Müller K, Franchi N (2025)
Publication Type: Conference contribution
Publication year: 2025
Pages Range: 1-7
Conference Proceedings Title: 2024 IEEE 100th Vehicular Technology Conference
DOI: 10.1109/VTC2024-Fall63153.2024.10758005
Subnetworks with different levels of autonomy are an integral part of 6G, helping to improve local connectivity, resilience, and security. These are especially interesting for cooperative robotics on shop floors, where multiple robots can come together and directly exchange task-specific data in near real-time without stressing the parent network. The same sub-network capabilities are currently implemented into Automated Guided Vehicles (AGVs), allowing non-time-critical data from near monitoring equipment to be offloaded via AGVs instead of streaming them directly to the shop floor's parent network, relieving the parent network in the process. However, to do so, an efficient approach is needed, which (1) can predict future encounters between AGVs and monitoring nodes on a node-to-node basis with minimal traffic overhead, (2) does not require heavy computation like machine learning on highly constraint nodes, and (3) does not need access to AGVs' subnetwork controllers and routing tables.Introducing ConTST (CONnection forecasting with Time Series Transformers), a semi-passive, probabilistic approach for predicting future data offloading possibilities to autonomous mobile subnetworks on a node-to-node basis from only successfully received bundles in the network edge. We utilize a single, small probabilistic time series transformer model in the network edge, minimal additional metadata per bundle, and some tracing bundles to successfully estimate future encounters from past encounters. Nodes optimize their offloading strategy greedily by only utilizing precomputed encounter tables from the edge services, making this approach compatible even with the most constrained IoT devices. We extensively evaluate our approach on our shop floor simulator and show that it is resilient against outliers, capable of predicting future encounters, and can reliably optimize overall offloading Quality of Service (QoS) metrics like bundle latency or lost bundle ratio.
APA:
Müller, K., & Franchi, N. (2024). ConTST: Predicting Mobile Subnetwork Encounters in Dynamic Factories with Time Series Transformers. In 2024 IEEE 100th Vehicular Technology Conference (pp. 1-7). Washington DC, US.
MLA:
Müller, Kilian, and Norman Franchi. "ConTST: Predicting Mobile Subnetwork Encounters in Dynamic Factories with Time Series Transformers." Proceedings of the IEEE Conference on Vehicular Technology (VTC2024-Fall), Washington DC 2024. 1-7.
BibTeX: Download