Hydrodynamic detection based on multilayer perceptron and optimization using dynamic mode decomposition

Fish and pinnipeds have the ability to sense flow changes through lateral lines or whiskers, so as to find prey or partners, and avoid natural enemies or obstacles. This provides an idea for the possibility of hydrodynamic detection. Due to the complex nonlinearity between the shape and motion parameters of the target and its induced flow fields, it is a new trend to construct the relationship between the wake characteristics and the target using machine learning. In this work, a Multi-Layer Perceptron (MLP) neural network is constructed based on the wake flow field around a cylinder to realize the hydrodynamic detection ability. Specifically, the instantaneous wake flow field is used to reduce the redundancy of the data and improve the usability of the method. The predictions of the cylinder position and the Reynolds number are realized. The influences of the interval of measuring points, the number of time slices and the number of hidden layers on the prediction error are studied. By adjusting the sampling method and the network structure, the detection accuracy is optimized. A multi-data fusion method combing MLP and DMD (Dynamic Mode Decomposition) is proposed at last, which is able to further improve the detection accuracy. The error limits of the predicted Reynolds number and the position using this approach are effectively reduced. This technique provides an alternative for target detection by using the flow field.