In general, training fault samples of aero-engines are very rare and only collected under one or a few operating conditions. However, due to diverse operating conditions and fault severities, testing fault samples may have higher diversity and larger distribution region than training fault samples, leading to a high missing detection rate. To address this issue, a hyperplane-oriented over-sampling technique (HOOST) is developed to synthesize training fault samples with higher diversity and larger distribution region. To be specific, HOOST not only adopts the interpolation strategy to synthesize in-distribution samples, but also adopts the extrapolation strategy to synthesize out-of-distribution samples. Moreover, the sampling factors in the extrapolation strategy are automatically set under the guidance of the initial classification hyperplane, rather than randomly selected, in order to improve the reasonableness of synthetic samples. Finally, the developed HOOST is integrated with a self-attention encoder-decoder. Extensive experiments are conducted, which not only validate the performance of the developed HOOST on an actual aero-engine dataset, but also demonstrate its generalizability on three other public industrial datasets.