This letter presents a radial basis function neural network (RBF NN) based methodology to investigate the dynamics modeling and fault detection (FD) problems for soft robots. Finite element method (FEM) is first used to derive a mathematical model to describe the dynamics of a soft trunk robot. An adaptive dynamics modeling approach is then designed based on this FEM model by incorporating model-reduction and RBF NN techniques. This approach is capable of achieving accurate identification of the soft robot’s highly-nonlinear dynamics, with the identified knowledge being obtained and stored in constant RBF NN models. Finally, a model-based FD scheme is proposed with the modeling results, which can achieve efficient FD for the soft robot whenever it encounters an unknown fault. Note that the proposed methods are generic and usable for general soft robots. Validation of these methods is performed through both computer simulation and physical experiments.