From Behavior to Natural Language: Generative Approach for Unmanned Aerial Vehicle Intent Recognition

This article introduces a novel cross-modal neural network model that aims to convert long-term temporal behavior data into natural language to achieve unmanned aerial vehicle (UAV) intent recognition. Our generative intent recognition model effectively utilizes the inherent redundancy present in long temporal behavior data by incorporating a sequence compression module, which enables the cross-modal generation and alignment of intents while preserving the integrity of the standard Transformer architecture. Importantly, we observe that this approach mitigates the negative impact of imbalanced database distribution by mapping intent categories onto the modality of natural language. Furthermore, we propose three comprehensive pretraining tasks specifically designed for time series data, thoroughly examining their interconnections and analyzing the impact of a hybrid pretraining framework on the accuracy of intent recognition. Our experimental results demonstrate the superiority of our proposed generative UAV intent recognition model, along with the hybrid pretraining initialization method, compared to conventional classification models. Simultaneously, the intent recognition method exhibits heightened temporal sensitivity and robust resilience, enabling it to deal with complex UAV confrontation and interference environment.