Abordagens de aprendizado de máquina para reconhecimento de padrões em processos de manufatura

Abstract

Machine learning approaches for pattern recognition in manufacturing processes are increasingly prevalent in the context of Industry 4.0. This scenario allows factories to develop new methodologies for monitoring and controlling the quality of their products, obtaining better indicators of operational efficiency, and offering increasingly competitive products in the market. The interpretability of machine learning models can facilitate understanding how models make decisions and can assist in tracking for defect detection and anomalies. This work presents results from supervised and unsupervised machine learning models for defect and anomaly detection in the manufacturing process within the context of a multinational industry. The goal of this work is to develop machine learning models for multi-class and binary attribute classification, as well as anomaly detection models. The data come from the tire uniformity process. The Random Forest, Gradient Boosting Decision Tree, Light Gradient Boosting Machine, Logistic Regression, Support Vector Machine, Multi-layer Perceptron, K-Nearest Neighbor, and Gaussian Processes Classification models were considered for classifying tire performance in compliance with production standards. The Local Interpretable Model-Agnostic Explanations and SHapley Additive exPlanations methodologies were used for model interpretation. In the context of multi-class classification, the Random Forest model yielded robust and satisfactory results for classifying tire performance. The model also showed adequate results for defect detection in binary classification; however, the Logistic Regression model achieved comparable results according to the McNemar’s statistical test. Logistic Regression was used to determine reference values for each selected predictor variable for tire manufacturing in compliance with quality standards. For anomaly detection, the following models were considered: Isolation Forest, Local Outlier Factor, One Class Support Vector Machine, and Elliptic Envelope. The Isolation Forest model provided satisfactory results for anomaly detection. With the help of the SHapley Additive exPlanations technique, it was possible to identify the variables with the greatest influence on the model. These different approaches are relevant tools and provide robust solutions to ensure the efficiency of quality control management in manufacturing processes.