Abstract:

In the process of civil aviation repair and maintenance， a large number of text maintenance records containing rich fault features have been accumulated. However， due to complexity of the maintenance text itself， intelligent diagnosis has not been realized， and the data utilization rate is low. An aircraft fault diagnosis method based on civil aircraft maintenance text data is proposed using the Bidirectional Encoder Representation from Transformers （BERT） based on the pretrained language model and the Light Gradient Boosting Machine （LightGBM）， so as to solve the fault cause in the text maintenance records and then assist maintenance personnel to make correct maintenance decisions. Firstly， the multi-head self-attention mechanism is introduced into the feature extraction architecture of the Bert-based transformer fault diagnosis model to fully capture the bi-directional semantics in the context and pay more attention to key words. Secondly， to improve the diagnosis speed， the parameters of the model are reduced， and the LightGBM model is integrated to achieve fault cause classification of maintenance texts. Finally， the improved model is compared with other commonly used text analysis models. In the fault diagnosis based on civil aircraft maintenance text， the accuracy of the model is 38.99%， 22.98% and 18.16% higher than that of TextCNN model， LSTM model and BiLSTM model respectively， and the diagnostic speed of BERT LightGBM model is 0.91% higher than that of BERT model. It shows that the proposed method is effective and superior in fault diagnosis of aircraft maintenance text.

Key words: aircraft maintenance text, fault diagnosis, BERT, LightGBM, multi-head attention mechanism, parameter optimization