Abstract: Gait phase prediction holds significant importance in the control of assistive robotic devices, such as exoskeletons. The control unit is required to discern the gait phase to supply the necessary power during operation. Given that current gait phase prediction methods based on the Inertial Measurement Unit (IMU) do not fully leverage the relationship between joints and bones, this study presents a gait phase prediction approach for an exoskeleton robot using a Channel Attention-enhanced Directed Graph Neural Network (CA-DGNN) to enhance prediction accuracy and reliability. Initially, a device for collecting human lower limb posture information is developed to gather walking gait data and construct a skeleton model of the lower limbs. Subsequently, a gait phase prediction model based on CA-DGNN is established to extract motion characteristics of human gait phases and predict the gait phase at a future moment based on current data. Lastly, the impact of the sliding window size on the algorithm's performance is analyzed. The experimental results show that compared to other algorithms, the prediction accuracy of CA-DGNN is 97.88%, which is better than other four algorithms such as CNN, RNN, TCN and LSTM. This work aims to present an innovative idea and method for gait phase prediction in exoskeleton robots, thereby advancing the accuracy and robustness in such robotic systems.