<p>In recent years, a large number of machine learning-based intrusion detection (ID) schemes for unmanned aerial vehicles (UAVs) networks have been proposed. However, most of existing works rely on either simulated datasets or general-purpose datasets that are not relevant to UAVs, while actual UAV intrusion detection (UAV-ID) datasets collected by real UAVs are very limited, thus, they have received little research attention. Based on a recently released actual UAV-ID dataset (including denial-of-service, replay, evil twin, and false data injection attacks), this work presents the first attempt of combining principal component analysis (PCA) feature extraction and ensemble learning methods for UAV intrusion detection systems. In particular, the PCA is employed to extract important features from the original data, which are then fed into the proposed ensemble learning model. This model consists of four different base machine learning models, namely Decision Tree, Random Forest, Gradient Boosting and eXtreme Gradient Boosting, whose outputs are fed into a meta-classifier called Multi-Layer Perceptron to make a final decision on the attack type. Thanks to this design, our proposal leverages the advantages from different base models as well as helpful features extracted by PCA to improve the detection accuracy. Experimental results show that the proposed method achieves superior performance over baselines, including existing feature selection schemes and base models without ensembling in terms of key performance metrics such as Recall, Precision, Accuracy and F1-score. Our method also outperforms baselines in both multi-class and binary classification tasks. Last but not least, our results reveal that the PCA feature extraction is more effective than its feature selection counterparts when integrated with the proposed ensemble learning model.</p>