Abstract: P- and S-wavefield separation is a key step in multicomponent VSP data processing. Conventionally, this is achieved through signal analysis or polarization projection to decompose different wave types. However, factors such as strong vertical velocity variations and borehole deviation significantly increase the complexity of wavefield separation. To address this issue, a deep neural network-based method is proposed. This method integrates elastic wave theory with Helmholtz decomposition to construct P- and S-wave labels for multicomponent VSP data under complex acquisition conditions. A U-Net architecture is then leveraged to extract P- and S-wavefield features in 2D VSP data and consequently achieve model-independent wavefield separation. Experimental results from synthetic and field data show that the deep learning method effectively separates P- from S-waves in 2D VSP data when using training samples constructed with target features. Furthermore, by analyzing the impact of factors such as offset and deviation on network performance, targeted strategies are proposed to improve the network's generalization ability. These findings provide effective support for P- and S-wave separation in multicomponent VSP data.