Reproduce “On-device Anomaly Detection in Conveyor Belt Operations” on OpenResearch

About this paper

Luciano S. Martinez-Rau, Yuxuan Zhang, Bengt Oelmann, Sebastian Bader

Researchers at Mid Sweden University introduced two on-device pattern recognition methods for real-time anomaly detection in conveyor belt operations. These approaches achieve enhanced duty cycle detection and classification accuracy by independently handling these tasks, demonstrating robustness against machine drift and high energy efficiency on low-power microcontrollers.

Key insights

• Decoupling duty cycle detection from operation mode classification using a simple speed threshold significantly improves detection accuracy and robustness, especially under conditions involving data drift.
• Employing a second machine learning model for duty cycle classification can better manage the complexity and variability found in long-term operational datasets compared to fixed heuristic rules.
• Quantizing the models to 8-bit integers minimally affects classification accuracy but substantially reduces energy consumption on low-power microcontrollers, enabling practical TinyML deployment.

Summary by alphaXiv.