Constrained control allocation for dynamic ship positioning using deep neural network

Abstract

Dynamic positioning (DP) is one of the key technologies towards ship autonomy. Ships in DP mode use thruster devices to maintain position and perform low-speed maneuvering. The motion controller issues force requests according to the measured ship state and motion objectives. These requests must be translated into individual thruster commands. Due to constraints in the thrusters, such as inertia and limited angles of operation, state-of-the-art control allocation methods apply constrained optimization techniques. Although such methods readily capture the handling of constraints, they may require significant computational resources in searching for optimized commands in real time. Here we show that a neural network may be applied to offer an effective evaluation of the mapping between motion controller requests and executable thruster commands. An Autoencoder-like neural network is trained with data generated using knowledge about the configuration of the thrusters. Custom loss functions shape the weights of the network, such that the overall motion objectives and thruster constraints are met. Then, the network is applied to perform low-speed maneuvering and stationkeeping in a simulator. Comparison relative to a state-of-the-art variable angle, constrained control allocator indicate similar dynamic performance with reduced peak power consumption.