運用深度學習探討金屬零件外觀瑕疵檢測-以S公司為例;Using deep learning techniques for surface defects detection of metal parts - take S company as an example

NCU Institutional Repository > 管理學院 > 資訊管理學系碩士在職專班 > 博碩士論文 > Item 987654321/86588

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/86588

题名:	運用深度學習探討金屬零件外觀瑕疵檢測-以S公司為例;Using deep learning techniques for surface defects detection of metal parts - take S company as an example
作者:	邱旭輝;CHIU, HSU-HUI
贡献者:	資訊管理學系在職專班
关键词:	金屬表面瑕疵檢測;深度學習;類神經網路;Metal surface defect detection;Deep Learning;Neural Network
日期:	2021-07-27
上传时间:	2021-12-07 13:00:15 (UTC+8)
出版者:	國立中央大學
摘要:	品質管理是製造業的一個重要課題，好的品質管理可以降低營運成本、減少風險、獲取良好的客戶滿意度。近幾年圖形處理器（GPU）技術持續突破，類神經網路演算法的不斷推陳出新，使得人工智慧運用在工廠瑕疵檢測上成為新的一種選擇。運用深度學習進行瑕疵檢測，需要大量的瑕疵樣本進行模型訓練，才能提高其瑕疵辨識正確率。金屬加工業生產技術成熟，產量大、良率高，要蒐集大量的瑕疵樣本需要耗費相當多的時間與人力。本研究透過二階段方式，第一階段先使用二元分類模型進行「良品」、「瑕疵品」分類，證實僅需要250張良品樣本與250張瑕疵樣本進行模型訓練，就能以高於99 % 正確率判斷良品與瑕疵品，證實該訓練模型可實際運用於生產環境中，並可大幅縮短初期瑕疵樣本的蒐集時間，透過系統分階段上線，快速降低部分外觀品質檢驗之人力或檢驗設備，有效降低瑕疵品流出之機會。透過階段一判定後之瑕疵工件，數量已大幅減少，再利用「人工視檢」或「自動光學檢測」設備來進行覆判，並進行瑕疵樣本蒐集及瑕疵標註，用來進行第二階段瑕疵類別辨識模型訓練。雖然初期會因瑕疵樣本數不足，實際檢測正確率不高，但仍可協助覆判人員預先選擇瑕疵分類及瑕疵位置，透過覆判人員對錯誤資料的校正與分析，持續蒐集瑕疵樣本照片，不斷的進行模型訓練及調整，來逐步改善其辨識的正確率。透過加入二元分類作業，快速減少部分檢驗人力需求，降低長期資料蒐集成本，讓深度學習運用在金屬表面瑕疵檢測的門檻降低，促使成為傳統「人工目視」檢驗或「自動光學檢測」的另一種選擇。 ;Quality management is an important topic in the manufacturing industry. Excellent quality management provides advantage in reduce operating costs, risks and obtain good customer satisfaction. Recent years, technology breakthrough for graphics process unit (GPU) and the innovation of neural network algorithms have made the artificial intelligence become a new option for quality inspection in manufacturing process. Apply deep learning for defect inspection requires large amount of sample to create model in order to increase its accuracy for identity defect. The challenge for collecting large scale of sample in metal-processing industry is its long lead time and consume manpower resources due to its mature for production technology in mass production and high quality yield. This study demonstrate two stage method that could enhance the disadvantage for the long lead time and manpower usage. In the first stage, using binary classification model to proceed “Good” and “Defect” separation has proved that by only study and create training model from 250 samples of good products and 250 samples of defect products is sufficient to provide accuracy quality inspection by above 99%. It is prove that this training model can apply in real mass production and greatly shorten the collection time for defect samples. By launches system in different stage can quickly reduce the inspection manpower or equipment and also reduce the chance to miss. After inspection in stage 1, the defect amount has already significant reduced. By using “Manual Inspection” or “Automatic Optical Inspection – AOI” will also support to confirm and determine the defect samples after stage 1 and the result can use in the stage II for defect category identification model training. Although the actual detection rate is not accuracy enough due to the insufficient number of defect samples in the initial stage, however, it can still assist the reviewers to pre-select the defect classification and defect location. Through the correction and continue analysis of the wrong data and defect photo images and continue adjust the study model to reinforce the accuracy step by step. By adding binary classification operations, it can quickly reduce the inspection man-power needs and long-term data collection. It also creates advantage in lower the barrier for metal surface defect inspection thru the deep learning process and making it an alternative selection for traditional “Manual Inspection” or “Automatic Optical Inspection – AOI”.
显示于类别:	[資訊管理學系碩士在職專班 ] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	89	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....