| 摘要: | 現今的噴射混凝土的噴塗工程，包含建築牆面、隧道工程的噴塗作業，大部分都是由人力完成，在作業的過程中會因為噪音、粉塵、嚴峻的施工環境等對施工人員造成不可逆的傷害；建築物以及隧道工程亦有檢修年限的問題，依舊需要人力檢修以及補強。
本研究預計研發一款混凝土缺陷自動修補機器人，此機器人經由微控制器(Micro Control Unit ,MCU)控制各機構的行為，在手機或電腦端以無線網路(Wi-Fi)連接微控制器進行遠端控制，互動介面則是客製化網頁，使用HTML以及JavaScript編寫溝通伺服器端以及客戶端；在噴射混凝土噴槍上方搭載視覺辨識模組，基於機器學習 (Machine Learning)以及影像辨識(Visual Recognition)技術蒐集圖片進行訓練替代人力檢測，視覺辨識訓練使用卷積神經網路(Convolutional Neural Network, CNN)演算法，多次調整訓練模型訓練後正確率達到94.2%、隨機圖片預測正確率97.28%，F1-Score在訓練模型達到0.9、隨機圖片預測達到0.96，並設計演算法完成牆面自動掃描，微控制器配合視覺辨視模組回報牆面缺陷狀況，當缺陷被辨識到，微控制器將遠端控制噴射混凝土幫浦控制盒的微控制器扳動開關，即可自動化完成缺陷自動噴塗修補作業。
本研究整合機器人自動控制、噴射混凝土、視覺辨識系統、物聯網系統以及網頁遠端控制，經由實驗證實可以自動化完成缺陷辨識以及噴射混凝土的噴塗作業，開發出未來自動化缺陷辨識補強機器人的雛形。
;Nowadays, shotcrete operation, including the spraying of building walls and tunnels, is done mainly by human resources. During the operation, it will cause irreversible damage to the construction personnel due to noise, dust, and a severe construction environment. Buildings and tunnel projects also have the problem of the maintenance period, which still requires manual maintenance and reinforcement.
This research is expected to develop an automatic repairing robot for concrete defects. This robot controls the behavior of each device through a microcontroller and connects the microcontroller with a wireless network (Wi-Fi) on the mobile phone or computer. Remote control, the user interface is a customized web page using HTML and JavaScript to contact the communication server and client. A visual recognition module is carried on the shotcrete spray gun, based on Machine Learning and Visual Recognition technology to collect pictures for visual recognition to replace manual work. Visual recognition training uses Convolutional Neural Network (CNN) algorithm. Then designed, an algorithm automatically scans the target wall, and the microcontroller cooperates with the visual recognition module to report the defect status of the wall. When the defect is identified, the microcontroller will remotely control the microcontroller of the shotcrete pump control box and flip the switch to automate complete the defect automatic spray repair operation. After adjusting the training model many times, the accuracy rate reached 94.2%, the random image prediction accuracy rate was 97.28%, and the F1-Score reached 0.9 in the training model and 0.96 in random image prediction.
This research develops the prototype of the future automated defect identification reinforcement robot. This research integrates automatic robot control, shotcrete, visual recognition system, Internet of Things system, and web remote control. It has been confirmed through experiments that defect identification and shotcrete spraying can be automatically completed. |
