Deep Temporal Convolutional Network for Multisensory Fusion With Imbalanced Data of Flight Regime Recognition

Regime recognition (RR) is critical to the establishment of component fatigue lives for an aircraft. However, it is still challenging to recognize flight regimes accurately and precisely from multisensory signals and an imbalanced number of regimes. In this article, a deep temporal convolutional network (DTCN) with a dynamic sampling strategy is proposed to address these challenges. To extract the sequential patterns of multisensory signals, the temporal convolution is developed to model the sequential data. The dilated convolution operation can mine sequential features from long sequential data, which contributes to the detection of sequential patterns and an improvement in accuracy for RR. Since the numbers of regime samples are imbalanced, a dynamic sampling strategy is combined with DTCN, which would increase the ability to recognize the small number of regimes. The proposed DTCN with dynamic sampling is validated by using the flight parameters collected from the health and usage monitoring system of helicopters, and current results show its potential for RR in practice.