Inteligência artificial para diagnóstico e prognóstico na engenharia mecânica : metodologia, aplicações e análise de explicabilidade em injeção plástica e rolamentos

Abstract

The increasing integration of Artificial Intelligence (AI) in mechanical engineering has led to highly accurate predictive models. However, many of these applications still operate as ’black boxes,’ lacking a methodological analysis to validate the physical coherence of the learned knowledge and ensure trust in their outcomes. This work addresses this gap by proposing a risk-oriented methodological framework for the application and, crucially, the explainability—the ability to interpret how models generate their predictions—of Machine Learning techniques in two distinct case studies: predicting the quality of parts produced by plastic injection molding and prognosing the Remaining Useful Life (RUL) of bearings. The central aim is to develop predictive models whose reliability is assessed not only by performance metrics but also by the coherence of their learned knowledge and alignment with engineering requirements. In the first case study, an optimized Artificial Neural Network was developed to predict the occurrence of defects in manufactured products. The model achieved 99.39% accuracy and demonstrated excellent recall for critical defect classes, a design choice guided by the application of the Failure Mode and Effects Analysis (FMEA) tool. Analysis with SHAP (SHapley Additive exPlanations) validated the model’s learning by identifying physically relevant process variables. The second case study addressed bearing prognosis. Although sensitivity analysis confirmed that the model learned physically coherent relationships, predictive performance was limited by data scarcity. The results evidence that the successful application of AI in engineering critically depends on data quality and a methodology that integrates risk analysis, technical validation, and governance principles. Explainability proved to be fundamental for building trust and promoting the necessary transparency to meet the requirements of AI management systems, such as those advocated by the ISO 42001 standard.