Classificação multiclasse baseada em séries temporais multivariadas do estágio de operação de ferramentas de corte de torno

Abstract

The majority of mechanical components went through a machining process during their manufacturing. Therefore, manufacturing processes with inadequate condition tools are likely to induce unexpected operational interruptions, accidents, product quality, and economic losses. Accordingly, the ability to classify fault imminences can result in cost reduction, along with productivity and safety increase. This work presents a low-cost data acquisition system and aims to discuss a model based on digital signal processing, the SelfOrganised Direction Aware Data Partitioning Algorithm (SODA) and machine learning techniques, including time series Feature Extraction based on Scalable Hypothesis tests (TSFRESH) and Principal Components Analysis (PCA), to solve this problem. Taking into consideration the real-time monitoring, features selection and features relevance analysis are essential as it can lead to a dimensionality and data flow rate reduction. The model proposed in this work can identify the patterns that distinguish the cutting tool’s flank wear in a multi-class scenario as adequate, intermediate, and inadequate conditions, achieving satisfactory performances in all cases and allowing to prevent fault occurrences.