Collision Detection and Body-Part Classification in Collaborative Robots Using a Multi-Task Deep Learning Model With Hybrid Time-Series Augmentation

Ensuring collision safety in human–robot collaboration has emerged as a critical challenge in the Industry 5.0 era. This study proposed an intelligent deep learning framework that can detect robot–human collisions and simultaneously identify the human body part involved using only the robot’s internal sensors (six-axis joint torque sensors and tool center point coordinates) with no external sensing. To address the severe class imbalance in collision data, we introduced a hybrid time-series augmentation approach combining a synthetic minority oversampling technique based on the furthest neighbor algorithm and time-series generative adversarial network. The framework uses a multi-task deep neural network (integrating a temporal convolutional network, a bidirectional long short-term memory, and an attention mechanism) to perform binary collision detection and multi-class collision location classification within a single model. We experimentally validated the system on an industrial collaborative robot (Company Y HC10 model) using a humanoid dummy. Collisions were induced on the dummy’s head, chest, shoulders, and arms at speeds of 100, 200, and 300 mm/s, generating an internal sensor dataset for model training and evaluation. The proposed method achieved 99.5% accuracy and 100% recall in collision detection and 86.2% accuracy with an 88.5% macro F1-score in collision location classification, significantly outperforming existing learning-based collision detection approaches in sensitivity and adding the new capability of collision location identification. These results demonstrate the feasibility of a real-time sensorless safety system, enhancing human-centric robot safety through software alone, without external hardware, in alignment with Industry 5.0 objectives.