CQI-Based Interference Prediction for Link Adaptation in Industrial Sub-networks

Technical Summary: CQI-Based Interference Prediction for Link Adaptation in Industrial Sub-networks

Problem Statement
In industrial sub-networks (SNs) supporting High-Reliability and Low-Latency Communication (HRLLC), traditional Link Adaptation (LA) faces significant challenges. These environments feature dense deployments and limited sub-bands, leading to substantial inter-SN interference with heavy-tailed probability density functions (PDFs). Furthermore, strict latency constraints (0.1–1 ms) and protocol delays render standard feedback mechanisms, such as Hybrid Automatic Repeat Request (HARQ) and Acknowledgement (ACK) signals, impractical. Consequently, Channel Quality Indicator (CQI) reports available at the Sub-network Controller (SNC) are often outdated, quantized, and compressed. Existing LA solutions struggle with these conditions: Gaussian-based assumptions fail to model the heavy-tailed nature of interference, while data-driven approaches like Gaussian Process Regression (GPR) suffer from prohibitive computational complexity ($O(L^3)$) and lack robustness to outliers. Additionally, many current methods require additional signaling overhead or rely on perfect interference knowledge at the User Equipment (UE) side, which is not feasible in this context.

Methodology
The authors propose a novel predictive LA framework that operates entirely at the SNC using only available CQI reports. The core of the methodology involves three integrated components:

1. Vector Discrete-Time State-Space Model (vDSSM): The interference power vector (IPV) is treated as a latent variable. The system models the evolution of IPV and its relationship to observed CQI through unknown process and measurement functions. This model accounts for quantization errors, SINR estimation errors, and protocol-induced delays.
2. Sparse Student-t Process Regression (SPTPR): To address the heavy-tailed nature of interference and computational constraints, the authors replace standard GPR with SPTPR. This Bayesian non-parametric approach utilizes a Student-t distribution, which is robust to outliers and better captures sporadic traffic events than Gaussian assumptions. By employing "inducing points," the method reduces computational complexity from $O(L^3)$ to $O(L_E^2 \cdot L)$, where $L_E \ll L$.
3. Modified Unscented Kalman Filter (MUKF): To handle the non-linear dynamics of the SPTPR models and compensate for protocol delays, a MUKF is employed. Unlike linear approximations, the MUKF uses a modified unscented transform with sigma-points derived from the Student-t distribution to accurately propagate mean and covariance. The algorithm recursively refines interference predictions by aligning delayed CQI observations with the predicted state, effectively compensating for the time lag without requiring additional signaling.

Key Contributions

• Novel vDSSM Framework: This work presents the first vector dynamic state-space model for LA in industrial SNs that moves beyond Gaussian assumptions to explicitly model heavy-tailed interference.
• Robust and Scalable Inference: The integration of SPTPR allows for the robust estimation of interference power under sparse, heavy-tailed conditions while significantly reducing computational complexity compared to non-parametric baselines.
• Delay Compensation without Overhead: The proposed scheme compensates for protocol delays and outdated CQI entirely at the SNC. It eliminates the need for dequantizers or additional signaling overhead from the Sensor-Actuator (SA) pairs, adhering to 3GPP signaling constraints.
• Data-Driven Approach: The method learns process and measurement models directly from observed CQI data, avoiding the need for complex analytical channel models that may lack generalizability in dynamic industrial settings.

Numerical Results
Simulations were conducted in a 25 × 25 m² factory hall with 20 SNs, operating at 6 GHz with 3GPP InF-DL channel models. The proposed SPTPR-MUKF was compared against a "Genie" LA (upper bound), a supervised LSTM predictor, a Conventional Moving-Average (MA) predictor, and a GPR-based approach.

• Performance: The proposed method achieves a Block Error Rate (BLER) below the 90th percentile target of $10^{-6}$, even with protocol delays up to 10 ms. Its performance is comparable to the supervised LSTM baseline and the "Genie" LA, significantly outperforming the MA and delayed estimate methods.
• Complexity: The SPTPR-MUKF achieves over 10× lower computational complexity compared to similar non-parametric baselines.
• Scalability: The method maintains high spectral efficiency and low BLER as the number of interferers increases and in multi-SA scenarios (up to 4 SAs per SN).
• Robustness: The system effectively mitigates the impact of protocol delays and outliers, maintaining performance where other methods degrade.

Significance and Claims
The paper claims to be the first work to utilize a vDSSM-based approach for LA in industrial SNs that incorporates heavy-tailed interference prediction beyond Gaussian assumptions. It highlights that by leveraging CQI reports directly through a sparse Student-t process and a modified Kalman filter, the system can achieve near-optimal link adaptation performance without incurring the signaling overhead or energy costs associated with transmitting interference estimates from the edge devices. The authors position this work as a critical step toward enabling single-transmission reliability in dense, interference-prone industrial environments where retransmissions are not feasible. Future work is noted to focus on real-time deployment and generalization across diverse scenarios.