A robust ultrasonic SLAM algorithm based on multi-constraint graph optimization was proposed to address the issues of positioning drift and poor mapping quality in conventional ultrasonic SLAM systems caused by low sensor resolution in complex environments. To ensure successful line feature extraction, temporal information from consecutive M ultrasonic detections was fused, and a split-merge algorithm was employed to achieve stable line feature extraction. A graph model incorporating odometry motion constraints and ultrasonic observation constraints was constructed, where the weights of each constraint are dynamically adjusted based on sensor uncertainty. The joint optimization problem of pose and features was solved using sparse matrix decomposition and the Levenberg-Marquardt algorithm. Simulation results demonstrate that compared to odometry-based dead reckoning and extended Kalman filter SLAM methods, the average positioning accuracy is improved by approximately 83.8% and 63.5%, respectively, with an average map error of only 1.19 cm and a map coverage rate reaching 96%. An experimental platform was established to further validate the algorithm in real-world environments. This work provides a novel approach for low-cost ultrasonic SLAM systems, suitable for resource-constrained devices or extreme scenarios such as smoke-filled environments.