Traditional methods for extracting gear fault features require manual setting of extraction rules, relying heavily on experience. Moreover, they have poor adaptability to variable working conditions. On the other hand, traditional neural networks are unable to effectively capture temporal correlations and fail to extract features adequately. Aiming at the above problems, this paper uses a method based on recurrent units and self-attention mechanism to conduct fault diagnosis on the central transmission gear of a turboshaft engine. This method first extracts global feature information from the original vibration signal data through a bidirectional gated recurrent unit (BiGRU) and conducts learning and training in both forward and backward directions to more fully capture the temporal relationships in the signals. Then, a self-attention mechanism is introduced to preferentially select the main feature information and discard the irrelevant feature information to improve efficiency. Finally, fault classification is carried out through the Softmax function. Experimental results show that the diagnostic accuracy of this method can reach more than 97%, and its diagnostic effect is better than models such as RNN, LSTM, GRU, and BiGRU.