Aircraft EWIS Safety Risk Topic Recognition Method Based on Enhanced BERT-LDA

To identify critical topics of safety risk in the aircraft electrical wiring interconnection system (EWIS), this article proposes an unsupervised text topic recognition method, leveraging the event synopses from the aviation safety reporting system. The approach involves utilizing latent Dirichlet allocation (LDA) for constructing topic feature vectors, coupled with the bidirectional encoder representations from transformers (BERT) to generate semantic feature vectors, thereby obtaining EWIS fused feature vectors. Addressing challenges, such as information loss during transfer, random clustering center determination, and disunity of optimization objectives during the feature extraction and cluster assignment processes of the fused vectors, an enhanced BERT-LDA model based on the deep embedded clustering (DEC) algorithm is proposed. The model incorporates clustering layers after the encoders using DEC to mitigate information loss and explores optimal clustering center determination through Kullback–Leibler divergence parameter adjustment. It also involves iterative training of the encoders and clustering layers through stochastic gradient descent to harmonize the optimization objectives of feature extraction and clustering. Comparative studies are conducted to demonstrate the superiority of the proposed method, evaluating various word vector embedding methods, state-of-the-art methods, BERT, and variant models, before and after DEC improvement. The results indicate that, after applying DEC, the performance of evaluation metrics for all models improved significantly. Compared with other methods, the enhanced BERT-LDA methods exhibit superior compactness within the same topic cluster and greater separation among different topic clusters in EWIS safety risk topic recognition.