Deep reinforcement learning-based optimization of sparse sensor placement for supersonic isolator flowfield monitoring

Neural networks are typically employed to establish a mapping relationship between local observational quantities and the entire flowfield, thereby enabling flowfield prediction. In supersonic inlets, the optimization of sparse sensor placement involves whether the flowfield prediction model (FPM) can accurately perceive the flowfield using the minimal number of sensors. In this study, deep reinforcement learning (DRL) is used to optimize the equidistantly arranged sensors on the wall of a supersonic isolator. A neural network model that takes wall pressure sensors data as input and produces schlieren images of the flowfield as output serves as the FPM, which also functions as the environment in DRL. The agent, based on the Actor-Critic model, explores in this environment with goal of minimizing the number of sensors while maintaining accurate flowfield prediction. After each interaction, the FPM demonstrates excellent performance in the test set with the agent-explored scheme serving as input. Using the effective sensor scheme provided by DRL, among the structure similarity index measure of all samples in the test set between predicted and real flowfield schlieren images, both the mean and median are no <0.75, and the lower bound is no <0.65. These schemes generate a series of flowfield schlieren images with the assistance of the FPM. The images are then processed by the YOLO model, a target detection method, to determing shock train leading edge (STLE) locations. Based on the tracking results of the STLE, a better scheme can be further selected. Through the analysis of sensors data and STLE locations, it can be concluded that the sensors selected by DRL and located near the STLE exhibit broadband and high-frequency oscillation characteristics. This characteristic facilitates the neural networks in establishing a relationship between the sensors and the full flowfield.