Superpressure balloons in the stratosphere are emerging as a promising platform for extending network connectivity to remote and disaster-affected regions. Yet, their dependence on altitude control and absence of horizontal maneuverability render them highly susceptible to wind drift, leading to unstable coverage. Coordinating multiple balloons further challenges current multi-agent reinforcement learning (MARL) methods, as wind-induced moderation effects can simultaneously disrupt all agents. Here we present a novel Adaptive Role-Switching Cooperative Learning (ARSCL), a dual-policy MARL framework that achieves resilient, cooperative coverage through adaptive role specialization. ARSCL assigns complementary roles —a main policy maintaining primary coverage and a compensatory policy introducing diversity to counteract correlated drift—linked by a dynamic role-switching mechanism that adapts to spatial deviation. A drift-prevention module further constrains learning to feasible trajectories. Simulations demonstrate that ARSCL achieves an average coverage duration exceeding 94% within a 25 km radius over a 24-hour period, significantly outperforming state-of-the-art MARL baselines in both efficiency and resilience.

5G base station networks generate numerous alarms daily. With the increasing demand for digital services, it is vital to inspect and rectify anomalies to uphold user satisfaction. This study explores the potential of unmanned aerial vehicle (UAV) empowered opportunistic inspection based on alarm data. We formulate the inspection routing problem as a prioritized traveling salesman problem (PTSP) encompassing two categories of base stations. Priority is assigned to stations generating more alarms, while others are subject to opportunistic inspection. To expedite large-scale opportunistic inspection routes, we introduce a novel transformer-based parallelizable routing algorithm (TPRA). TPRA is an intelligent optimization that orchestrates multiple parallelized constrained reinforcement learning algorithms. Through balancing spectral clustering, the large-scale graph is segmented into manageable subgraphs. For each subgraph, the prioritized inspection routing problem is formulated as a constrained Markov decision process and optimized by transformer-based reinforcement learning in parallel. The optimized subgraphs are then merged using an adaptive large neighborhood search approach. Through parallel computing, our approach achieves as much as 75% reduction in computation time, while concurrently generating shorter routes. The approach is implemented in real-world cases to validate its efficacy. Note to Practitioners—The rapid expansion of 5G infrastructure underscores the critical need for advanced technology and maintenance strategies. Base stations are often placed at high altitudes to ensure line-of-sight connectivity, which poses difficulties for maintenance, particularly in challenging terrains. UAVs offer a promising solution for faster and safer inspection and rectification. The designed approach utilizes reinforcement learning in parallel to optimize UAV inspection routes in an opportunistic manner. This method strategically prioritizes inspection routes based on the real-time base station alarm data, ensuring a swift and effective response to potential issues. Trained in simulated scenarios, the model requires few adjustments for real-world deployment, making it readily implementable in 5G networks. Beyond the potential of the 5G network, the approach also unlocks new value across various types of service in the low-altitude economy.

Inspection and maintenance strategies are crucial for ensuring system serviceability. Due to stochastic deterioration and limited observability, sequential decision-making often relies on established approaches such as Partially Observable Markov Decision Processes (POMDPs). However, such approaches impose constraints on periodic decision epochs and exponentially distributed sojourn times, leading to redundant early-stage inspections and increased overall costs. To address these limitations, we propose a novel Predictive Semi-Markov Decision Process (PSMDP) that leverages Remaining Useful Life (RUL) predictions to reduce unnecessary inspections. Our approach incorporates an online scheme that utilizes inspection data for RUL prediction and proactively schedules subsequent maintenance decisions. When applied to infrastructure management, such as bridge maintenance, our PSMDP demonstrates an average cost reduction of 8.9% across various deterioration scenarios compared to other sequential decision-making models. Note to Practitioners—Effective infrastructure maintenance is essential for ensuring public safety and economic growth. However, limited budgets necessitate more efficient strategies. Current periodic inspection strategies often lead to wasteful expenditures due to unnecessary inspections during the early stages of deterioration. To address this challenge, we propose a predictive and proactive approach that leverages Remaining Useful Life (RUL) estimation to schedule inspections and optimize sequential maintenance decisions. By intelligently planning inspections, practitioners can reduce redundant inspections by 89.5% and save overall costs by 8.9% while retaining the required reliability. This approach facilitates a paradigm shift from condition-based maintenance to predictive maintenance.

In the era of digitalization, ubiquitous sensing technologies have paved the way for predicting the remaining useful life (RUL) of assets or systems. In both practical and theoretical fields, enabled by machine learning algorithms, predictive maintenance (PdM) has attracted significant attention. Among machine learning algorithms, deep learning benefits from its multilayer architecture for performing feature engineering. It provides high-quality results in an efficient manner and has become a prevalent approach. However, only predicting the expected RUL is insufficient. For practically implementing PdM approaches, both the overestimating and underestimating prediction risks should also be analyzed and mitigated before making maintenance decisions. In this article, we propose a deep Gaussian process approach to predict the expected RUL and estimate the associated variance. The approach adopts the multilayer architecture such that the predicted result is robust against the selection of kernel functions. Several novel evaluation metrics are introduced to evaluate the predicted RUL distribution. To realize a complete framework of PdM, enabled by the RUL distribution, we propose a distribution-based cost minimization algorithm to dynamically optimize the predicted maintenance thresholds. The overall approach is tested with two practical datasets.

Aging bridge infrastructure may jeopardize safety, and its economic impact has raised concerns worldwide. Reinforced concrete is a major component of bridge infrastructure, whose condition is directly related to reliability and safety. However, evaluating and maintaining reinforced concrete remains a complex issue, as it is often exposed to a dynamic and stochastic environment. In this article, we provide an integrated approach to capturing the characteristics of reinforced concrete deterioration by synthesizing knowledge from the physical model and a stochastic model. Then, a semi-Markov decision process is constructed for optimizing the maintenance of reinforced concrete. To improve the computation efficiency of the approach, we design the policy iteration algorithm with finite evaluations to reduce computation cost. The designed approach can further calibrate the temporal and complex dynamic information of the reinforced concrete deterioration and has a better performance than that of the Markov decision process. Finally, we explore the sensitivity of the maintenance policy under the changing of cost rates and durations of the maintenance activities.

Complex industrial assets such as power transformers are subject to accelerated deterioration when one of its constituent component malfunctions, affecting the condition of other components, which is a phenomenon called fault propagation. In this paper, we present a novel approach for optimizing condition-based maintenance policies for such assets by modelling their deterioration as a multiple dependent deterioration path process. The aim of the policy is to replace the malfunctioned component and mitigate accelerated deterioration at minimal impact to the business. The maintenance model provides guidance on determining inspection and maintenance strategies to optimize asset availability and operational cost.