Federated edge learning (FEEL) is a promising distributed learning technique\nfor next-generation wireless networks. FEEL preserves the user's privacy,\nreduces the communication costs, and exploits the unprecedented capabilities of\nedge devices to train a shared global model by leveraging a massive amount of\ndata generated at the network edge. However, FEEL might significantly shorten\nenergy-constrained participating devices' lifetime due to the power consumed\nduring the model training round. This paper proposes a novel approach that\nendeavors to minimize computation and communication energy consumption during\nFEEL rounds to address this issue. First, we introduce a modified local\ntraining algorithm that intelligently selects only the samples that enhance the\nmodel's quality based on a predetermined threshold probability. Then, the\nproblem is formulated as joint energy minimization and resource allocation\noptimization problem to obtain the optimal local computation time and the\noptimal transmission time that minimize the total energy consumption\nconsidering the worker's energy budget, available bandwidth, channel states,\nbeamforming, and local CPU speed. After that, we introduce a tractable solution\nto the formulated problem that ensures the robustness of FEEL. Our simulation\nresults show that our solution substantially outperforms the baseline FEEL\nalgorithm as it reduces the local consumed energy by up to 79%.\n