The complexity of unmanned aerial vehicle (UAV) systems and the diversity of their fault modes present significant challenges to their reliability, stability, and safety. To address the issue of incomplete fault UAV data samples, a fault simulation dataset was constructed using a predefined fault injection method. This dataset is based on four models of faults: bias faults, drift faults, lock faults, and scale faults, allowing equivalent simulation of the drone in fault-free states, actuator failures, and sensor failures. Furthermore, the dataset was evaluated using deep learning networks. Simulation results demonstrate that the three deep learning architectures—WDCNN, ResNet, and QCNN—validate the completeness and effectiveness of the construction method and the fault simulation dataset in this paper. In terms of precision, WDCNN achieved over 82%, ResNet exceeded 90%, and QCNN surpassed 92%. The methods proposed in this study provides a complete dataset and evaluation method for data-driven research on UAV fault diagnosis.