Introduction
In the evolution toward 6G, industrial automation is shifting from centralized control to highly localized, ultra-reliable communication environments. A key enabler of this transition is the In-Factory Subnetwork (InF-S)—small, low-power radio cells integrated directly into industrial robots, production modules, or moving parts. These subnetworks must support extremely tight control loops where sensors and actuators communicate in sub-millisecond cycles.
Traditional network performance metrics, such as raw throughput or packet delivery ratio, are often insufficient for these industrial applications. Instead, the stability of a robotic control system depends on the "freshness" of the data it receives. This paper addresses the challenge of maintaining data freshness through a proactive resource allocation framework designed to minimize the Age of Information (AoI) in dynamic factory environments.
The Challenge: Beyond Reactive Management
Age of Information (AoI) measures the time elapsed since the last successfully received update was generated. In a factory setting, high AoI can lead to desynchronization in robotic arms or safety failures. However, factory floors are notoriously difficult radio environments; mobility, physical obstructions, and electromagnetic interference create non-stationary channel conditions.
Standard resource allocation methods are typically reactive, responding only after a packet is dropped or a delay is detected. In the context of 6G InF-S, this delay in response is often too late to prevent an AoI violation. To maintain sub-millisecond reliability, the network must become predictive, anticipating channel fluctuations and interference before they impact data freshness.
Proposed Methodology: Bayesian Proactive Allocation
The authors propose a decentralized, online learning framework that allows each subnetwork to manage its radio resources (power levels and frequency blocks) autonomously. The core of this framework is Bayesian Ridge Regression (BRR).
1. Bayesian Ridge Regression (BRR)
Unlike traditional regression models that provide a single point estimate, BRR treats model weights as random variables. This allows the system to predict not only the expected future AoI but also the uncertainty (variance) associated with that prediction. This uncertainty is a critical signal for the resource manager.
2. Active Learning and the Exploration-Exploitation Trade-off
The framework employs an active learning strategy to balance two competing objectives:
- Exploitation: Selecting resource configurations (power and frequency) predicted to minimize the probability of an AoI violation.
- Exploration: Selecting configurations where the model has high uncertainty. By "exploring" these unknown states, the system gathers data to improve its future predictive accuracy.
3. Decentralized Architecture
To minimize overhead and latency, each subnetwork runs its own BRR model locally. The models are updated at every time slot, allowing the system to adapt to the non-stationary nature of the factory floor in real-time.
Experimental Setup
The framework was evaluated using Monte Carlo simulations modeled after 3GPP standards for indoor factory environments. The simulation environment consisted of a 20m x 20m area featuring 20 mobile subnetworks. Key parameters included:
- Channel Model: Rician fading to simulate line-of-sight conditions.
- Time Slots: 3ms duration.
- AoI Threshold: 10ms (the limit before a violation is recorded).
- Sampling Rates: Tested at 1 Mbps and 2 Mbps.
Key Results and Insights
Massive Reduction in Violations
The proactive BRR-based method demonstrated significant performance gains. At a 1 Mbps sampling rate, the system achieved a 98% reduction in the probability of AoI violations compared to default random resource allocation. It also consistently outperformed "Greedy" algorithms, which focus solely on immediate AoI without accounting for model uncertainty or future interference.
The "Goldilocks" Dataset Size
A critical finding of the study is that more data is not always better in dynamic environments. Because factory conditions change rapidly, keeping a long history of training data can actually degrade performance. The authors found that keeping approximately 200 samples provided the optimal balance; beyond this, the model begins to fit "stale" patterns that no longer reflect the current state of the environment.
Impact of Sampling Rates
As the data sampling rate increases (e.g., from 1 Mbps to 2 Mbps), the bottleneck for AoI shifts from channel access to queuing time. The study showed that proactive management becomes even more critical at higher rates, as the system must account for the buildup of packets in the buffer to prevent staleness.
Conclusion
This research highlights that in the 6G era, predictability is a prerequisite for reliability. By utilizing Bayesian Ridge Regression and active learning, In-Factory Subnetworks can transition from reactive troubleshooting to proactive resource management.
Future work will focus on automating the tuning of exploration-exploitation parameters and addressing the hardware constraints of implementing Bayesian models on low-power industrial chips. Additionally, the complex interplay between inter-subnetwork and intra-subnetwork interference remains a vital area for further optimization.
Summary updated