應用慣性感測器並以類神經模型預測人形機器人跌倒之研究

A research on applying IMUs to predict fall of a humanoid robot based on the Neural Network model

洪政群(HUNG CHENG-CHUN)

指導教授 ： 陳俊達

國立臺灣師範大學/科技與工程學院/機電工程學系/碩士(2019年)

https://doi.org/10.6345/NTNU201901052

摘要

在近年來老年化問題趨於嚴重的社會當中，但一個完整的療程，並不只是在進行完手術後即結束，是包括後續的運動以及追蹤和觀察，因此需要患者在家中復健時的密切配合，可是很多患者會因為各種因素，使得居家運動不夠確實，無法達到應有的效果，所以一套使患者與醫師之間能互相溝通，並能記錄患者資訊的居家機制是非常重要的。研究之系統主要利用六軸慣性感測器來進行膝關節角度量測將資訊傳輸至電腦，且運用網際網路把資料存入後端資料庫當中，並將此系統運用在機器人步行軌跡，模擬醫院之醫師進行觀看，即時掌握和監控每位患者的復健情形，並給予患者回饋，以機器人取代患者的方式與醫師之間達到相輔相成的效果。
本研究提出一個基於穿戴六軸慣性感測裝置的跌倒預測方法，相對於跌倒偵測是發生跌倒後的訊號，跌倒預測是藉由量測跌倒前的訊號且在尚未發生著地前的訊號，來避免跌倒的發生。由於跌倒是一個連續時間的動作，因此可視為是一個連續性動作的分類問題，在本研究中，我們利用類神經網路來做此分析完成跌倒預測，而我們主要利用Bioloid機器人來模擬跌倒動作的發生以避免用人體來實驗的不客觀性，因向前跌倒很容易會造成嚴重的傷害，所以本文以向前跌倒來作為我們的研究，在實驗部分，為了完成慣性感測器的資料讀取，我們以arduino為開發環境來做並把讀出的數值傳給python來做類神經的實驗，此外我們建立的類神經網路有六成以上的準確率，除了可以達到不錯的結果外，也確實可以提早完成跌倒預測的效果。

關鍵字

膝關節角度量測 ； 類神經網路

並列摘要

In a society where ageing problems have become more serious in recent years, a complete course of treatment is not just the end of surgery, it includes follow-up exercises as well as tracking and observation, so it requires close attention when patients are rehabilitated at home. Cooperate, but many patients will not be able to achieve the desired effect because of various factors, so it is very important to have a home mechanism that allows patients and physicians to communicate with each other and record patient information. The research system mainly uses a six-axis inertial sensor to measure the knee angle and transmit the information to the computer, and uses the Internet to store the data in the back-end database, and applies the system to the robot walking trajectory to simulate The doctors of the hospital watched, instantly grasped and monitored the rehabilitation situation of each patient, and gave the patients feedback, and the robots replaced the patients to achieve complementary effects with the doctors.
This study proposes a fall prediction method based on a six-axis inertial sensing device. The fall detection is a signal after a fall, and the fall prediction is a signal before the fall and the signal before the ground has occurred. To avoid the occurrence of falls. Since the fall is a continuous time action, it can be regarded as a classification problem of continuous action. In this study, we use the neural network to do this analysis to complete the fall prediction, and we mainly use the Bioloid robot to simulate the fall action. The occurrence of this is to avoid the objectivity of experimenting with the human body. Because falling forward is easy to cause serious injury, this article takes the forward fall as our research. In the experimental part, in order to complete the reading of the inertial sensor. Take, we use arduino as the development environment to pass the read value to python for the nerve-like experiment. In addition, we have established a neural network with more than 60% accuracy, in addition to good results. It is also true that the effect of the fall prediction can be completed early.

並列關鍵字

knee angle measurement ； G-sensor ； neutral network ； tensorflow

參考文獻

[1] “內政部統計部,”取自:https://www.moi.gov.tw/stat/chart.aspx?ChartID=S0401.

[2] 劉育佑, “以感測裝置追蹤膝關節復健情形的有效方法,” 2016.

[3] 鄭宇倫, 基於穿戴式三軸加速度感測器的步態週期分割與特徵分析, 2016.

[4] 黃榮興, 應用步行演算法則於Bioloid機器人動態行為之實現, 2010年6月.

[5] Saba Bakhshi ,Mohammad H. Mahoor,Bradley S.Davidson, “DEVELOPMENT OF A BODY JOINT ANGLE MEASUREMENT SYSTEM,” 33rd Annual International Conference of the IEEE EMBS, p. 4, 30 8 2011.