Adapting Audio-Spectrogram Transformers for Industrial Time-Series Sensor Data

Abstract

Launcher manufacturing is undergoing a rapid transformation, driven by the need for increased production robustness, cost reduction, and enhanced sustainability. In particular, the adoption of automation and digitalisation in production processes has become essential to remain competitive in today’s space industry. This thesis contributes to this transition by developing a data-driven anomaly detection system for robotic drilling and riveting, a critical step in the assembly of launcher structures. Drawing inspiration from the field of audio classification, in which transformer-based models have recently achieved state-of-the-art results, this research proposes a novel adaptation of the spectrogram transformer architecture to the domain of industrial time-series sensor data. The resulting deep learning-based approach enables automated detection of anomalous behaviour through sensor signals, offering a scalable solution to process monitoring.