A Novel Acoustic Weighing Method Based on Deep Learning Model With Joint Local and Global Information

Acoustic weighing expands the possibilities of contactless mass measurement, avoiding contact contamination and loss, especially for tiny objects. Current acoustic weighing models focus on the mechanism analysis of the relationship between oscillation frequency and mass. However, the nonlinear effects of the acoustic field cause significant deviations in the oscillatory behavior of suspended objects, resulting in difficulty in accurately calibrating the frequency parameters, which poses challenges for accurate acoustic weighing. To address these challenges, this article proposes an acoustic weighing method based on a deep learning model with joint local and global information named LGNet. The oscillating trajectories of an object in an acoustic force field encode information about wave–matter interactions and can be used to characterize the object. The designed LGNet model can directly establish the connection between the oscillating signals and the mass of the object, which helps to construct a novel high-accuracy acoustic weighing method. The experimental results show that the designed network model has high accuracy, and the root-mean-square error (RMSE) and mean absolute error (MAE) of this network are 0.195 and 0.129 mg, respectively. Moreover, the proposed method is readily practical for measuring tiny objects, further promoting the horizon of nonlinear acoustic systems in research and applications.