| 摘要: | 在鋼結構工程中，利用螺栓連接鋼材部件是一種十分常見的方法，螺栓可以保證各個結構部件之間穩固相連。然而，隨著氣候、外力、地震等自然或人為因素影響，螺栓可能會因此產生脫落、鬆動等缺陷情形，導致鋼構造物的安全性降低。目前螺栓檢測工作，大多通過檢測人員利用橡膠槌等工具敲擊螺栓以人工判斷其鬆動情形，這些檢測工作繁瑣費力，且可能遇到安全風險或是檢測困難的情形。若能夠以小型儀器對於缺陷情形進行自動化檢測，不僅能夠提高檢測效率，也能夠降低檢修人員進行檢修工作時的困難。因此，本研究結合Tiny Machine Learning的理念以及微控制器，開發可用於攀爬檢測機器人之輕量級即時多目標螺栓缺陷影像檢測系統，用於對螺栓的缺陷樣態進行檢測。螺栓的樣態被分為正常(Normal)、鬆動(Loosen)、欠缺(Miss)三種，利用FOMO (Faster Objects, More Objects)演算法進行視覺檢測任務，並將訓練後的模型部署於視覺檢測模組中，結合磁吸攀爬機器人進行實際螺栓檢測測試。本研究所訓練之模型在驗證集中獲得74.8%的F1分數，並且於測試集獲得72.9%的F1分數，而部署於視覺檢測模組上的量化模型則在驗證集中獲得72.4%的F1分數。除此之外，實際進行室外箱梁檢測試驗，在10個Case中，表現最好的Case所獲得的精確率及召回率分別為89%及82%，最差的則是57%與67%，而所有Case之精確率及召回率平均值皆為77%和76%。通過該缺陷檢測系統之結果，並且能夠及時回傳影像供使用者查看螺栓辨識結果，供檢測人員在遠端就能操控設備執行檢測任務，且整個系統更為輕量化和低能量消耗，並且能夠配置於不同的載具當中進行檢測任務。未來，螺栓缺陷檢測能夠變的更為容易且降低人力需求，並且整體檢測工作所需的儀器成本能夠下降，輕量化和高速檢測更是未來螺栓檢測工作的目標。;In steel structure engineering, using bolts to connect steel components is a common method that ensures a secure connection between structural elements. However, due to natural or human factors such as climate, external forces, and earthquakes, bolts may experience defects such as loosening or detachment, leading to a decrease in the safety of steel structures. Currently, bolt inspection work is mostly performed by inspectors manually checking for loosening by using rubber hammers and other tools. These inspection tasks are labor-intensive, time-consuming, and may involve safety risks or difficulties in inspection.
To address this, this study combines the concept of Tiny Machine Learning with microcontrollers to develop a lightweight real-time multi-target bolt defect image detection system for climbing inspection robots. This system is used to detect the defect patterns of bolts, which are categorized into "Normal," "Loosen," and "Miss." The Faster Objects, More Objects (FOMO) algorithm is used for the visual detection task, and the trained model is deployed in the visual detection module, combined with a magnetic climbing robot for actual bolt inspection tests.
The trained model achieved an F1 score of 74.8% on the validation set and an F1 score of 72.9% on the test set, while the quantized model deployed on the microcontroller achieved an F1 score of 72.4% on the validation set. In actual outdoor box beam tests, the best-case achieved precision and recall rates of 89% and 82%, respectively, while the worst-case achieved 57% and 67%. The average precision and recall rates for all cases were 77% and 76%, respectively.
The defect detection system can provide real-time feedback on the bolt recognition results, allowing inspectors to remotely control the device for inspection tasks. The entire system is lightweight, low-energy consuming, and can be configured on different carriers for inspection tasks. In the future, bolt defect detection is expected to become easier and reduce manpower requirements. The overall instrument cost for inspection work can be reduced, and the goals of lightweight and high-speed detection are pursued in future bolt inspection tasks. |
