Unsupervised time–frequency environment perception model for underwater vehicle in irregular ocean

Abstract

Dynamic ocean environments are difficult to monitor and affect the motion balance of underwater vehicles. Existing unsupervised environment perception methods have poor performance and less consideration of environment and vehicle interaction effects. This paper proposes a novel unsupervised time–frequency environment perception model for underwater vehicles in the irregular ocean. First, the observed ocean information is transformed into the time–frequency space based on the wavelet transform. After that, an adaptive dual-threshold diagnosis model is proposed to automatically extract and locate ocean features and depress the background noise. Then, the environment-vehicle interaction model is derived to quantitatively predict the state change of underwater vehicles caused by environmental features. Finally, extensive experiments on the novel buoyancy-driven underwater glider and publicly South China Sea dataset are conducted to verify the effectiveness of the proposed method. The result shows that the diagnosis model can better perceive the ocean environment, and the performance outperforms traditional unsupervised statistical and supervised learning methods. The thermocline detection accuracy is 100%, and the location accuracy is 98.73%. The environment-vehicle interaction model can accurately describe the state changes of underwater vehicles under environmental influence, with an average accuracy of 94.88%.