A Q-learning-assisted self-healing wireless sensor network for industrial carbon capture monitoring: a controlled HIL study

Abstract

Wireless sensor networks (WSNs) used for carbon capture and storage (CCS) monitoring must maintain low latency, high packet delivery ratio, rapid recovery after node or link failure, and stable energy consumption under harsh industrial conditions. Existing static and reactive adaptive routing approaches often treat fault recovery, rerouting, and energy optimization as separate functions, which limit their ability to provide coordinated resilience in CCS-like environments. This study develops and evaluates a CCS-oriented system-level integration of Q-learning-assisted routing, fault-aware recovery, and energy-aware communication within a hybrid mesh-star WSN. The routing agent jointly considers link quality, node health, residual energy, queue length, latency, hop count, and energy utilization factor (EUF) during routing and recovery decisions. The reward coefficients were selected through an independent constrained grid-search calibration using simulation scenarios that were separated from the final HIL evaluation. The framework was evaluated using controlled simulation and a 50-node Hardware-in-the-Loop (HIL) testbed under nominal, node-failure, high-load, and attenuation scenarios. Compared with the reactive adaptive baseline, the proposed framework reduced end-to-end latency by 53.6%, from 140 ms to 65 ms, improved fault recovery time by 65.9%, from 8.5 s to 2.9 s, increased PDR to 99.1%, improved EUF to 0.95, and reduced power consumption by 22.2%, from 1.53 kW to 1.19 kW. Under 25% node failure, the framework maintained a PDR of 99.1% with a recovery time of 2.9 s. Additional reward-weight sensitivity and extended-duration experiments were included to assess parameter robustness and operational stability beyond the original 600-s sessions; the proposed method remained top-ranked across all tested profiles, while ±20% coefficient perturbations changed pooled PDR by no more than 0.4 percentage points and p95 latency by no more than 8 ms; the 21,600 s mixed-stress sessions maintained 98.3% PDR, 98 ms p95 latency, and 95.6% recovery success. The techno-economic estimation indicated a 2.17-year payback period, a 36% reduction in annual OPEX, and an estimated 70 tCO 2 /year reduction under the stated assumptions. The validation was performed in a controlled HIL and simulation environment; therefore, the reported industrial implications should be interpreted as deployment potential rather than evidence from a full-scale CCS plant. The contribution is a CCS-oriented integration and validation framework rather than a fundamentally new reinforcement-learning algorithm.