Utvärdering av maskininlärningsmodeller för att minimera driftstopp genom operatörsuppgiftsschemaläggning i CNC-produktionslinjer
Year: 2025
Type: dissertation
Abstract: In computer numeric control (CNC) production lines, recurring tasks such as tool changes and quality inspections are still manually performed by operators. When CNC machine (MC) cycles become misaligned, these tasks may overlap, leading to increased downtime, reduced throughput, and greater operator workload. This thesis explores how machine learning can support operator task scheduling by predicting the timing and overlap of these interventions. To address limitations in historical MC data, a simulation program was developed based on a mathematical model of MC behavior, operator activities, and automation logic. The simulation generated synthetic signals that were used to train and evaluate machine learning (ML) models capable of forecasting operator-dependent events. The study compares three model architectures Long short-term memory (LSTM), Convolutional neural network (CNN), and Temporal convolutional network (TCN)—using performance metrics including F1-score, precision, recall, Hamming loss, and AUC. The results show that no single model consistently outperformed the others, but architectural differences affected performance depending on task frequency and machine position in the line. LSTM models demonstrated advantages in handling long-range dependencies, particularly under continuous learning conditions. However, low label frequency and data sparsity limited overall performance. Despite these challenges, the findings suggest that predictive scheduling with ML can enhance production line efficiency if supported by better signal tracking and model specialization. The study also outlines future research directions, including model modularization, graph-based learning, and reinforcement learning for task optimization.
Author: Özsan, Kaan
Language: English
Subfield: Industrial and Manufacturing Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6997238464