Abstract: Traditional ground contact detection methods are often affected by sensor noise and dynamic complexity in unstructured terrains. Thus, they rely heavily on foot force sensors, and this over-reliance restricts their applicability to certain robot scenarios. To address these limitations, we propose a supervised learning based ground contact event detection algorithm for bipedal robots. Further, we design an algorithm to assist the invariant extended Kalman filter in obtaining accurate, inexpensive ground contact signals, enabling precise state estimation. First, we establish a state machine model for ground contact events and analyze the contact event types and their transitions. Second, we collect sensor data from the robot, including the inertial measurement unit, encoders, current sensors, foot force sensors, and kinematic-derived foot height and vertical velocity (z-axis). We use mutual information for feature selection, thereby retaining 16-dimensional features. Finally, we sample historical features and use them to construct graph embeddings via a clustering method. These graph embeddings are subsequently fed into a one-dimensional convolutional neural network to extract temporal information. This enables the regression analysis of foot contact forces and the classification of ground contact events. The proposed method is experimentally compared with current-based detection and neural network-based classification approaches. Results reveal that the proposed method performs excellently during flat-ground walking and demonstrates strong generalization and robustness, significantly outperforming traditional methods in complex terrains. Overall, this method overcomes the reliance of conventional ground contact detection approaches on foot force sensors and provides an efficient and inexpensive solution for robot ground contact detection.