Optimized Physics-Informed Neural Network with Deep Learning-Driven Attention Mechanisms for Efficient Task Offloading and Resource Management in Edge–Cloud Computing

As edge-cloud-based Mobile Crowdsourcing systems swiftly expand, the issues of effective offloading and resource allocation have become essential for reducing energy usage and delay time, and for meeting strict deadline requirements. Nevertheless, current methods, such as deep reinforcement learning and heuristic optimization, have severe drawbacks, including high convergence instability, poor deadline awareness, and high energy consumption, leading to increased deadline miss rates under heavy workloads. To address them, this study proposes a new Physics-Informed Focal Self-Attention Neural Network-Bermuda Triangle Optimization Framework (PIFSANN-BTOF). The energy-delay cost model is developed to capture the trade-off between local and edge execution. Supervised learning of spatial crowdsourcing data is pre-processed and normalized across 500 workers and 1500 tasks in the real world. The proposed PIFSANN combines physics-informed constraint learning, focal self-attention to simulate task urgency and global load, and mutual learning to improve convergence stability. Priority scheduling on the stack and Bermuda Triangle Optimization ensure efficient resource matching. Experimental performance is better, showing the best final rate of 0.92 of a maximum task completion, lowest energy consumed of 520 Joules with a reduction of 52% over mobile-only execution, average delay of 3.1 seconds, deadline non-adherence of 0.04, and convergence with a final total loss of 0.24, demonstrating better performance than current state-of-the-art methods.