In unmanned navigation of agricultural machinery, it’s crucial to make accurate positioning. Although Real-time Kinematic Global Positioning System (RTK-GPS) and other accurate positioning technologies had made significant progress, when the unmanned vehicles were in the region with poor satellite signal, such as near buildings, woods, etc., its positioning became poor. To solve this problem, this study proposed an advanced Kalman Filter algorithm based on RTK-GPS, Inertial measurement unit (IMU), and encoder for multi-sensor fusion which achieved position prediction under the condition of poor satellite signals. An unsupervised Auto-Encoder model based on convolutional neural networks was proposed in this study. Auto-Eecoder and Auto-Dncoder had asymmetrical structures with three-layer convolution to generate the denoised data to optimize state equation and measurement equation noise covariance Q and R of Kalman Filter. Using the optimized Q and R matrices, the state equation and measurement equation were constructed. The Nvidia Jetson Nano development board was used as the experimental platform to test in open-air and shade-covered environments. The results showed that the proposed algorithm had better positioning accuracy than RTK-GPS in the two environments and could still accurately predict the vehicle position in real-time with poor signals.