Multi-Class Anomaly Detection in an Industrial Robot Arm

Abstract

This paper introduces a specialized Parallel Inception-style 1D-CNN (PINC) architecture designed for real-time condition monitoring of the Universal Robots UR3e platform. Traditional deep learning models for anomaly detection often require substantial computational resources. In contrast, our proposed PINC model uses per-channel parallel feature extraction to identify multi-scale physical anomalies with high accuracy. Using the CASPER dataset, the architecture achieves a diagnostic accuracy of 99.1% while maintaining a low memory footprint of just 152 KB. These findings show that the PINC model effectively captures vibrational signatures without depending on high-latency cloud processing. The results offer a scalable and efficient hardware solution that enables localized, real-time fault detection in human-robot shared workspaces.