineering components providing the functionality for this product feature can be considered candidate components to be removed in next-generation product designs. Figure 6.6 shows 1 product feature that was initially part of the 1st-generation Xbox platform, characterized as Obsolete in the 2nd-generation Xbox platform. The 2nd-generation Xbox platform (Xbox 360) launched in 2005 without an internal hard drive, making this product feature obsolete to the Xbox 360 platform. Microsoft opted for modular hard drives (External HD in Fig. 6.6) that could be replaced or upgraded at a certain price (Farkas 2009). This enabled Microsoft to discontinue an obsolete component (the internal hard drive), while creating a product family of Xbox 360 s that served different consumer market segments based on the technical capabilities of the Xbox 360 platform [platforms came with no External Hard Drive, 20 GB External Hard Drive and 100 GB Hard Drive (Farkas 2009)].

The methodology proposed in this chapter aims to address the link between the evolution of product feature relevance and the implications to product platform and product family design. Specifically, this work aims to:

• Translate a product feature classification from the market-driven domain to the detailed engineering domain.

• Determine the optimal product platform sharing decisions based on the market-driven evolution of product features and customer preferences.

3.2 Level 2: Engineering Design Optimization

3.2.1 Mapping Product Feature Space to Engineering Design Space

From the resulting Data Mining Predictive Model from Level 1 of the methodology, product features will either:

• 1. Be part of the predictive model and therefore considered relevant product features.

• 2. Be omitted from the predictive model and be classified as:

  • (a) Standard

  • (b) Nonstandard

  • (c) Obsolete

Level 2 of the product feature classification framework plays a vital role in mapping market-driven, product feature preference trends to engineering design specifications. The ability of a product family to address market-driven demand is highly dependent on the evolution of product feature preferences over time. While component commonality decisions in product family design aim to provide the optimal configuration of product platforms within a product portfolio, mathematical models often omit evolving product feature preferences in the market space, hereby increasing the risk of product failure when launched. Unlike individually designed products, the market failure of a product family (e.g., due to an unwanted product feature) could result in the redesign of an entire product family (as opposed to just a single product) due to the components shared between product variants. In many real life scenarios, the characteristics of the product feature space (as defined by the customer) significantly differs from the technical characteristics of the engineering product design space. For example, a large-scale data set containing product preference data may contain a product feature such as 8 h battery life. For the same product feature however, the technical components used to achieve such functionality of a product’s feature would be described by more technical terms such as 60 Whr 6-Cell Lithium-Ion Battery (Dell Inc. 2012).

The aim is to first map the Standard, Nonstandard, and Obsolete product feature classifications to the Functions of a component(s) in a product family, as shown in Fig. 6.7. The Standard classification includes all features included in the Data Mining Model, in addition to the SF irrelevant feature classifications. The Nonstandard classification includes all features classified as NF, and the Obsolete classification includes all features classified as OF. It is important to note that a feature can have one and only one classification.

Fig. 6.7

Component function identification

The textual descriptions of each product feature will then be compared with the textual description of all functions in a product family to determine which components functions are providing the market-driven product preferences. Each component function F _i existing in a product family is assumed to be defined by a synonym set {s₁,s₂,…,s_s} that describes its technical purpose to the product/product family as a whole. A product feature-function matrix is then created as Table 6.1.

Table 6.1 Product feature-function comparison matrix

Latent Semantic Analysis is employed to make semantic comparison between the vector of terms characterizing a product feature and those of a product function. LSA not only compares the original vector of textual terms but also their semantic meaning and makes the assumption that terms with similar meanings will occur close to each other. Therefore, despite the fact that a product feature term “charge” and the product function “battery” are not identically the same, LSA may quantify the related meaning between the two.

Table 6.1 can be represented by one of two vectors

• Semantic term vector (each row of Table 6.1):

  $$ t_i^T=[{c_{i,1 }}\ldots {c_{i,n }}] $$

  (6.11)

• Product feature-function Comparison (each column of Table 6.1):

  $$ {f_j}=\left[ {\begin{array}{llll} {{C_{1,j }}} \\. \\. \\. \\{{C_{m,j }}} \\\end{array}} \right] $$

  (6.12)

Table 6.1 can be defined as X, where c _i,j represents the frequency/occurrence of a particular term in the description of either a product feature or product function.

$$ X=\left[ {\begin{array}{llll} {{c_{1,1 }}} & \cdots & {{c_{1,n }}} \\\vdots & \ddots & \vdots \\{{c_{m,1 }}} & \cdots & {{c_{m,n }}} \\\end{array}} \right] $$

(6.13)

The singular value decomposition (SVD) of X can therefore be represented as (Deerwester et al. 1990):

$$ X={T_0}{S_0}{{D^{\prime}}_0} $$

(6.14)

where

• X: is the term (t) by function/feature (f) matrix (i.e., X = t × f)

• T ₀: represents the term (t) by rank (m) matrix, having orthogonal, unit-length columns (T ₀ ’T ₀  = I)

• S ₀: is the diagonal matrix of singular values (m × m)

• D ₀: is the rank (m) by function (f) matrix, having orthogonal, unit-length columns (D ₀’D ₀ = I) (i.e., D ₀ = m × f)

• m: is the rank of X $(\leq\min (t,d))$

LSA therefore provides lower-dimension estimates of the original high-dimension space which then enables a comparison of the semantic meaning (beyond just simple term matching) between a product feature and a product’s function using similarity metrics such as cosine similarity (Tucker and Kang 2012).

Once the market-driven product features have been mapped to specific product functions using SVD, a detailed function-component analysis must be performed. A component-function matrix representation is used to quantify the relationships/interactions between components. A majority of the literature reviewed in this work typically focus on component sharing optimization within and between a family of products. By employing the market-driven product feature classification methodology presented in the previous section, engineering design decisions relating to product family optimization can be guided by emerging product preferences in the market space. The DSM has been employed extensively in the design community to represent interactions among products/processes in a design process (Browning 2001). Examples of interactions captured by the DSM framework include Spatial (associations relating to the physical location of elements), Energy (energy transfer/exchange between elements), Information (data/signal exchanges between elements), and Material (material exchange between elements) (Pimmler and Eppinger 1994).

The first step is to determine which functions of a product relate to specific component(s). There may be some components that perform more than one specific function, even across product variants as can be seen in Fig. 6.8. For example, unlike the 60 Whr 6-Cell Lithium-Ion Battery that supplies electrical energy to a product as its function, other components such as a DVD super drive component provides multiple functions such as reading media, recording media, and erasing media. Each of these would be considered a unique function of this component. Atomicity is a desired property of the functional decomposition process. It is assumed that designers have a database of components with their individual functions; therefore, the component-function relationship in Table 6.2 will ensure that design teams understand the interactions among components within a product variant and across product variants existing in a product family.

Fig. 6.8

Function-component analysis in a product family

Table 6.2 Component-function interaction matrix

where

• 1: represents a component critical to achieving a specific function

• 0: represents a component that is complimentary to achieving a specific function

The product family optimization problem can be solved using a quasi-separable, bi-level optimization model, where the coordination level handles the shared variables (components common to the system) and the product platform level handles the individual product variant optimization problems (each with local objective functions such as cost minimization) (Kim et al. 2003; Tosserams et al. 2006). three binary feature classification variables are included in the objective function to help guide the optimization problem. The Standard Feature (SF), Nonstandard Feature (NF), and Obsolete Feature (OF) are modeled as follows.

Optimization Level 1: Product Family Sharing Level

Minimize

$$ {{\boldsymbol{\varepsilon}}_y} $$

(6.15)

Subject to:

$$ g1:S{F_p}\cdot \sum\nolimits_{{k\in Q}} {\left\| {{\mathbf{y}_p}-\mathbf{y}_{p,k}^{Eng }} \right\|}_2^2-{\mathbf{\upvarepsilon}_y}\leq 0 $$

(6.16)

Here,

• p: product feature from the market-driven Data Mining Predictive Model.

• SF _p : binary variable for Standard Feature classification (1 if the product feature is deemed relevant by the Data Mining Predictive Model and 0 otherwise).

• y _p : linking variable at the product family sharing level that maintains consistency between k product variant values. The component or variable y _p corresponds to function-component map for the specific product feature p.

• y _p,k ^Eng: the shared variable/component value associated with product variant k providing product feature p. This is constant at each iteration in the above formulation that is subsequently updated at the product variant optimization level after each iteration.

• k: the kth candidate product variant that has been identified for component sharing.

• Q: the total number of products attempting to share design variables/components y _p,k ^Eng.

• ε _y : deviation tolerance between linking variables that is minimized in the objective function.

Optimization Level 2: Product Variant Optimization Level

Minimize

$$ F{(x)_{{Varian{t_k}}}}={f_k}+\left\| {y_p^U-{y_{p,k }}} \right\| $$

(6.17)

Subject to:

$$ {{\mathbf{g}}_k}({{\mathbf{x}}_k}_{,}\ {{\mathbf{y}}_{p,k }})\ \leq \mathbf{0} $$

(6.18)

$$ {{\mathbf{h}}_k}({{\mathbf{x}}_k}_{,}\ {{\mathbf{y}}_{p,k }}) = \mathbf{0} $$

(6.19)

Here,

• p: product feature from the market-driven Data Mining Predictive Model.

• NS _p : binary variable for Nonstandard Feature classification (1 if the product feature is classified as Nonstandard Feature by the Data Mining Predictive Model and 0 otherwise).

• OF _p : binary variable for Obsolete Feature classification (1 if the product feature is classified as an Obsolete Feature by the Data Mining Predictive Model and 0 otherwise).

• f _k : local product design objective function (s).

• g _k : inequality design constraints.

• h _k : equality design constraints.

• x _k : design variables local to product variant k. x _k is a function of the product feature variables NS_p and OF_p. NS_p and OF_p are presented here in a general form as the mathematical formulation and inclusion in the optimization model will be highly dependent on the structure of the product family model.

• y _p ^U: linking variable target value cascaded down to the Level 2 from Level 1; a constant value at each iteration that is subsequently updated with each successful iteration.

• y _p,k : linking variable at Level 2 that attempts to match the target linking variable value y _p ^U used to achieve the product feature p.

• k: the kth candidate product variant that has been identified for component sharing.

For each unique product feature that is included in the product family optimization model has to satisfy the equality constraint:H1: SF_p + NS_p + OF_p = 1, indicating a single state during the product family optimization model (variable/component sharing, module, or exclusion from the optimization model).

4 Case Study of a Family of Aerodynamic Particle Separators

Particulate Matter (PM)/particle pollution is a complex mixture of very small particles such as acids, organic chemicals, metals, soil, or dust particles (US EPA 2012). Severe health problems can be caused to the heart, lungs, and other organs when PM sizes are 10 μm in diameter or smaller.

Aerodynamic particle separators are devices developed to separate Particulate Matter from the clean air stream, typically by employing centrifugal forces on the particles (Zhang 2005). The case study presented in this methodology is based on an aerodynamic particle separator market segment including applications such as agriculture, industrial, and manufacturing processes. The global market for air cleaning technologies has exceeded $7 Billion and continues to rise (Parker 2006). The diverse operating conditions and preferences of customers make product standardization a challenge. Customized solutions for aerodynamic particle separators are typically used to solve the wide range of market segments (Fig. 6.9).

Fig. 6.9

Aerodynamic particle separator market segments [adapted from (Barker 2008)]

This case study aims to investigate the feasibility of employing the proposed Data Mining-Driven product family design methodology to help:

• Quantify the evolution of product feature characteristics over time.

• Develop a Data Mining predictive model of relevant product features for future product family designs.

• Classify product features as Standard, Nonstandard, or Obsolete.

• Investigate how product feature classification influences product family sharing and optimization decisions.

For more details regarding this case study, please see references Barker (2008) and Tucker et al. (2010).

4.1 Level 1: Temporal Market-Driven Preferences

Table 6.3 above presents a snapshot of the structure of the data set for the aerodynamic particle separator for one instant in time (t _i), where:

Table 6.3 Sample data set for aerodynamic particle separator

• Q: air flow rate (m³/s)

• ΔP _max : maximum allowable change in pressure drop/airflow restriction (Pa)

• L _max : total allowable length of the system (m)

• AF _max : maximum allowable face area perpendicular to air flow direction (m²)

• N_max: maximum number of aerodynamic particle separator units in one module (#)

• F(d _p ): particle size distribution (%)

• ρ _p : particle density (kg/m³)

• T_air: air temperature (^°C)

• P _air : air pressure (kPa)

The data set contains nine features relating to the aerodynamic particle separator (five of which are related to the physical design of the system while the remaining four are related to the environmental conditions that the system will perform under). The product features will help guide the product family optimization process by suggesting candidate components for sharing or displacement. The Environmental Features will serve as the design constraints of the model. The class variable here is Efficiency which is defined as the total amount of particulate matter that a system is able to separate from clean air.

The data in Table 6.3 is mined for emerging product feature trends in the market space by quantifying the relevance of product features over time. A Data Mining predictive model is generated that helps guide Level 2 of the product family design methodology: Engineering Design Optimization.

4.2 Level 2: Engineering Design Optimization

The aerodynamic particle separator has a fan system downstream (attached to the radius r3 in Fig. 6.10) and operates by pulling contaminated air from upstream into the system. The contaminated air (composed of clean air and particulate matter) enters the vane section (Fig. 6.10) causing the particles to rotate based on the angle of the vane section (Barker 2008). The straight section in Fig. 6.10 is designed to increase the separation (centripetal acceleration and inertia) between the particulate matter and the clean air. The clean air particles enter the converging region of the particle separator, leaving the particulate matter to collect in the storage bunker in Fig. 6.10.

Fig. 6.10

Uniflow aerodynamic particle separator design [Augmented from Tucker et al. (2010)]

Figure 6.10 presents 2 approaches to the product family design, Scale based and Module based. Scale-based product family is where a product platform is “stretched” or “shrunk” in one or more dimensions in order to satisfy a market need (Simpson et al. 2006). Design/scaling variables are formulated in the product family optimization model to achieve such scalability. Module-based product family design on the other hand creates product variants by adding, replacing, or substituting one or more functional modules from a product platform. A product architecture is considered modular if there is a clearly defined mapping of functional elements to physical structures (1-1 or many-1) (Simpson et al. 2006).

The design objectives of each aerodynamic product variant will be influenced by the market-driven Data Mining model. For scale-based product family design, the efficiency of the system can be influenced by altering (stretching or shrinking) the design variables that make up the physical system (such as length of straight region, inner and outer radii, etc.)

The optimization approach (module vs. scale based) is left up to the design team based on the technical resources and available mathematical models.

Optimization Level 1: Product Family Sharing Level

The Product Family Sharing Level will coordinate the component sharing among product variants of the aerodynamic particle separator by minimizing the tolerance deviation variable of each shared component.

Minimize

$$ {\mathbf{\upvarepsilon}_y} $$

(6.20)

Subject to:

$$ g1:S{F_p}\cdot \sum\nolimits_{{k\in Q}} {\left\| {{\mathbf{y}_p}-\mathbf{y}_{p,k}^{Eng }} \right\|_2^2-{\mathbf{\upvarepsilon}_y}\leq 0} $$

(6.21)

Here,

• p: product feature from the market-driven Data Mining Predictive Model.

• SF _p : binary variable for Standard Feature classification (1 if the product feature is deemed relevant by the Data Mining Predictive Model and 0 otherwise).

• y _p : linking variable at the product family sharing level that maintains consistency between k product variant values. The component or variable y _p corresponds to function-component map for the specific product feature p.

• y _p,k ^Eng: the shared variable/component value associated with product variant k providing product feature p.

• k: the kth candidate aerodynamic particle separator product variant

• Q: The total number of products attempting to share design variables/components y _p,k ^Eng.

• ε _y : deviation tolerance between linking variables that is minimized in the objective function.

Optimization Level 2: Aerodynamic Particle Separator Variants

The engineering design model for the aerodynamic particle separator can be mathematically represented as:

kth Aerodynamic Particle Separator

Minimize:

$$ F{({\boldsymbol x})_{variant(k) }}=Cos{t_k}-{\xi_k}+\left\| {y_p^U-{y_{p,k }}} \right\| $$

(6.22)

where

$$ \zeta \left( {x,{d_{pi }}} \right)=1-\exp \left( {\frac{{{\rho_p}d_{pi}^2{C_c}Qtan(\alpha ){L_s}}}{{9\eta (r_2^2-r_1^2)}}} \right)\cdot \exp \left( {\frac{{{\rho_p}d_{pi}^2{C_c}(V_t^2{G_t}(x)+V_z^2{G_r}(x)}}{{\eta {V_z}}}} \right) $$

(6.23)

$$ {\xi_k}=\sum\nolimits_{i=1}^N {\zeta ({\boldsymbol x},{d_{pi }})\cdot F({d_{pi }})} $$

(6.24)

Here,

• ${\xi_k}$: efficiency of aerodynamic particle separator variant k

• y _p ^U: linking variable target value cascaded down to the Level 2 from Level 1; a constant value at each iteration that is subsequently updated with each successful iteration

• y _p,k : linking variable at Level 2 that attempts to match the target linking variable value y _p ^U used to achieve the product feature p

• k: the kth candidate product variant that has been identified for component sharing

• C _c : Cunningham slip correction factor

• d ${_{{{p_i}}}}$: diameter of particle i), μm

• F(d _p ): particle size distribution

• G _t (x): efficiency model geometric relationship between design variables, tangential acceleration

• G _r (x): efficiency model geometric relationship between design variables, radial acceleration

• ρ _p : particle density, kg/m³

• η: air viscosity, Pa∙s or kg∙m/s

• Q: air flow rate, m³/s

• V _t : tangential velocity of particle mixture

• V _z : axial velocity of particle mixture

• r ₁ : inner tube radius

• r ₂ : inner tube radius

• α: vane discharge angle

• L _S : maximum pressure drop

Subject to:

$$ {{\mathbf{g}}_k}({{\mathbf{x}}_k}_{,}\ {{\mathbf{y}}_{p,k }})\ \leq \mathbf{0} $$

(6.25)

$$ {{\mathbf{h}}_k}({{\mathbf{x}}_k}_{,}\ {{\mathbf{y}}_{p,k }}) = \mathbf{0} $$

(6.26)

5 Results and Discussion

5.1 Level 1: Temporal Market-Driven Preferences

Figure 6.11 presents the Data Mining Predictive Model based on the temporal market-driven preferences relating to the aerodynamic particle separator. The results in Fig. 6.11 can be interpreted by traversing down each individual branch in the tree until a class variable (Efficiency) is reached (rectangular box). The ovals in Fig. 6.11 represent the product feature that is deemed relevant to predicting the market preferences for aerodynamic particle separator efficiency.

Fig. 6.11

Results from the data mining predictive model [attained using Weka 3.6.6 (Frank et al. 2010)]

Four unique paths can be attained based on the results in Fig. 6.11:

• Lmax > 0.49 then efficiency > 85–90 %

• Lmax < 0.49 and Q > 3 and Delta_pmax > 1560 then efficiency > 95 %

• Lmax < 0.49 and Q > 3 and Delta_pmax < 1560 then efficiency > 90–95 %

• Lmax < 0.49 and Q < =3 then efficiency > 85–90 %

Figure 6.11 also provides designers with the appropriate product feature classification (Standard, Nonstandard, and Obsolete) of all product features existing in the market space. The next step is to quantify the relationship between product features and product function so that designers can understand how evolving market-driven preferences guide next-generation product platform and product family design decisions.

5.2 Level 2: Engineering Design Optimization

Mapping Product Feature Space to Engineering Design Space: Table 6.4 presents the result, employing Latent Semantic Analysis to quantify the relationship between the market-driven product feature space and the product family design space. As described in Sect. 6.2.2, the textual description of each product feature is compared with the functional description of each product module/component providing this function. Table 6.4 represents the similarity of a product feature to a component function as measured on a 0–1 scale, where values closer to 1 indicate a stronger relationship to the functionality of the product component/module, while 0 represents a weaker relationship. As can be seen from Table 6.4, the product features are strongly coupled across the entire product architecture. Such insight will help designers understand the market effects of adding, removing, or replacing specific functionality relating to a product. For an Obsolete product feature classification, designers would need to ensure that the removal of a particular component/module does not have negative market demand implications. The results in Table 6.4 are consistent with the absence of an Obsolete product feature classification from the Data Mining model in Fig. 6.11. The Standard and Nonstandard product classifications from Fig. 6.11 support the findings from Table 6.4, indicating that all product features are relevant to market success at this time. The challenge arises when designers are trying to optimize product family sharing and platforming decisions, which will now be guided by the product feature preference trends in the market space.

Table 6.4 Mapping product features of the aerodynamic particle separator to the engineering design space

Figure 6.12 presents the results from the Data Mining-Driven Product Design methodology. The resulting product feature classifications from Fig. 6.11 helps guide the product family design process by first quantifying the functional relationships between product features and product design variables (Table 6.4) and then suggesting candidate modules/components for sharing decisions. In Fig. 6.12, the vane component is considered a candidate for commonality across product variants due to its relation to the product features deemed relevant by the Data Mining Predictive model in Fig. 6.11. The Nonstandard product feature suggests modularity in the design of product variants by including multiple individual products that are housed in a design case. This modularity approach will enable enterprise decision makers to quickly address market needs by increasing/decreasing the number of units housed in a casing in an attempt to meet emerging customer preferences. The Nonstandard product feature classification means that there is volatility in the market space, wherein a product may have to undergo modifications (modular or scalable) in the future. Figure 6.12 reveals that of the 4 product segments suggested by the market space, the designers can only satisfy the performance requirements of 3 of those markets, as Product Variant 2’s design (Target efficiency > 95 %) is infeasible at the engineering design level. Such design insights enable design teams to focus on developing a product portfolio that both capitalizes on product standardization through guided commonality decisions, and at the same time, providing customized solutions to the market that meet customer expectations (efficiency requirements).

Fig. 6.12

Results from data mining-driven product family design

6 Conclusions

This chapter introduces a market-driven, product family design framework based on product feature classification framework as it relates to engineering component selection and product family design. Product features are classified as Standard, Nonstandard, or Obsolete, with a given classification having different implications in the product family design process. A component-function matrix is presented to quantify the relationships and interactions among components as it relates to product specific features. By employing Natural Language Processing techniques, the product feature space can be mapped to the engineering design space for optimal product platform decisions that incorporate market-driven objective. The methodology aims to aid design teams in the efficient modeling of customer preferences and designing of subsequent product portfolios.