2D工程圖之深度學習的整合應用

製造業是經濟基礎與工業化的產業主體，在產品設計需求多樣化，且製造精度要求越來越高的環境下，工程圖設計的精準度要求也相對受到重視，而製造工程統計中，完整的專案計畫所用的工程圖可能超過百張之上，在如此龐大的圖面資訊中，繪製、修改、儲存、管理都不盡相同，尤其圖面呈現的方式常依個人習慣而有所不同，在格式、符號無法標準化下，會導致後續解讀不一致的現象。故合理圖紙標註決定產品品質、後期檢測和加工成本，且能有效表達產品功能，並使圖紙解釋具唯一性，保證產品可製造性，以及重複量測和再現性，減少人為誤判的情況發生。然而，當圖面特徵訊息一多，圖面解讀與分析往往造成產品驗證時間拉長，耗時又費力使成本增加，衍生製造加工另一類問題。本研究採用幾何尺寸公差(Geometric Dimensioning and Tolerancing, GD&T)，對應ASME Y14.5 2018規範描述工程圖紙語言，並使用Darknet SDK(Software Development Kit, SDK)來訓練YOLO(You Only Look Once)模型，以2D工程圖作為訓練資料，來對工程圖進行影像分割，以網路模型對2D圖紙辨識出尺寸、公差、功能框及幾何符號等圖面特徵訊息物件，再藉由座標讀取各物件相對應的位置，顯示正確的數值，並以實際案例，使用多張有混合特徵的工程圖訓練模型，達到辨識能力能優於單一特徵的識別成效，提高深度學習的辨識準確度，使工程圖及影像識別實用化，讓整體效能在2D工程圖上能辨識出七成正確特徵訊息，並直接導入數據庫，且能縮短產品驗證時間，避免數值因人工不慎填寫錯誤，造成後續品質管控問題發生。

關鍵字

2D工程圖；影像分割；深度學習；圖形辨識； ASME Y14.5

並列摘要

Manufacturing industry is the foundation of economics and industry. Since the diversify and precision of product have become higher and higher, the accuracy of engineering drawings is getting more and more attention. There are over hundreds of drawings are used in a life cycle of product. Amount this huge information, each person has different ways to illustrate, modify, storage and manage. It might lead to inconsistent interpretation when the format and symbol could not be standardized. Therefore, the reasonable notations determine product quality, examination and produce cost. It also helps showing products function effectively and make sure the one and only interpretation to assure the manufacturability, repeatability, and reproducibility, which can reduce misjudgments. However, excessive information for interpretation and analysis cause time consuming and increasing cost is another manufacture problem. Thus, this research applies Geometric Dimensioning and Tolerancing (GD&T) with ASME Y14.5 2018 regulation to illustrate drawing language and use Darknet Software Development Kit to train the model called You Only Look Once (YOLO). Using 2D drawing as training data to segment image and recognize feature objects including size、tolerance、functional diagram and geometric symbol with online model and then get the relative location with coordinate to show the correct figures. This research uses multiple engineering drawing training models with mixed features to achieve the better identify ability than single feature. Therefore, the accuracy of deep learning of identification can be improved and make it more practical. Over 70% correct features can be identify correctly and directly import into data base which will shorten products examination period and avoid manual errors of filling wrong values and resulting in quality control issue happen.