Dynamic Situational Awareness Discovery for Goal-Driven Operation of Resource-Constrained Connected Systems
Abstract
The rapid evolution of wireless communication technologies and their recent convergence with artificial intelligence have empowered diverse Industrial Internet of Things (IIoT) systems, necessitating dynamic device collaboration for realizing diverse operational goals that go beyond the conventional focus on data rates. However, with the growing system scale, dynamic network conditions, and constrained resources, it becomes extremely challenging to optimize decision-making in IIoT systems. Driven by application goals in highly dynamic and complex environments, future IIoT systems are expected to rapidly and efficiently discover situational awareness, particularly location, time, network topology, and available resources to optimize adaptive decision-making. Motivated by these, this thesis designs top-down discovery mechanisms of situational awareness in future dynamic IIoT systems, i.e., rapid discovery of spatio-temporal situational awareness, energy-efficient network topology virtualization and adaptation, and dependency-aware resource allocation under dynamic conditions.
Firstly, to achieve rapid situational discovery in the spatio-temporal domain for IIoT, a distributed, optimal-measurement-geometry-directed integrated localization and synchronization (ILAS) framework for large-scale dynamic connected systems is proposed. Specifically, this framework comprises a collaborating node selection algorithm, where collaborating nodes forming the optimal measurement geometry are selected to achieve the best estimation accuracy with controlled complexity, and a sequential-state-stacking belief propagation (3SBP) implementation algorithm, where computational complexity is further reduced by limiting matrix inversions and square roots to only a subset of the overall stacked states. Simulation results demonstrate a significant enhancement in the robustness of the ILAS estimation and a reduction in computational complexity compared to baseline schemes.
Secondly, to simplify the discovery of situational awareness under dynamic conditions, a virtual network topology (VNT) construction and adaptation framework is proposed. The proposed VNT construction transforms complex, time-varying physical parameters into stable performance indicators, thus reducing the dimensionality of network parameters while preserving their impact on system operation. Based on this simplified network monitoring approach, we develop a VNT adaptation algorithm for rapid optimal network operations to maximize the effectiveness of data collection, where effectiveness is defined as the sum of collected data weighted by spatio-temporal priority. Numerical simulations and benchmark comparisons demonstrate a significant performance improvement in maximizing data collection effectiveness using the proposed framework.
Thirdly, we investigate the use of situational awareness in improving network topology management and resource allocation. This study is applied in the scenario of task offloading for integrated air-ground networks, where we propose a dependency-aware strategy. Specifically, full tasks demanded by various applications define dependencies among dynamic nodes, which are modeled as dynamic directed acyclic graphs. We define the goal of task offloading as maximizing the average success rate of completing full tasks and formulate topology management, bandwidth, and computing resources allocation as a distributed optimization problem. A multi-agent deep reinforcement learning (MADRL)-based algorithm is developed to solve this problem, where agent-information embedding is introduced to enhance learning performance. Numerical results demonstrate that the proposed strategy significantly improves the average full task success rate compared to baseline approaches in complex and dynamic environments.