HBTASP: Two-stage temperature-aware multiprocessor scheduling for dynamic DNNs in real-time defect detection
Hui Li, Feixiang Han, Kai Wang, Shanying Zhu · Expert Systems with Applications · 2026
Real-time defect detection is critical for industrial quality control, yet dynamic deep neural networks (DNNs) on heterogeneous multiprocessors face significant scalability and thermal reliability challenges. Existing scheduling methods either assume static task execution times or reactively throttle processors when temperature constraints are violated. We propose HBTASP (Heuristic-Based Temperature-Aware Semi-Partitioned Scheduling), a two-stage temperature-aware scheduling framework that comprises a load-balanced task assignment stage (HBTASP-Assign) and a per-processor co-scheduling stage (HBTASP-Schedule). The co-scheduling stage also includes an auxiliary low-voltage mechanism that executes eligible lower-weight subtasks at a feasible low voltage to reduce modeled processor temperature when the thermal state permits, creating thermal headroom for subsequent higher-weight subtasks subject to deadline and modeled thermal-feasibility tests. Experiments on a real-world steel defect dataset demonstrate that HBTASP maintains low mandatory-region service-failure rates and competitive deadline-aware detection performance in the primary workload grid. Additional tests characterize its behavior under alternative arrival patterns, overload, and RC-parameter perturbations. These results support HBTASP as a practical scheduling option for deadline-constrained automated inspection in manufacturing environments.