Deep-Learning-Based Emergency Stop Prediction for Robotic Lower-Limb Rehabilitation Training Systems

Robotic lower-limb rehabilitation training is a better alternative for the physical training efforts of a therapist due to advantages, such as intensive repetitive motions, economical therapy, and quantitative assessment of the level of motor recovery through the measurement of force and movement patterns. However, in actual robotic rehabilitation training, emergency stops occur frequently to prevent injury to patients. However, frequent stopping is a waste of time and resources of both therapists and patients. Therefore, early detection of emergency stops in real-time is essential to take appropriate actions. In this paper, we propose a novel deep-learning-based technique for detecting emergency stops as early as possible. First, a bidirectional long short-term memory prediction model was trained using only the normal joint data collected from a real robotic training system. Next, a real-time threshold-based algorithm was developed with cumulative error. The experimental results revealed a precision of 0.94, recall of 0.93, and F1 score of 0.93. Additionally, it was observed that the prediction model was robust for variations in measurement noise.