Deep learning-based flow sensing and control for autonomous underwater vehicles

Abstract

This study developed a deep learning-based object identification (OI) model and control (PPO-based AFC) system to improve the navigation system function of autonomous underwater vehicles (AUV). A two-dimensional numerical simulation simplifying AUV and the underwater environment was used to develop a deep learning-based OI model and PPO-based AFC system suitable for AUV with various underwater motions. During the simulation, sensors arranged side by side on NAC Afoil, which simplified AUV, moved in a flow field with foil to create a time series data set, and performed deep learning using this data set containing information on the flow field. Research on the development of OI models was conducted with the inspiration of the function of lateral lines that enable hydrasonic imaging through the surrounding flow information of various aquatic animals, and this OI model was intended to contribute to the performance improvement of AUV's vision system. We trained a deep-learning neural network to develop an OI model that predicts the location of foils and objects based on two types of sensory data, such as velocity and pressure obtained through the sensor array located on the foil surface. The predictive performance of the OI model derived using the LSTM algorithm showed better predictive performance than that of the OI model derived through other neural networks (ANN, DNN). The PPO-based AFC system was studied with the inspiration of the optimal propulsion seen in the flight and swimming of birds and aquatic animals, and this PPO-based AFC system was used to improve the propulsion efficiency of AUV. To this end, we placed a flow actuator on the foil surface and trained an agent that controls the jet sprayed from the actuator using the PPO algorithm. The PPO neural network was improved to control several actuators individually delicately, and the PPO-based AFC system derived through reinforcement learning improved the propulsion efficiency of the foil. In addition, to increase the practicality of the OI model and PPO-based AFC system, a sensor importance test study was conducted to reduce the number of sensors applied to the foil. Through the sensor importance test, it was possible to derive the minimum number of sensors capable of hydraulic imaging by removing sensors with high redundancy and selecting critical sensors. The sensor importance test used feature selection methods like LASSO, Elastic Net, and RSR. By removing duplicate sensors, the OI model could predict the exact location of the foil and the object, and the PPO-based AFC system was also able to maintain the same high propulsion efficiency as before. Through this study, the practicality of the OI model and PPO-based AFC system could be increased, and the foundation for developing underwater navigation technology for sensor-based AUVs could be laid.