Original scientific paper
https://doi.org/10.30765/er.3431
A constrained generative framework for scaling metrology data in multi-component automotive assemblies
Faizan Ali
orcid.org/0009-0007-2945-2149
; Technische Hochschule Ingolstadt, AImotion Bavaria, Ingolstadt, Germany
*
Ameya Punekar
orcid.org/0009-0004-3655-1581
; Technische Hochschule Ingolstadt, AImotion Bavaria, Ingolstadt, Germany
Juergen Bock
orcid.org/0000-0002-1210-1576
; Technische Hochschule Ingolstadt, AImotion Bavaria, Ingolstadt, Germany
* Corresponding author.
Abstract
Automotive pilot production yields limited metrology data, which hinders the development of robust quality-assurance models. This paper presents a constrained generative framework that synthesises multi-component assembly data while ensuring adherence to production boundaries. The pipeline combines Bayesian hyperparameter optimisation with rejection sampling and evaluates three generative architectures: a conditional tabular generative adversarial network, a tabular variational autoencoder, and a Gaussian copula model. The framework was tested on a roller-slide door assembly comprising eight components and approximately 1,200 real measurements, across 360 experimental conditions spanning three generative models, three scaling factors and five random seeds. Constraint enforcement was implemented as a post-hoc filter applied outside the generator, so that the same compliance layer wraps all three architectures without modification. Full adherence to the encoded empirical bounds and the torque-ordering rule was confirmed in every condition. Hyperparameter optimisation proved critical, increasing the mean Kolmogorov–Smirnov pass rate from 49.0 % to 94.0 % for the adversarial model, from 63.5 % to 89.0 % for the variational autoencoder, and from 55.2 % to 79.0 % for the copula model. The results confirm that tuned generative models can successfully scale scarce metrology datasets.
Keywords
synthetic data generation; manufacturing metrology; constraint enforcement; rejection sampling; hyperparameter optimisation; tabular generative models
Hrčak ID:
350093
URI
Publication date:
4.8.2026.
Visits: 0 *