Time-Series Anomaly Detection for Industrial Screwdriving Task with Machine Learning Algorithms
Author
Vucicevic, Ozren
Term
4. term
Education
Publication year
2023
Submitted on
2023-06-01
Pages
78
Abstract
In modern automated factories, it is important to detect when a screwdriving operation does not run as expected. This project investigates whether machine learning (ML)—computer methods that learn patterns from data—can serve as a reliable tool to identify such irregularities (anomalies) in industrial screwdriving. To capture the process from multiple angles, the team collected time-series measurements from screwdriver sensors, signals from a UR10 industrial robot arm, and audio recordings from a microphone. The data were then preprocessed to clean and format them, and feature selection was applied to choose the most informative attributes, strengthening the models’ ability to learn relevant patterns. In the final phase, classification algorithms were developed to distinguish between different classes of screws; together, these steps form an approach to anomaly detection in screwdriving processes. The research was conducted at Aalborg University in partnership with VELUX, providing a real-world manufacturing context. Looking ahead, the aim is to improve the quality and efficiency of their automated screwdriving line by reducing manufacturing anomalies. If successful, the project would support the viability of integrating ML into industrial applications and pave the way for future work to optimize the algorithms and extend the approach to other manufacturing domains.
I moderne, automatiserede fabrikker er det vigtigt at opdage, når en skruningsproces ikke forløber som forventet. Dette projekt undersøger, om maskinlæring (ML) – computerprogrammer der lærer mønstre fra data – kan bruges som et pålideligt værktøj til at finde sådanne uregelmæssigheder (anomalier) i industrielle skruningsopgaver. For at beskrive processen fra flere vinkler blev der indsamlet tidsseriedata (målinger over tid) fra skruetrækkerens sensorer, data fra en UR10-robotarm samt lydoptagelser fra en mikrofon. Dataene blev derefter forbehandlet for at rense og formatere dem, og der blev udført feature selection (udvælgelse af de mest informative træk i data) for at styrke modellernes evne til at lære relevante mønstre. I den afsluttende fase blev der udviklet klassifikationsalgoritmer, som kan skelne mellem forskellige klasser af skruer; samlet set udgør disse trin en fremgangsmåde til anomalidetektion i skruningsprocesser. Arbejdet er udført på Aalborg Universitet i samarbejde med VELUX, som har bidraget med en virkelig produktionskontekst. På sigt er målet at forbedre kvalitet og effektivitet i deres automatiserede skruningslinje ved at mindske fabriksafvigelser. En vellykket gennemførsel vil underbygge, at ML kan integreres i industrielle anvendelser, og den baner vejen for videre forskning i at optimere algoritmerne og udbrede metoden til andre produktionsområder.
[This apstract has been rewritten with the help of AI based on the project's original abstract]