Aircraft onboard systems typically have limited labeled fault samples and large amounts of unlabeled data. To better utilize the information contained in limited labeled fault samples, a deep learning-based semi-supervised fault detection method is proposed, which leverages a small number of labeled fault samples to enhance its performance. A novel sample pairing strategy is introduced to improve algorithm performance by iteratively utilizing fault samples. A comprehensive loss function is employed to accurately reconstruct normal samples and effectively separate fault samples. The results of a case study using real data from a commercial aircraft fleet demonstrate the superiority of the proposed method over existing techniques, with improvements of approximately 16.7% in AP, 9.5% in AUC, and 19.2% in F1 score. Ablation studies confirm that performance can be further improved by incorporating additional labeled fault samples during training. Furthermore, the algorithm demonstrates good generalization ability.