LSTM 신경망을 활용한 하지 작업동작 시퀀스 예측 방안

다국어 초록 (Multilingual Abstract)

Objective: This study investigates the possibility of utilizing deep learning algorithm to predict motion intention of workers based on joint kinematic data. It is to secure motion intention-response simultaneity for efficient and safe worker-exoskeleton robot interaction. Background: Industrial exoskeleton robots are emerging as a solution to reduce work-related musculoskeletal disorders. The main key for successful development of an active exoskeleton robot is to understand and respond accurately to the wearers motion intention. The control system of a wearable robot needs to recognize in advance the exact moment to start activating its joints. Method: Sequential data for 6 representative motion types (walking, left lunge, right lunge, stoop, squat, Asian squat) of the lower extremity joints were collected using inertial measurement unit (IMU) sensors. A Long Short-Term Memory (LSTM)-based model, which is one of the representative deep learning techniques, was designed to train the data. The accuracy and speed of the model for predicting the motion intention of the subjects were analyzed. Results: The classification model showed 86% of accuracy in average which is satisfactory considering its small size of training data set. Using only the initial data from one motion cycle, it was confirmed that the model can predict which type of motion is in progress 75 to 100ms earlier than follow-up movement of the successive joint. Conclusion: This study confirmed the possibility of using artificial intelligence technique in predicting motion intention of workers based on earlier joint kinematic data. It is expected to develop a more sophisticate prediction model in the further study based on multimodal data sets gathered from various sensors such as EMG and Foot Pressure Sensor. Application: The prediction system for motion intention of workers based on deep learning algorithm could be a solution to secure the simultaneity in human-robot interaction.

목차 (Table of Contents)

• ABSTRACT
• 1. Introduction
• 2. Method
• 3. Results
• 4. Conclusion
• References