Due to the strong nonlinearity of material used in soft gripper, it is difficult to establish a precise mechanism model for measuring the bending angle of the soft gripper. To address this challenge, a mechanism and data-driven soft sensor model for the bending angle of the soft gripper is proposed. The model consists of a mechanism model and an adaptive block increment stochastic configuration networks (ABSCN) compensation model, which includes information scent evaporation and inertia weights. The mechanism model parameter is identified using least squares, and ABSCN is used to predict compensation for high order unmodeled dynamics. By adaptively optimizing the number of incremental block configurations in block incremental stochastic configuration networks (BSC), the compactness of the model is improved, and the training time is reduced. Finally, through simulation experiments and comparison with real data using a hybrid model, it is shown that the proposed method significantly improves the accuracy.