Speaker
Samuel Liu (劉恒惠), VP of Engineering, Uniring Tech Co., LTD
Shao-Hang Hsieh (謝少航), Director, Viscovery Pte. Ltd.
Samuel Liu (劉恒惠), VP of Engineering, Uniring Tech Co., LTD
Shao-Hang Hsieh (謝少航), Director, Viscovery Pte. Ltd.
Topic
Facility Inspection Robot
廠務巡檢機器人
Abstract
(1) Inspection Robotics allow its customers to inspect without manpower. It integrates camera, sensors and AGV with our core technologies including“Indoor Positioning and Tracking Algorithm" and“Artificial Intelligence". By integrating these technologies, the inspection robot can increase the accuracy of data analysis and image recognition.
The inspection robot collects the data on-site and transfer real-time data to the visual recognition system. If there is any anomaly, the system will alarm and send notification to related department. The FMCS center can do real-time control, monitor and record the data. This can free staff from monotonous repetitive work to perform more productive work.
However, implementing autonomous robots in practice environment still has some challenges to overcome. Most working environments are not friendly to autonomous robots, and some situation may cause issue while moving. When building a large-area factory map, the cumulative errors in the process can easily cause distortion of the map. For the trackless robots, how to maintain the correct positioning ability is a key challenge for on-site inspection robots. We delivered variety corresponding solutions to our Indoor Positioning and Tracking Algorithm. These solutions can provide a stable data and information for robot while it’s working in complex real-world environments.
巡檢機器人主要是取代現有的人力巡檢，他整合了攝影機(Camera)、感應器(Sensors)與無軌式自動導引車(AGV)，並應用人工智慧(A.I.)的技術來提升資料分析與影像辨識的準確性。
巡檢機器人將現場巡檢所蒐集到的資料即時回傳主機進行分析並將結果通知廠務值班中心(FMCS Center)，進行遠端即時監控與記錄留存，可有效提昇巡檢的效率與節省人力。
然而，在實務上導入巡檢機器人仍然有些挑戰需克服。大多數工作場域對巡檢機器人的使用並不友善，某些設施或區域可能對機器人的移動造成不便。建立大面積廠域的地圖時，過程中的累績誤差容易造成地圖失真。對於無軌式AGV而言，當局部環境的變化與地圖差異過大時，如何保有正確的定位能力，更是一項挑戰。面對這些問題，我們提出了對應的處理方法，以期能讓無軌式AGV的應用能夠確實減輕人力的負擔。
(2) Breakthroughs in Deep Learning have opened up a wide range of new possibilities in intelligent manufacturing. As promising as it is, general purpose classifiers cannot recognize items or actions of interest in factories. Custom-trained classifier require thousands of annotated training data per class, which is highly infeasible in the semiconductor industry. The workaround in AOI have been using golden templates and making oversensitive algorithms that miss few but require humans to filter out false alarms. This same setup cannot be applied to surveillance or factory inspection robots for critical missions. This talk shares the preliminary results of training highly accurate factory inspection robots capable of detecting dangerous actions, inappropriately placed objects, and other dangerous situations, by using abundant auxiliary data, transfer learning, and minimal training data from actual factory sites.
深度學習技術的突破，為智慧工廠的各式應用帶來新的契機。深度學習技術雖被寄予厚望，但通用型的視覺辨識演算法無法直接偵測工廠中各種品管、公安等問題。重新訓練客製化的深度學習辨識核心時，一個分類動輒需要上千個標記過的訓練樣本，對於重視保密、製成優化的半導體產業而言，是不實際、不可行的。半導體業的AOI光學檢測，過往採用golden template的折衷作法，讓稍有不同的成品即交由人工複檢篩除誤判。但同樣的設置，是無法複製套用到智慧監控或巡檢機器人這類的應用，因會嚴重降低真正發生事故時監控人員的敏銳度與反應速度。此演講會分享：如何利用大量場外資料結合遷移學習，讓巡檢機器人僅需少量訓練樣本，即可達到比擬重新訓練分類器水準，精確偵測出：危險非標準作業動作、限制區域出現物品、漏液、漏氣等狀況。
Facility Inspection Robot
廠務巡檢機器人
Abstract
(1) Inspection Robotics allow its customers to inspect without manpower. It integrates camera, sensors and AGV with our core technologies including“Indoor Positioning and Tracking Algorithm" and“Artificial Intelligence". By integrating these technologies, the inspection robot can increase the accuracy of data analysis and image recognition.
The inspection robot collects the data on-site and transfer real-time data to the visual recognition system. If there is any anomaly, the system will alarm and send notification to related department. The FMCS center can do real-time control, monitor and record the data. This can free staff from monotonous repetitive work to perform more productive work.
However, implementing autonomous robots in practice environment still has some challenges to overcome. Most working environments are not friendly to autonomous robots, and some situation may cause issue while moving. When building a large-area factory map, the cumulative errors in the process can easily cause distortion of the map. For the trackless robots, how to maintain the correct positioning ability is a key challenge for on-site inspection robots. We delivered variety corresponding solutions to our Indoor Positioning and Tracking Algorithm. These solutions can provide a stable data and information for robot while it’s working in complex real-world environments.
巡檢機器人主要是取代現有的人力巡檢，他整合了攝影機(Camera)、感應器(Sensors)與無軌式自動導引車(AGV)，並應用人工智慧(A.I.)的技術來提升資料分析與影像辨識的準確性。
巡檢機器人將現場巡檢所蒐集到的資料即時回傳主機進行分析並將結果通知廠務值班中心(FMCS Center)，進行遠端即時監控與記錄留存，可有效提昇巡檢的效率與節省人力。
然而，在實務上導入巡檢機器人仍然有些挑戰需克服。大多數工作場域對巡檢機器人的使用並不友善，某些設施或區域可能對機器人的移動造成不便。建立大面積廠域的地圖時，過程中的累績誤差容易造成地圖失真。對於無軌式AGV而言，當局部環境的變化與地圖差異過大時，如何保有正確的定位能力，更是一項挑戰。面對這些問題，我們提出了對應的處理方法，以期能讓無軌式AGV的應用能夠確實減輕人力的負擔。
(2) Breakthroughs in Deep Learning have opened up a wide range of new possibilities in intelligent manufacturing. As promising as it is, general purpose classifiers cannot recognize items or actions of interest in factories. Custom-trained classifier require thousands of annotated training data per class, which is highly infeasible in the semiconductor industry. The workaround in AOI have been using golden templates and making oversensitive algorithms that miss few but require humans to filter out false alarms. This same setup cannot be applied to surveillance or factory inspection robots for critical missions. This talk shares the preliminary results of training highly accurate factory inspection robots capable of detecting dangerous actions, inappropriately placed objects, and other dangerous situations, by using abundant auxiliary data, transfer learning, and minimal training data from actual factory sites.
深度學習技術的突破，為智慧工廠的各式應用帶來新的契機。深度學習技術雖被寄予厚望，但通用型的視覺辨識演算法無法直接偵測工廠中各種品管、公安等問題。重新訓練客製化的深度學習辨識核心時，一個分類動輒需要上千個標記過的訓練樣本，對於重視保密、製成優化的半導體產業而言，是不實際、不可行的。半導體業的AOI光學檢測，過往採用golden template的折衷作法，讓稍有不同的成品即交由人工複檢篩除誤判。但同樣的設置，是無法複製套用到智慧監控或巡檢機器人這類的應用，因會嚴重降低真正發生事故時監控人員的敏銳度與反應速度。此演講會分享：如何利用大量場外資料結合遷移學習，讓巡檢機器人僅需少量訓練樣本，即可達到比擬重新訓練分類器水準，精確偵測出：危險非標準作業動作、限制區域出現物品、漏液、漏氣等狀況。