Piezoelectric Energy Harvesting Coupled with Energy-Aware Deep Reinforcement Learning for Extended-Endurance Autonomous UAVs

Abstract

 Battery endurance limits commercial quadcopter UAVs to 15–25 minutes per charge. This paper integrates two complementary approaches to the problem. First, an Euler-Bernoulli beam finite element analysis (FEA) is performed on the DJI F450 arm to characterise piezoelectric PZT-5A energy harvesting across six patch locations and three motor operating conditions. Second, an energy-aware deep reinforcement learning (DRL) navigation framework is developed, comparing DQN, PPO, and SAC across five random seeds over 200,000 training steps with a battery-constrained reward. The FEA, implemented in open-source Python (NumPy/SciPy) and verified analytically to within 0.03%, shows that arm-root placement (P3, 15% span) harvests 0.0600 mW on average and 0.1393 mW at maximum throttle — a factor of 75 improvement over motor-mount placement. A four-arm deployment recovers 144 mJ per 10-minute mission. The DRL framework augments the reward with the FEA-derived harvest function; SAC achieves 82.2±2.7% navigation success with 24.2±1.8% battery use, statistically superior to all baselines (ANOVA F=93.96, p < 0.001). Together, the integrated system powers all proximity sensors from harvest during maximum-throttle climb phases, effectively removing avionics load from the primary battery. All results are confirmed by 43 unit tests (43/43 PASS).