Monitoring propeller failures is vital to maintain the safe and reliable operation of quadrotor UAVs. The simulation-to-reality UAV fault diagnosis technique offer a secure and economical approach to identify faults in propellers. However, classifiers trained with simulated data perform poorly in real flights due to the wind disturbance in outdoor scenarios. In this work, we propose an uncertainty-based fault classifier (UFC) to address the challenge of sim-to-real UAV fault diagnosis in windy scenarios. It uses the ensemble of difference-based deep convolutional neural networks (EDDCNN) to reduce model variance and bias. Moreover, it employs an uncertainty-based decision framework to filter out uncertain predictions. Experimental results demonstrate that the UFC can achieve 100% fault-diagnosis accuracy with a data usage rate of 33.6% in the windy outdoor scenario.