Ultra-broadband natural frequency using automatic resonance tuning of energy harvester and deep learning algorithms

• Automatic resonance tuning (ART) harvester of two clamped piezoelectric beams is proposed. • Optimization COMSOL module is employed to optimize the frequency range. • The proposed ART design adjusts its natural frequency in range (27–137 Hz). • A novel CNN is trained and tested with FEM results of the harvester. • Predict harvester outputs not evaluated numerically using FEM. Previous broadband energy harvester techniques met many challenges like output power with a sharp peak, small enhancement in bandwidth, and large dimensions and weights. This paper introduces the Automatic Resonance Tuning (ART) technique of two piezoelectric beams to manage these challenges. The energy harvester of two clamped beams automatically adapts their natural frequencies corresponding to the ambient vibration using (sliding masses over the beams). The optimization using COMSOL was conducted to determine the frequency ranges of the low-frequency beam and high-frequency beam and maximize the output power. The bandwidth of the optimized ART harvester is broadened from 27 to 137 H z , ultra-broad bandwidth ( 110 H z ). Our Finite Element Method (FEM) results were validated with experimental results that exhibited excellent convergence. Usually, the dataset of voltage and power is collected by the FEM. Voltages and power evaluated using FEM for some positions are used as the convolutional neural network (CNN) input. CNN predicts the most of masses' positions over the harvester due to the complexity of repetition implementation FEM in several positions. Then, the CNNs are trained for new wide masses position prediction. The mean square error (MSE) of the training dataset is 2.5601 × 10 - 7 μ w and the performance of the CNN training is 97.62 % accuracy ( P % ), 95.38 % regression rate ( R % ), and 93.78 % F-score ( F % ), at epoch 1000 , which shows the effectiveness of the proposed approach.