以卷積神經網路檢測塑膠射出成型製品瑕疵

An Application of Convolutional Neural Network in Detecting Defects of Plastic Injection Molding Products

徐晧恩(Hsu Hao-En)

指導教授 ： 黃啟祐

國立臺灣師範大學/科技與工程學院/工業教育學系/碩士(2021年)

https://doi.org/10.6345/NTNU202101633

摘要

一個產品的成功與否與企業的成敗息息相關，而品質管理是影響產品成功的重要因素，也是形塑企業競爭力的重要環節。塑膠射出成型適用於生產形狀複雜的加工工藝，而使用普遍原因是因為塑膠射出成型能在較短的週期時間大量生產出產品，且具有非常高的產品精度，所以在消費型電子產品上被廣泛使用，其中因為產品需要非常精密的尺寸與外觀，使品質把關變得非常嚴格。基於完善的瑕疵檢測系統可以幫助產線更快的判定產品的好壞，藉以減少人力成本以及檢驗成本等動機與目的。本研究基於卷積神經網路，使用傳統卷積神經網路以及You Only Look Once (YOLO) V3做為本研究模型架構開發塑膠射出成型產品不良品瑕疵檢測系統。首先本研究選定Rombit所開發之工人安全與防疫追蹤穿戴裝置Romware ONE作為研究對象，拍攝一定數量的相片，經過預處理後將資料分別導入傳統卷積神經網路以及基於YOLO V3所開發模型中訓練，直到模型達到理想情況。本研究結果使用傳統卷積神經網路發生過度擬和情況，使實際使用不如預期。而YOLO V3建構之模型經過訓練與調整最終達到F1分數為0.9775，mAP50為96.8254，因此在偵測產品缺陷上面有良好表現。未來可能再細分更多缺陷類別，達到更精確的檢測效果。

關鍵字

卷積神經網路 ； YOLO V3 ； 物件偵測 ； 自動光學檢測 ； 塑膠射出成型

並列摘要

The success of a product is closely related to the success or failure of the enterprise, and quality management is an important factor affecting the success of the product, but also an important link of shaping the competitiveness of enterprises. Plastic Injection Molding (PIM) is suitable for the production of complex shape machining process, and use the common reason for plastic injection molding can in a relatively short period of time the product in mass production and high precision products, so is widely used in consumer electronics, which because of the size and appearance of the products need to be very precise, make quality control is very strict. Based on the perfect defect detection system can help the production line to determine the quality of products more quickly, so as to reduce labor costs and inspection costs and other motives and objectives. This study based on convolutional neural network, uses traditional convolutional neural network and You Only Look Once (YOLO) V3 as the model framework to develop a defect detection system for plastic injection molding products. First, this research selects Romware ONE, a wearable device developed by Rombit, as the research object, and takes a certain number of photos. After preprocessing, the data is imported into the traditional convolutional neural network and the model developed based on YOLO V3 for training until the model reaches the ideal situation.
The result of this study is that the traditional convolutional neural networks are over-fitting and the actual use is not as expected. After training and adjustment, the model constructed by YOLO V3 finally achieved a F1 score of 0.9775 and a mAP50 score of 96.8254, so it has a good performance in detecting product defects. In the future, more defect categories may be subdivided to achieve more accurate detection results.

並列關鍵字

Convolutional Neural Network ； YOLO V3 ； Object Detection ； Automated Optical Inspection ； Plastic Injection Molding

參考文獻

Abiodun, O. I., Jantan, A., Omolara, A. E., Dada, K. V., Mohamed, N. A., & Arshad, H. (2018). State-of-the-art in artificial neural network applications: A survey. Heliyon, 4(11), e00938.

Albawi, S., Mohammed, T. A., & Al-Zawi, S. (2017). Understanding of a convolutional neural network. Paper presented at the 2017 International Conference on Engineering and Technology (ICET), Antalya, Turkey.

Aloysius, N., & Geetha, M. (2017). A review on deep convolutional neural networks. Paper presented at the 2017 International Conference on Communication and Signal Processing (ICCSP), Chennai, India.

Benjdira, B., Khursheed, T., Koubaa, A., Ammar, A., & Ouni, K. (2019). Car detection using unmanned aerial vehicles: Comparison between faster r-cnn and yolov3. Paper presented at the 2019 1st International Conference on Unmanned Vehicle Systems-Oman (UVS), Muscat, Oman.

Bryce, D. M. (1997). Plastic injection molding: material selection and product design fundamentals. Dearborn, MI: Society of Manufacturing Engineers.