Publicação

Exploring automotive quality correlations through explainable machine learning what-if simulation

Ver documento

Detalhes bibliográficos
Resumo:High-dimensional variability in manufacturing processes presents significant challenges for quality control, demanding predictive strategies capable of capturing complex parameter dependencies. Machine learning (ML) offers robust mechanisms for this purpose, but reliance on black-box models often limits interpretability and hinders producing stakeholders’ identification of meaningful correlations for model optimization. This paper introduces an interactive what-if simulation platform designed to explore structural quality correlations in automotive assembly through explainable ML techniques, enhancing transparency and enabling uncertainty quantification. The platform is based on a modular Digital Twin (DT) architecture aligned with the ISO 23247 standard, guiding expert and non-expert users through correlation-driven feature selection, regression modelling and SHapley Additive exPlanations (SHAP) based post-hoc explanations. A case study using real inspection data from a vehicle assembly line demonstrates the tool’s capacity to support variable relevance assessment, dimensionality reduction, and model interpretability. Furthermore, an uncertainty-aware SHAP analysis enhances confidence in the model’s prediction stability, reinforcing the platform’s suitability for quality-driven decision support and integration into future DT ecosystems.
Autores principais:Oliveira Júnior, Alexandre de
Outros Autores:Calvo-Rolle, JoséLuis; Pires, Rui; Leitão, Paulo
Assunto:Explainable Machine Learning Correlation Analysis Uncertainty-aware Simulation Industrial Metaverse Digital Twin
Ano:2025
País:Portugal
Tipo de documento:comunicação em conferência
Tipo de acesso:acesso aberto
Instituição associada:Instituto Politécnico de Bragança
Idioma:inglês
Origem:Biblioteca Digital do IPB
Descrição
Resumo:High-dimensional variability in manufacturing processes presents significant challenges for quality control, demanding predictive strategies capable of capturing complex parameter dependencies. Machine learning (ML) offers robust mechanisms for this purpose, but reliance on black-box models often limits interpretability and hinders producing stakeholders’ identification of meaningful correlations for model optimization. This paper introduces an interactive what-if simulation platform designed to explore structural quality correlations in automotive assembly through explainable ML techniques, enhancing transparency and enabling uncertainty quantification. The platform is based on a modular Digital Twin (DT) architecture aligned with the ISO 23247 standard, guiding expert and non-expert users through correlation-driven feature selection, regression modelling and SHapley Additive exPlanations (SHAP) based post-hoc explanations. A case study using real inspection data from a vehicle assembly line demonstrates the tool’s capacity to support variable relevance assessment, dimensionality reduction, and model interpretability. Furthermore, an uncertainty-aware SHAP analysis enhances confidence in the model’s prediction stability, reinforcing the platform’s suitability for quality-driven decision support and integration into future DT ecosystems.

Atividades financiadas

Carregando projetos financiados...