The proliferation of Internet of Things (IoT) applications has significantly increased edge devices and data volume at the edge of network, making them ideal candidates for federated edge learning (FEEL). However, the limited spectral resources of edge base stations (BSs) restrict device participation, exacerbating the impact of data heterogeneity, negatively affecting model convergence, even leading to model drift. To mitigate the impact of BS access restrictions on the performance of FEEL, this paper introduces a novel FEEL architecture that is empowered by a mode of direct inter-device communication, Device-to-Device (D2D) communication. Considering the negative influence of data heterogeneity and the potential packet error rate (PER) under this architecture, a nonlinear integer programming problem is formulated. Subsequently, we proposed an elegant algorithm termed FedD2D that jointly optimizes device pairing and scheduling in D2D communications while incorporating fairness constraints to counteract the negative consequences of above factors. Ultimately, the experimental results demonstrate the superiority of proposed, specifically reflected in the increased device participation and the reduction of the impact of communication unreliability, thereby enhancing the performance of FEEL.