Multi-Head Attention-Based Hybrid Deep Neural Network for Aeroengine Risk Assessment

Abstract

Existing deep-learning models have limited applicability to aeroengine risk assessment owing to insufficient feature extraction capabilities and low robustness. This paper presents a hybrid deep neural network based on a Time2Vec time-embedding layer and multi-head attention mechanism for the proactive risk assessment of aeroengines. The proposed model uses quick access recorder data as input to identify risks associated with different types of failures and outputs two sets of labels: risk level and risk cause. The base of the proposed model combines a convolutional neural network and bidirectional long short-term memory, which are used to automatically extract temporal and spatial features from the input data to represent the system state and capturing irregular temporal trends. The Time2Vec layer facilitates automated processing of sequential data to make it easier for these deep-learning models to recognize patterns in the dataset. The multi-head attention mechanism further enhances the ability of the proposed model to capture and allocate information weights effectively. In comparative experiments, five benchmark models were compared with the proposed model, which demonstrated the best classification accuracy and computational efficiency as well as the most robustness against imbalanced data samples.