~/g8dr1ch/research

Research, published
and explained.

Peer-reviewed work from the studio, each paper walked through in plain steps with its original figures.

thesis_00.md2024

Pengembangan Sistem Identifikasi Similaritas Produk Engineer-to-Order dengan Metode Dekomposisi Matriks SVD

Developing a similarity identification system for engineer-to-order products with SVD matrix decomposition
Bachelor thesis · Industrial Engineering, Bandung Institute of Technology · the work the paper below grew from

itb_digilib ↗
↓ peer reviewed and published
paper_01.md2025

Similarity Identification System Design for Engineer-to-Order Products with SVD Matrix Factorization

Ignatius Vito W.P. Hapsara, Mohammad M. Isnaini, Jeisa R. N. Tarigan
Procedia CIRP 134 (2025) 657–662 · CC BY-NC-ND 4.0 · Bandung Institute of Technology

@article{hapsara2025svd,
  author  = {Hapsara, Ignatius Vito W.P. and Isnaini, Mohammad M. and Tarigan, Jeisa R. N.},
  title   = {Similarity Identification System Design for Engineer-to-Order Products with SVD Matrix Factorization},
  journal = {Procedia CIRP},
  volume  = {134},
  pages   = {657--662},
  year    = {2025},
  doi     = {10.1016/j.procir.2025.02.179}
}
read_paper ↗
// how it works
encode.md01

Encode the product

Every product is described by two parameters. Its material structure comes from the material use voucher, mapped into a binary matrix of 609 materials and components by 9 types. Its features come from a six-segment hierarchical code, Y-S0-S1-S2-S3-Z, nine digits long: 132210040 reads as a standard 40 kl water truck. The code is then encoded as a matrix too.

Fig. 2. Parameter transformation
Fig. 2. Parameter transformation. From Ignatius Vito W.P. Hapsara, Mohammad M. Isnaini and Jeisa R. N. Tarigan, Procedia CIRP 134 (2025) 657–662, CC BY-NC-ND 4.0.
decompose.md02

Decompose with SVD

Each parameter matrix is factorised into three parts. The middle singular matrix Σ carries the spectrum, how strongly the structure is expressed; the left and right singular matrices U and VT carry its row and column patterns. Reducing to these components filters noise out of the comparison.

Eq. 1. Singular value decomposition of parameter j for product p
Eq. 1. Singular value decomposition of parameter j for product p. From Ignatius Vito W.P. Hapsara, Mohammad M. Isnaini and Jeisa R. N. Tarigan, Procedia CIRP 134 (2025) 657–662, CC BY-NC-ND 4.0.
measure.md03

Measure similarity

Σ is compared with the Frobenius norm, simF = [1 + ‖Σp − Σp′‖F]−1; U and VT with cosine similarity. A general equation X combines every method i and parameter j with weights that each sum to 1, and a Hadamard term for parameters that interact. The weights are tuned through a design of experiments with stakeholders.

Fig. 4. Similarity matrix processing
Fig. 4. Similarity matrix processing. From Ignatius Vito W.P. Hapsara, Mohammad M. Isnaini and Jeisa R. N. Tarigan, Procedia CIRP 134 (2025) 657–662, CC BY-NC-ND 4.0.
Eq. 5 and 6. The general similarity equation and its simplified form
Eq. 5 and 6. The general similarity equation and its simplified form. From Ignatius Vito W.P. Hapsara, Mohammad M. Isnaini and Jeisa R. N. Tarigan, Procedia CIRP 134 (2025) 657–662, CC BY-NC-ND 4.0.
Eq. 11 and 12. The trial weighting used for the network maps
Eq. 11 and 12. The trial weighting used for the network maps. From Ignatius Vito W.P. Hapsara, Mohammad M. Isnaini and Jeisa R. N. Tarigan, Procedia CIRP 134 (2025) 657–662, CC BY-NC-ND 4.0.
map.md04

Map the network

X becomes a product network: each node is a truck model, each line a similarity, thicker for closer products. Across 36 models the clusters fall out on their own: open trailers, low-beds, fuel trucks and water trucks. A new product with no material list borrows one from its nearest neighbour by feature code. Rebuilding the map from a 609 × 3 matrix instead of 609 × 9 kept the same structure.

Fig. 5 and 6. Product network maps at thresholds 0.42 and 0.47
Fig. 5 and 6. Product network maps at thresholds 0.42 and 0.47. From Ignatius Vito W.P. Hapsara, Mohammad M. Isnaini and Jeisa R. N. Tarigan, Procedia CIRP 134 (2025) 657–662, CC BY-NC-ND 4.0.
system.md05

The whole system

The modules chain together: initial recognition and codification feed similarity measurement, which feeds product mapping and material recommendation, all inside one Python environment.

Fig. 1. System framework design
Fig. 1. System framework design. From Ignatius Vito W.P. Hapsara, Mohammad M. Isnaini and Jeisa R. N. Tarigan, Procedia CIRP 134 (2025) 657–662, CC BY-NC-ND 4.0.