Feature-Efficient LSTM for Underwater Target Intention Recognition

Due to the unique characteristics of underwater environments, intention recognition of underwater targets requires high-performance algorithms. A LSTM network incorporating the Adaptive Attention Mechanism is designed in the paper to integrate features and enhance network performance. Furthermore, Meta-action combination feature with comprehensive spatial information consideration is developed, rendering the model more suitable for intention recognition in underwater environments. The ablation experiment confirms that the Enhanced LSTM network improves the focusing ability and adaptability of the model, comprehensively considers the spatial distribution of underwater scenes, and improves the accuracy and robustness of intention recognition. Additionally, comparative analyses with SVM demonstrate that the network developed in this paper is adeptly suited for the time-series-based recognition of underwater target intentions.