Aiming to address the problem of severe inter-dimensional coupling interference of three-dimensional force sensors， a decoupling algorithm based on the sparrow search algorithm–extreme learning machine （SSA–ELM） is proposed considering a fiber Bragg grating （FBG） three-dimensional force sensor with a double-layer cross beam structure as the research object. First， the sensing and force measurement principle of FBGs is studied， the mapping relationship between wavelength drift and the force of the three-dimensional force sensor is revealed， and its structural coupling characteristics are analyzed. Then， a calibration experiment system is constructed to perform calibration experiments. Finally， a nonlinear decoupling model of extreme learning machine （ELM） is established， and the sparrow search algorithm （SSA） is used to optimize the model to obtain the optimal initial weight and threshold of the network. Considering the decoupling accuracy and efficiency， the optimal number of ELM hidden layer nodes and optimal number of SSA iterations are determined. After decoupling， the maximum type I error is 1.18%， the maximum type II error is 1.14%， and the decoupling training time is 1.778 6 s. At the same time， in order to verify the decoupling effect， the decoupling results of the SSA–ELM algorithm are compared with least squares and ELM algorithm. The experimental results show that the SSA–ELM algorithm has a short decoupling training time， can more effectively construct the dimensional coupling relationship of the three-dimensional force， reduce the type I and II errors of the sensors， and has a good nonlinear decoupling ability.