| 摘要: | 在臺灣的建築法規範下，若是因為外牆磁磚掉落導致的人身傷害或是財務毀損，根據建築法第91條第1項的規定，無論是建商、建物所有權人、使用人、管理人，甚至是公寓大廈的管理委員會，都將必須承擔起法律責任，並且可能遭受罰款和刑事責任。然而，傳統的檢測和維修方式卻面臨許多挑戰，不僅效率低，還需要大量的人力和資源的投入才能完成工作，現今社會缺工的問題越來越嚴重，尤其是對於這種高風險且需要專業技能的工作人力更顯短缺，這無疑是增加了維護建築安全的困難度。除此之外，這種方法在執行過程中可能會出現疏漏，不僅增加了意外的風險，也可能延長了修復工作的時間。因此，尋找一種能提高效率且減少人力投入的新方法，變得至關重要。
本研究旨在研發一款仿生攀爬機器人—GLIMBOT (Gecko-Leech Inspired Climbing Robot)，應用於外牆磁磚缺陷檢測，以協助檢測人員遠端遙控進行非破壞性檢測。靈感源於仿生學，模擬壁虎和水蛭的行為模式。攀爬機器人的重要性能之一是吸附能力，因此，在有限的吸附力下，嚴格控制機器人的重量是關鍵。為此，本研究利用3D列印技術製造出機器人的主要部分，實現了輕量化且具有高韌性的機身。本研究亦導入了機器學習的元素，採用梅爾頻率倒譜係數(Mel-Frequency Cepstral Coefficients,MFCC)特徵提取方法來提取聲音的特徵，並透過使用卷積神經網路(Convolutional Neural Networks ,CNN) 進行訓練和學習，並將優化模型部署在AI聲學辨識模組上，經由實地攀爬敲擊辨識實驗，模型對於實際敲擊磁磚的識別準確率能達到74.62%，此方法能快速識別磁磚的缺陷，並將結果即時傳送給使用者。開發過程中，經歷了五個版本的設計，並專注於微調機構設計，釋出足夠空間以容納關鍵零件。最後透過進行吸力測試、摩擦力實驗以及攀爬試驗來得到相關參數，以此作為設計及驗證的依據。
;Under the construction regulations in Taiwan, if personal injury or financial damage is caused by the falling of exterior wall tiles, according to Article 91, Paragraph 1 of the Building Act, constructors, building owners, users, managers, and even apartment management committees are required to bear legal responsibility. They may also face fines and criminal liability. However, traditional inspection and maintenance methods face many challenges. They are not only inefficient but also require a large amount of manpower and resource investment to complete the job. The problem of labor shortage in today′s society is becoming increasingly serious, especially for high-risk jobs that require professional skills, which undoubtedly increases the difficulty of maintaining building safety. In addition, there may be omissions in the execution process of these methods, which not only increases the risk of accidents but also possibly extends the time for repair work. Therefore, it is crucial to find a new method that can improve efficiency and reduce manpower input.
This study aims to create a bionic climbing robot - GLIMBOT (Gecko-Leech Inspired Climbing Robot), applied for the detection of defects in exterior wall tiles, to assist inspectors in remote control for non-destructive testing. Inspired by bionics, it simulates the behavior patterns of geckos and leeches. One of the important capabilities of a climbing robot is its adhesion ability, so under limited adhesion, strictly controlling the weight of the robot is key. To this end, this study uses 3D printing technology to make the main parts of the robot, achieving a lightweight and highly resilient body. This study also introduces the elements of machine learning, uses the Mel-Frequency Cepstral Coefficients (MFCC) feature extraction method to extract sound features, and trains and learns through Convolutional Neural Networks (CNN), and deploys the optimized model on a microcontroller. Through field climbing and knocking identification experiments, the model can achieve an accuracy rate of 74.62% for the actual knocking on tiles. This method can quickly identify tile defects and send the results to users in real time. During the development process, five versions of the design were experienced, focusing on fine-tuning the mechanism design to release enough space to accommodate key components. Finally, through suction tests, friction experiments, and climbing tests, relevant parameters are obtained as the basis for design and verification. |
