Data Science Applied to the Reduction of Part Diversity in Industrial Portfolios

Eduardo Zonatto; Jéssica Eloá Poletto

DOI: 10.22167/2675-6528-202602850

Article derived from a Course Conclusion Work (TCC), with content based on the student’s original work and adapted to the editorial format of the E&S Magazine with the support of the ResumeAI tool, an artificial intelligence solution developed by Instituto Pecege for textual synthesis and organization.

Summary

The excessive diversity of components in industrial portfolios generates operational costs and inefficiencies throughout the entire production chain. An integrated data science pipeline was applied to characterize component usage patterns in a real industrial portfolio and to identify, quantitatively and reproducibly, opportunities for standardization and rationalization. The research was characterized as quantitative, with secondary data, using descriptive statistical analysis, clustering, and network modeling. 4,580 items directly linked to the product portfolio were analyzed, on which an analytical pipeline in Python was developed, including data preparation, outlier detection, exploratory analysis, dimensionality reduction, K-means clustering, weighted bipartite network modeling with centrality metrics, and synthesis by the Network Criticality Index with module detection. The results segmented the portfolio into four profiles: ultrastructural components (11.7%), which require individualized analysis; platform components (7.2%), candidates for consolidation in new projects; intermediate use components (15.8%), the main opportunity for immediate standardization; and restricted use components (65.3%), candidates for rationalization. Additionally, 6,658 components (42.0%) registered no usage, representing an immediate opportunity for simplification. The detection of seven modules revealed a modular organization that enables standardization initiatives with controlled scope, demonstrating the methodology’s capacity to identify optimization and portfolio management opportunities.

Keywords: Network analysis; Data science applied to engineering; Clustering; Part diversity; Component standardization.

1. Introduction

The excessive diversity of components in industrial portfolios is recognized as a factor generating operational costs and complexity throughout the entire product lifecycle. Each new component introduced into a portfolio incurs charges that go beyond the cost of the item itself, encompassing supplier qualification, dedicated inventory maintenance, inspection and traceability processes, and the cost of catalog retention. In high-variety manufacturing environments, such as that of commercial vehicles with multiple configurations, this effect is amplified by the overlap of functionally equivalent components that accumulate variants over product generations without active variety management (Fisher and Ittner, 1999; Ulrich and Eppinger, 2020).

The standardization of components and the deliberate sharing of parts among products are consolidated strategies to contain this growth and mitigate inefficiencies. Product engineering literature shows that modular designs and the use of common platforms reduce development costs, increase reuse, and simplify the supply chain (Jose and Tollenaere, 2005; Liu et al., 2010). The measurement of the degree of commonality between products, which represents the fraction of shared components relative to the total portfolio, is a critical indicator for product architecture decisions, with several quantitative indices proposed to capture its multiple dimensions (Thevenot and Simpson, 2006). Product architecture decisions, especially the choice between independent modules and shared components, determine a large part of the development, production, and service costs throughout the lifecycle (Ulrich and Eppinger, 2020). The reuse of components in mechanical designs is not just an engineering convenience, but a system decision with a direct impact on reliability, interchangeability, and maintenance cost (Budynas and Nisbett, 2019; Norton, 2020).

Despite the relevance of the topic, the identification of standardization opportunities is still often conducted empirically, supported by the experience of engineers and periodic catalog reviews. This process is limited by the scale of real industrial portfolios, which contain tens of thousands of components distributed across hundreds of products and assembly structures, making systematic manual analysis unfeasible. Although corporate data recording each structural link between components and products exists in these environments, it is rarely explored as a basis for quantitative portfolio management decisions.

The application of data science techniques to this problem allows transforming engineering operational records into structured and actionable knowledge. Clustering methods, such as K-means, identify components with similar usage profiles. Principal Component Analysis (PCA) reduces the dimensionality of the problem without loss of relevant information. Complex network modeling, in turn, captures structural properties of the portfolio that conventional tabular analyses do not reach (Newman, 2010; Barabási, 2016). In particular, the bipartite representation of component-product relationships as a weighted network allows calculating centrality metrics, such as degree, strength, Betweenness, Eigenvector, and PageRank, which quantify the systemic role of each component in the portfolio architecture (Newman, 2010; Barabási, 2016).

Recent approaches have consolidated the use of graphs and network theory metrics as support tools for modular design and the management of interdependencies in complex engineering systems, demonstrating that structural network properties, such as degree distribution, allow for the identification of high-impact components before any portfolio intervention (Sosa et al., 2007). Components with high betweenness centrality, for example, function as bridges between different portfolio segments and demand differentiated treatment in rationalization initiatives. The combination of these approaches produces a richer portfolio segmentation than any isolated technique: clustering provides interpretable groups for prioritization, network analysis reveals the systemic criticality of each component, and the integration of both, through a structural criticality index, allows for portfolio ordering by risk and opportunity simultaneously, differentiating itself from approaches that use only functional and geometric attributes (Jiao et al., 2007).

Given the above, this work is justified by the need for a quantitative and reproducible approach to managing the diversity of components in industrial portfolios, overcoming the limitations of traditional empirical methods. The objective of this work was to apply an integrated data science pipeline, combining exploratory analysis, clustering, and network modeling, to characterize the component usage patterns in a real industrial portfolio and to identify, quantitatively and reproducibly, opportunities for standardization and rationalization.

2. Material and Methods

This study was characterized as a secondary data survey, quantitative in nature. The methodological approach included descriptive statistical analysis, clustering techniques, and network modeling, according to the guidelines of Gil (2019) and Prodanov and Freitas (2013). The objective was to apply an integrated data science pipeline to characterize component usage patterns and identify opportunities for standardization and rationalization in a real industrial portfolio.

The research was conducted at a commercial vehicle manufacturer, using three corporate databases extracted from the company’s engineering management system. The databases were exported in tabular format and integrated using the unique component identifier (Partno), ensuring data traceability and consistency. The data collection period was not explicitly informed in the original TCC, but the data represents a static snapshot of the portfolio.

The first database, called the component base, contained the general registry of portfolio items, including attributes of development stage and validity status. The second database, the link base, recorded the direct structural links between components and products, using unique link identifiers (LinkID) and the adjusted quantity of each occurrence (adjusted_qty). The third database, the assembly base, described the links between components and intermediate subassemblies, being used to classify the type of component usage.

From the intersection of these three bases, 15,847 valid components were classified, restricted to the industrialization stage (Stage = P) and active status (Status = VALID). Of these, 4,580 components had direct links to products, forming Network A, which constituted the main study universe. The assembly base (Network B) was used only for classification purposes, not being subjected to network modeling.

The data treatment and analysis were structured in a sequential seven-step pipeline, implemented in Python 3. The libraries used included pandas and numpy for data manipulation, scikit-learn for standardization, Principal Component Analysis (PCA) and K-means, networkx for network construction and analysis, and python-louvain for community detection. The results were exported in CSV format for visualization in Power BI Desktop.

In the data preparation stage, the three databases were loaded, filtered, and integrated. The links database, which initially contained 692,465 raw records, was processed with the removal of null records and deduplication by pair (Partno, Product_ID), resulting in 18,818 unique component-product relationships. Each row represented a unique LinkID, indicating the structural recurrence of a component in a specific assembly.

For each component, three characterization variables were calculated: num_products, which expresses the number of distinct products in which the component is used; num_links, which counts the total of LinkIDs associated with the component across all products; and total_qty, which corresponds to the sum of the adjusted_qty values, representing the total physical volume consumed in the portfolio.

Before applying clustering techniques, the detection and isolation of components with extreme values, called “outliers”, were performed. The criterion adopted was the interquartile range (IQR) method, defined as the difference between the third (Q3) and the first (Q1) quartile of the variable’s distribution. A conservative factor k = 3.0 was applied independently to each of the three characterization variables.

A component was classified as an “outlier” when it exceeded the upper limit, calculated as Q3 + k × IQR. This procedure identified 538 components that were isolated from the analytical base and characterized separately, to avoid distortions in the K-means centroids. The cleaned base, used in subsequent analyses, comprised 4,042 components.

Exploratory analysis included the elaboration of descriptive statistics of the characterization variables, including distribution by reuse ranges and the calculation of Pearson correlations. For dimensionality reduction, Principal Component Analysis (PCA) was applied to the num_products and num_links variables from the clean database, after standardization by the z-score method. The total_qty variable was excluded due to its high redundancy with num_links.

The clustering was performed by the K-means algorithm on the two-dimensional projection of PCA. The determination of the number of clusters (K) was based on the elbow method (intra-cluster inertia) and the silhouette coefficient. K = 3 was adopted, justified by engineering criteria to distinguish single-use, intermediate, and high reuse components, with a silhouette coefficient of 0.588.

As a validation method, the DBSCAN algorithm was applied to the same PCA space, with parameters eps = 0.5 and min_samples = 5. This validation confirmed the absence of a clear multimodal structure in the data, corroborating the choice of K-means as the main clustering method. The group centroids were ordered by the median value of num_products to ensure consistency between runs.

The relationship between components and products was modeled as a weighted bipartite network. Nodes of type C represented components, and nodes of type P represented products. Each C-P edge was assigned a weight equal to the number of LinkIDs for the corresponding pair, reflecting the intensity of the structural relationship. The resulting network included 4,580 component nodes, 41 product nodes, 18,818 unique edges, and 692,455 total weighted links.

Five centrality metrics were calculated for each network component, capturing distinct dimensions of structural importance: “Degree”, “Strength”, “Betweenness”, “Eigenvector”, and “PageRank”. The betweenness centrality calculation was performed by approximation, with 500 nodes sampled. All metrics were normalized by the Min-Max method in the [0, 1] range to allow comparison and composition.

To synthesize the centrality metrics into a single indicator of structural criticality, the Network Criticality Index (NCI) was defined as the arithmetic mean of the normalized metrics (Degree_norm, Strength_norm, Betweenness_norm, Eigenvector_norm, PageRank_norm). The detection of structural modules was performed by the Louvain algorithm (Blondel et al., 2008), applied directly to the bipartite network to maximize modularity.

To verify the robustness of the module detection, the Leiden algorithm (Traag et al., 2019) was also applied to Network A. The substantial agreement between the results of both algorithms, with no identification of disconnected communities in the Louvain partition, confirmed that the methodological choice did not affect the identification of the structural modules.

In accordance with institutional ethical guidelines, the company name and sensitive data were preserved, ensuring the confidentiality of the information used in the study. Additionally, authorization was obtained from the company’s representative, through a Term of Consent, for the use of information and data.

3. Results and Discussion

The initial portfolio analysis revealed a universe of 15,847 valid components, of which 6,658, corresponding to 42.0% of the total, showed no record of use. This contingent of unused items represents an immediate and direct opportunity for portfolio simplification, as the absence of links is, in itself, a robust criterion for review and potential elimination, regardless of more complex analyses. The management of such inactive components generates unnecessary catalog maintenance costs and inventory potential, impacting the company’s operational efficiency.

Among the components with registered use, 4,580 items, representing 28.9% of the total valid, were directly linked to final products, constituting Network A, the main focus of this study. The link database, initially with 692,465 raw records, was processed by deduplication of component-product pairs, resulting in 18,818 unique relationships. This significant difference between the number of raw records and unique relationships highlights the high structural recurrence of the portfolio, where each component-product pair appears, on average, 36.8 times in the assembly structure. The most intensive record reached 42,774 occurrences for a single pair, indicating that certain components are deeply integrated into the architecture of multiple products.

Descriptive statistics and reuse distribution

The descriptive statistics of the characterization variables, calculated on the complete base and the clean base (after removing ultrastructural components), demonstrated the impact of “outliers” on central tendency estimates. On the complete base, the mean of num_links was 151.2, while on the clean base, after isolating 538 components, this mean dropped to 18.3. This variation highlights the importance of detecting and treating extreme values to ensure the accuracy of subsequent analyses, especially regarding the identification of usage and reuse patterns.

The distribution of the num_products variable in the clean dataset (n = 4,042) revealed a pronounced long-tail structure, characteristic of networks with heterogeneous degree distribution, as described by Barabási (2016). It was observed that 44.6% of the components are used in only one product, and 63.2% in at most two products. Only 4.6% of the components are present in ten or more products. This pattern indicates that the majority of the portfolio is composed of items with specific applications, with low potential for cross-reusability, which directs standardization strategies to specific segments.

Ultrastructural components

The 538 components classified as ultrastructural, representing 11.7% of Network A, were isolated due to their high criticality and potential for distortion in clustering analyses. These items concentrate a disproportionate share of the network’s connectivity. For example, the component with the highest criticality (Partno 980464) was present in 35 of the 41 analyzed products, accounting for 13.9% of all links in the database. This concentration pattern, where a few components support the majority of connections, is a distinctive feature of networks with heterogeneous degree distribution, as pointed out by Barabási (2016).

The identification and isolation of these ultrastructural components are crucial for portfolio management. From an engineering perspective, these items are not candidates for rationalization without in-depth individualized technical assessment, as their removal or replacement could significantly impact the structure of an expressive fraction of the portfolio. Differentiated management of these components is, therefore, a necessary condition for any effective standardization initiative, allowing actions to be strategically directed and minimizing systemic risks.

Principal Component Analysis

Principal Component Analysis (PCA) was applied to the variables num_products and num_links from the cleaned database (4,042 components), after z-score standardization. The total_qty variable was excluded from the analysis due to its high correlation with num_links (r = 0.738, p < 0.17, requiring individualized technical analysis due to the systemic risk of any change.

Together, the components with no registered use and those with restricted use total 9,649 items, representing 60.9% of the total valid portfolio. This contingent constitutes the main opportunity for portfolio simplification, with the potential for significant reduction in management costs, inventory, and supplier qualification processes. The 723 components with intermediate use represent the most immediate standardization opportunity, as they already have proven moderate reuse, indicating a real demand that justifies the investment in consolidation. Platform components (Group 2) are not candidates for elimination, but for consolidation and protection, with the creation of standardized component families from this group being the most efficient lever to reduce the entry of new components into the portfolio over time.

Limitations

The study presents some limitations that should be considered when interpreting the results. The analysis was restricted to Network A, which comprises the direct links between components and products, excluding Network B, which models the relationships through intermediate subassemblies. The inclusion of Network B could alter the classification of components that are intensely involved in assemblies but appear little in direct links to the final product. Additionally, the analyses were performed on a static temporal snapshot of the data, without considering the historical evolution of the portfolio, meaning that seasonal variations, projects under development, and components being phased out were not captured.

The choice of K = 3 in clustering, although justified by an engineering criterion for greater interpretive granularity, implies some arbitrariness at the boundary between groups. Components close to these boundaries should be treated with caution and evaluated individually before any rationalization decision. Furthermore, the Network Criticality Index (NCI), while useful as a prioritization index, does not replace the technical and economic feasibility analysis of standardization actions; it indicates where to direct attention, but not what to decide specifically.

An additional limitation lies in the use of physical consumption volume in the model. Each LinkID has an adjusted_qty attribute, indicating the quantity of that component in a specific assembly position. However, this granular value was not directly used as edge weights in the bipartite network; weights were defined by the number of LinkIDs for the component-product pair (num_links), treating each link with a unit weight. The total_qty variable, which aggregates adjusted_qty values per component, was considered in the exploratory analysis but excluded before PCA due to high correlation with num_links (r = 0.738, p < 0.001). Consequently, two components with the same num_links receive the same NCI, even if one consumes 1 part per link and the other 500, a difference that can be decisive for the inventory cost, logistics, and supplier qualification associated with each item. The current model captures where the component is used in the structure, but not how much is consumed in each position.

Contributions and future work

The main contribution of this work is methodological, demonstrating that the combination of clustering on the PCA space with weighted bipartite network modeling produces a richer portfolio segmentation than either of the two approaches in isolation. The use of variables derived from the network structure, such as num_links, as input for PCA, allowed capturing structural recurrence as a dimension independent of transversality, which is an advance over approaches that use only purely accounting attributes.

As natural extensions of this work, the analysis can be extended to Network B (subassemblies), allowing the classification of components by hierarchical level in the product structure. The incorporation of economic data, such as unit cost, supplier qualification cost, and purchase volume, is essential to transform structural prioritization into a financial impact analysis. The longitudinal application of the pipeline to monitor the evolution of diversity indicators over time is also a promising path, creating a baseline for evaluating the real effect of implemented standardization initiatives.

In summary, the integrated data science pipeline proved to be an effective tool for quantitatively characterizing component usage patterns in a real industrial portfolio, identifying clear opportunities for standardization and rationalization. Portfolio segmentation into distinct profiles, the structural criticality hierarchy revealed by NCI, and the detection of structural modules offer an actionable roadmap for diversity management, enabling more informed and strategic engineering and supplier management decisions, and directly addressing the objective of optimizing portfolio management.

4. Conclusion

The study sought to apply an integrated data science pipeline to characterize the usage patterns of components in a real industrial portfolio and identify opportunities for standardization and rationalization. It was found that 6,658 components had no recorded usage, representing an immediate simplification opportunity. Additionally, the analysis of 4,580 active components revealed the existence of 538 ultrastructural items, whose systemic criticality demands individualized evaluation. K-means clustering, applied to the principal component space, segmented the remaining portfolio into three distinct profiles: 2,991 components with restricted use, candidates for rationalization; 723 with intermediate use, constituting the main opportunity for immediate standardization; and 328 with high reuse, designated as platform components for consolidation in new projects. Bipartite network modeling and the Network Criticality Index (NCI) confirmed a criticality hierarchy, and the detection of seven structural modules indicated the feasibility of standardization initiatives with controlled scope. The main methodological contribution lies in demonstrating that the combination of clustering with network modeling, using structural recurrence variables as input for PCA, offers a richer and more actionable segmentation of the portfolio, overcoming the limitations of approaches based solely on accounting or functional attributes.

However, the study presented limitations, such as the restriction of the analysis to Network A, which models only direct links between components and products, and the static nature of the data, which did not capture the temporal evolution of the portfolio. The choice of K = 3 in clustering, although justified by engineering criteria, introduced some arbitrariness in the group boundaries, and the NCI, although a valuable prioritization index, does not replace the technical and economic feasibility analysis. An additional limitation was the non-use of physical consumption volume (adjusted_qty) as edge weight in the network, which may impact the prioritization of components with the same number of links but different consumption volumes. For future work, it is suggested to extend the analysis to Network B (subassemblies), incorporate economic data for a more complete financial impact analysis, apply the pipeline longitudinally for continuous diversity monitoring, and include the physical consumption volume as edge weight in the network. In summary, the integrated data science pipeline proved to be a robust tool for strategic component diversity management, providing a quantitative roadmap for optimizing the industrial portfolio.

Bibliographic References

Barabási, A.L. 2016. Network Science. Cambridge University Press, Cambridge, UK. Disponível em: <http://networksciencebook.com>. Acesso em: 10 mar. 2026.

Blondel, V. D.; Guillaume, J. L.; Lambiotte, R.; Lefebvre, E. 2008. Fast unfolding of communities in large networks. Journal of Statistical Mechanics: Theory and Experiment 2008(10): P10008.

Braha, D.; Bar-Yam, Y. 2007. The statistical mechanics of complex product development: empirical and analytical results. Management Science 53: 1127-1145.

Budynas, R. G.; Nisbett, J. K. 2019. Shigley’s Mechanical Engineering Design. 11ed. McGraw-Hill, New York, NY, USA.

Fisher, M. L.; Ittner, C. D. 1999. The impact of product variety on automobile assembly operations: empirical evidence and simulation analysis. Management Science 45: 771–786.

Gil, A. C. 2019. Métodos e Técnicas de Pesquisa Social. 7ed. Atlas, São Paulo, SP, Brasil.

Jiao, J.; Simpson, T. W.; Siddique, Z. 2007. Product family design and platform-based product development: a state-of-the-art review. Journal of Intelligent Manufacturing 18(1): 5-29.

Jose, A.; Tollenaere, M. 2005. Modular and platform methods for product family design. Journal of Intelligent Manufacturing 16: 371-390.

Liu, Z.; Wong, Y. S.; Lee, K. S. 2010. Modularity analysis and commonality design: a framework for the top-down platform and product family design. International Journal of Production Research 48: 3657-3680.

Newman, M. E. J. 2010. Networks: An Introduction. Oxford University Press, Oxford, UK.

Norton, R. L. 2020. Machine Design: An Integrated Approach. 6ed. Pearson, New York, NY, USA.

Prodanov, C. C.; Freitas, E. C. 2013. Metodologia do Trabalho Científico: métodos e técnicas da pesquisa e do trabalho acadêmico. Feevale, Novo Hamburgo, RS, Brasil.

Sosa, M. E.; Eppinger, S. D.; Rowles, C. M. 2007. A network approach to define modularity of components in complex products. Journal of Mechanical Design 129: 1118-1129.

Thevenot, H. J.; Simpson, T. W. 2006. Commonality indices for product family design: a detailed comparison. Journal of Engineering Design 17: 99–119.

Traag, V. A.; Waltman, L.; Van Eck, N. J. 2019. From Louvain to Leiden: guaranteeing well-connected communities. Scientific Reports 9: 5233.

Ulrich, K. T.; Eppinger, S. D.; Yang, M. C. 2020. Product Design and Development. 7ed. McGraw-Hill, New York, NY, USA.

Article originating from the Final Course Project of the Specialization in Data Science and Analytics of the MBA USP/Esalq

To learn more about the course, click here and access the MBX Academy platform