Odometry is a crucial component of Simultaneous Localization and Mapping (SLAM) technology. However, most existing odometry algorithms rely on single data sources such as point cloud data or image data, failing to fully leverage multi-data fusion to improve trajectory estimation accuracy. Additionally, these algorithms often exhibit insufficient accuracy in complex environments and feature-deficient scenarios. To address these issues, this paper proposes a deep network called 3DPointLIO, which fuses LiDAR data and Inertial Measurement Unit (IMU) data. Firstly, a feature pyramid network combined with a weight attention mechanism is introduced to reduce the impact of dynamic information in the environment and enhance the robustness of point cloud features. Secondly, in the IMU data processing network, a convolutional network is integrated with Gated Recurrent Units (GRUs) to mitigate noise in raw IMU data, and a Bidirectional Long Short-Term Memory (BiLSTM) network is employed to extract temporal features from the denoised IMU data. Finally, a pose estimation network composed of multiple linear layers is used to estimate translation and rotation. The proposed algorithm is validated on the open-source KITTI dataset. Experimental results demonstrate that, compared to the baseline model, the proposed odometry algorithm improves rotation estimation by 0.76° and translation estimation by 2.17%. Furthermore, it outperforms other common odometry algorithms in both rotation and translation estimation, particularly achieving higher accuracy in rotation estimation.