Learning-based adaptive neural control for safer navigation of unmanned surface vehicle with variable mass

This paper presents a novel approach to the precise control of variable mass unmanned surface vehicles (USVs) during payload deployment, where both mass and draught undergo unpredictable changes. We propose a draught observation method and an adaptive control strategy that leverages the strong coupling between the USV's motion states, mass, and draught. Our method employs a radial basis function neural network (RBF-NN) for real-time draught observation, using an offline training strategy based on gradient descent, combined with an adaptive online training strategy to improve observation accuracy. An adaptive control strategy based on the Backstepping method is then developed, incorporating real-time draught data from the RBF-NN to address unknown variations in mass and draught. The stability of both the RBF-NN observer and the adaptive control algorithm is rigorously verified using the Lyapunov method. Simulation results demonstrate that the proposed draught observation method achieves up to 30% faster convergence compared to traditional methods, with a significant improvement in observation accuracy. Furthermore, the adaptive control strategy effectively manages real-time adjustments in dynamic scenarios, maintaining robust control performance even under significant mass changes, where conventional approaches fail.