應用深度學習於內螺紋瑕疵檢測之研究

Application on The Deep Learning Technology for Internal Thread Defect Detection

許光城(Quang-Cherng Hsu) ； 丁凡琇(Fan-Siou Ding) ； 黃麗萍(Lih-Ping Huang)

《鍛造》 31卷3期 (2022/12) Pp. 22-31

摘要

螺帽在機械工業中應用廣泛，對於螺帽內螺紋之瑕疵檢測方式多由品管人員以目視進行檢測，此方法常因目視檢測時間過長造成眼睛疲勞進而發生漏檢之情形，因此本研究應用深度學習檢測內螺紋之瑕疵希望能取代人工目視檢測。因內螺紋空間位置限制，在影像處理上其邏輯法則及訂定條件較困難實現，本研究應用人工智慧技術中之深度學習，以兩種網路模型對影像資料進行訓練，最後做出分類預測。在資料擴增的部分，以隨機縮放之方式來進行來擴增內螺紋的訓練資料，探討Resnet50及ML.NET兩種不同模型對於訓練資料集的準確率，在內螺紋瑕疵檢測的部分，Resnet50與ML.NET之訓練準確率分別為100.00%及99.29%，而在測試資料的部分，Resnet50的網路模型對於無瑕疵、毛邊、殘油與顆粒瑕疵之F1 Score皆為100%；而ML.NET網路模型對於無瑕疵、毛邊、殘油與顆粒瑕疵之F1 Score則分別為99.67%、99.64%、99.28%及100.00%，由此可知Resnet50對於不同瑕疵特徵的判斷能力更佳。

並列摘要

Nuts are widely used in the machinery industry. The most common defect inspection method of the internal threads of the nuts is a visual inspection. This method is not only time-consuming but also leads to false-negative due to eyes fatigue. Therefore, deep learning technology was used to inspect the defect of internal thread on nuts in this study. However, due to the limitation of the space of the internal thread, it is difficult to use traditional image processing which is based on an algorithm base. Artificial intelligence technology was used in this research to reduce the complicated processing procedures in the program. Two different neural network models were used to train the image data, and finally make classification predictions. In the part of data amplification, random scaling is used to expand the training data of internal threads, and the accuracy of the two different models of Resnet50 and ML.NET on the training data set is discussed. About the part of internal thread defect detection, the training accuracy of Resnet50 and ML.NET is 100.00% and 99.29%, respectively. At the part of test data, the Resnet50 network model has 100% F1 Score for flawlessness, burrs, residual oil and particle defects; while the ML.NET network model is for flawless, burrs, residual oil and particle defects. F1 Score is 99.67%, 99.64%, 99.28% and 100.00%,it can be seen that Resnet50 has better judgment ability for different defect characteristics.