Cross-robot open-set fault detection via normal-only adaptation and residual latent-space alignment

Abstract

Reliable fault detection and health monitoring of industrial robots are important for smart manufacturing in Industry 4.0. However, robotic platforms differ in their mechanical design, sensing characteristics, and operating conditions. These differences make it difficult for data-driven models to generalize to unseen robots and working conditions. To address this challenge, this paper proposes a cross-robot framework for open-set fault detection using only normal data from the target robot. The framework uses a condition-normalized residual adaptation mechanism in the latent space. First, a structured representation is learned from the source robot. The learned representation is then adapted to the target robot using only normal samples. Compactness-based constraints and statistical modeling of normal behavior are used to detect unknown faults as deviations from the normal representation. The framework is evaluated on two industrial robots under variations in load, speed, and motion. Two evaluation strategies are also considered: sequential normal-only adaptation to welding motion and direct open-set detection without further adaptation. The results show consistent fault detection across different robots and operating conditions. A comparative benchmarking analysis further demonstrates the effectiveness of the proposed approach and its ability to maintain reliable fault detection performance under different loading and motion conditions. This demonstrates the potential of the framework for practical industrial health monitoring.