Aplicação de técnicas de aprendizado de máquina na predição da maturidade lean com base no modelo LESAT

Abstract

Lean maturity assessment in industrial organizations is frequently conducted through qualitative self-assessment instruments, which are subject to subjectivity, interpretive variability, and low replicability. This study proposes a methodological approach for predicting lean maturity levels from structured qualitative data and supervised machine learning techniques. The model aims to structure and standardize the subjectivity inherent in traditional lean maturity assessment approaches, promoting greater scalability and reliability in organizational diagnostics. The data collection instrument was developed based on the LESAT 2.0 dimensions, comprising 41 items distributed across lean organizational practices and tool adoption, applied to 43 professionals from Brazilian industrial organizations via structured questionnaire. The target variable was constructed through a weighted composite score and divided into three maturity classes: Initial, Developing, and Advanced. The assigned maturity levels served as labels for training two supervised classification algorithms: Decision Tree and Random Forest. The Decision Tree, with an accuracy of 0.674 in the LOOCV, generated classification rules whose decision hierarchy mirrors the theoretical progression of the LESAT 2.0 framework. The Random Forest model achieved an accuracy of 0.907 and a macro F1-score of 0.905 in the Leave-One-Out Cross-Validation, with variable importance analysis identifying the dimensions of organizational structure, transformation planning, implementation, and transformation nurturing as the main determinants of maturity level. In summary, the proposed approach demonstrates methodological feasibility and offers a more objective, scalable, and reproducible diagnostic instrument, contributing to the integration between Production Engineering and Data Science.