Answers are grounded in this paper and cite their sources.
Anomaly detection in cyber-physical systems (CPS) is difficult because system behaviour changes over time, and new attacks can appear after deployment. Most TinyML-based anomaly detection systems rely on static, pre-trained models, which cannot adapt once deployed. In addition, ground-truth labels are often unavailable during real CPS operations, restricting effective online learning. We present TinyLearnED, a weakly supervised framework for online anomaly detection. The framework supports online adaptation during the evaluated data streams without requiring post-deployment ground-truth labels. In particular, the framework exploits correlations among independent edge devices that observe the same physical environment. By exchanging binary anomaly decisions, these devices collaboratively generate pseudo-labels for training. Based on these pseudo-labels, each device performs local online fine-tuning, while a lightweight federated aggregation step averages final-layer parameters to smooth updates and improve resilience under the evaluated noisy and compromised-node conditions. During this process, the pre-trained feature extractor remains unchanged, and only a lightweight output layer is fine-tuned. We evaluate TinyLearnED on four XIAO ESP32S3 monitoring nodes using chronological replay of the Bolt 3D dataset, collected from one Prusa MK4 printer producing a single bolt geometry with triaxial accelerometer sensing. The results demonstrate that on-device adaptation is feasible without post-deployment ground-truth labels and yields task-dependent detection improvements in this testbed. Cross-geometry transfer and long-term deployment stability remain unevaluated.
No takes yet. Share an insight, caveat, or question.
Aljohani et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: