Physics-Guided Generative Adversarial Networks for fault detection of underwater thruster

In this paper, we propose a novel hybrid framework for fault detection in underwater thrusters based on the combination of a physical model and a Generative Adversarial Network (GAN). The proposed framework allows incorporating physical information within the architecture and training process of GAN, and relying only on a small dataset without fault samples to complete the training. Firstly, a Variational Autoencoder (VAE) is used to extract the informative distribution from the voltage signal, and this distribution is introduced into the generator of the GAN. Then, the thruster torque estimated by an Extended State Observer (ESO) is used as the real sequence of values for the discriminator of the GAN. Furthermore, a physical loss is aided for the training of the generator to improve the convergence of the GAN. The proposed hybrid framework is validated on a real dataset with multiple faults. The experimental results show that the proposed framework allows accurately detecting the different degrees of propeller damage faults and biofouling faults in underwater thrusters.