Sistem Deteksi Cacat Flash Untuk Kontrol Proses Friction Stir Welding (FSW) Pada Material Aa6061-T651 Dengan Metode Convolutional Neural Network (CNN)

Alamy, Ulya Ganeswara (2022) Sistem Deteksi Cacat Flash Untuk Kontrol Proses Friction Stir Welding (FSW) Pada Material Aa6061-T651 Dengan Metode Convolutional Neural Network (CNN). Other thesis, Institut Teknologi Sepuluh Nopember.

Text 6007201046-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only until July 2024. Download (2MB)

Abstract

Friction Stir Welding (FSW) merupakan metode pengelasan modern yang mana pengelasan ini terjadi dari gesekan suatu pahat dan material tanpa meleburkan material itu sendiri, sehingga tidak mengakibatkan bahan dasar material karena peleburan seperti las fusi. Meskipun proses FSW muncul sebagai proses pengelasan terdepan, tentunya diperlukan beberapa inspeksi bahan untuk menilai kualitas hasilnya, serta perlu adanya pendeteksian dini suatu cacat yang terjadi. Deteksi dini untuk mengendalikan cacat las dengan kecepatan dan objektivitas yang tinggi sangat diperlukan. Visual Inspection (VT) adalah metode yang sangat penting dan tahap awal sebelum material yang dilas akan diuji pada Destructive Test (DT). Selama ini VT hanya menggunakan penglihatan manusia, yang membutuhkan proses panjang dan sangat subjektif. Penelitian ini akan memberikan kontribusi pada metode VT untuk mengontrol hasil proses Friction Stir Welding (FSW) dengan mendeteksi cacat flash pada material AA6061-T651 menggunakan salah satu cabang deep learning yang berperan dalam pendeteksian suatu citra yaitu Convolutional Neural Network (CNN). Dengan demikian, cacat flash yang terjadi pada hasil proses FSW dapat dideteksi sedini mungkin dan diminimalisir. CNN berfungsi sebagai pengganti penglihatan manusia. Pemilihan CNN dinilai tepat untuk pendeteksian suatu citra karena prosesnya yang cepat dan mendeteksi fitur-fitur penting tanpa pengawasan manusia yang dilakukan dengan proses pembelajaran yang berkesinambungan. Sebanyak 620 citra dari proses FSW diolah menjadi dua kelompok dataset. Citra tersebut diproses dengan dua jenis arsitektur CNN, diantaranya AlexNet dan VGG16. Berdasarkan hasil VT oleh CNN, model AlexNet menunjukkan akurasi deteksi cacat flash sebesar 91,03%, sedangkan model VGG16 menunjukkan akurasi deteksi cacat flash sebesar 77,35%. Sementara performa model AlexNet menunjukkan nilai precision 0.8462, sensitivity 0.9778, F-measure 0.9072, dan accuracy 0.9230 serta model VGG16 memiliki nilai precision 0.6100, sensitivity 0.8188, F-measure 0.6991, dan accuracy 0.6610. Dari hasil tersebut, keberhasilan CNN model AlexNet dalam melakukan VT pada kontrol proses FSW cukup tinggi dan dapat berperan lebih besar dalam memantau hasil proses FSW. Oleh karena itu, kemungkinan cacat flash dapat diminimalisir dan dideteksi sedini mungkin. Sehingga model AlexNet dipilih untuk direkomendasikan pada penelitian selanjutnya.
========================================================================================================================
Friction Stir Welding (FSW) is a modern welding method where this welding occurs from the friction of a tool and material without melting the material itself, so it does not result in the basic material being melted like fusion welding. Even though the FSW process has emerged as a leading welding process, of course, several inspections of materials are needed to assess the quality of the results, as well as the need for early detection of any defects that occur. Early detection to control welding defects with high speed and objectivity is needed. Visual Inspection (VT) is a very important method and the initial stage before the welded material will be tested on the Destructive Test (DT). So far, VT only uses human vision, which requires a long and highly subjective process. This study will contribute to the VT method to control the results of the Friction Stir Welding (FSW) process by detecting flash defects in the AA6061-T651 material using one of the deep learning branches that plays a role in detecting an image, namely Convolutional Neural Network (CNN). Thus, flash defects that occur in the results of the FSW process can be detected as early as possible and minimized. CNN serves as a substitute for human vision. The selection of CNN is considered appropriate for the detection of an image because the process is fast and detects important features without human supervision which is carried out with a continuous learning process. A total of 620 images from the FSW process were processed into two groups of datasets. The image is processed with two types of CNN architecture, including AlexNet and VGG16. Based on the results of VT by CNN, the AlexNet model shows a flash defect detection accuracy of 91.03%, while the VGG16 model shows a flash defect detection accuracy of 77.35%. Meanwhile, the performance of the AlexNet model shows a precision value of 0.8462, sensitivity 0.9778, F-measure 0.9072, and accuracy 0.9230 and the VGG16 model has a precision value of 0.6100, sensitivity 0.8188, F-measure 0.6991, and accuracy 0.6610. From these results, the success of the CNN AlexNet model in performing VT on the FSW process control is quite high and can play a greater role in monitoring the results of the FSW process. Therefore, the possibility of flash defects can be minimized and detected as early as possible. So the AlexNet model was chosen to be recommended for further research.Keywords: Detection; Flash defect; Friction Stir Welding (FSW); Visual Inspection (VT); Convolutional Neural Network (CNN).

Item Type:	Thesis (Other)
Additional Information:	RTM 671.52 Ala s-1
Uncontrolled Keywords:	AA6061-T651; Cacat flash; Friction Stir Welding (FSW); Visual Inspection (VT); Convolutional Neural Network (CNN).
Subjects:	T Technology > TN Mining engineering. Metallurgy > TN879.6 Welding
Divisions:	Faculty of Industrial Technology and Systems Engineering (INDSYS) > Mechanical Engineering > 21101-(S2) Master Thesis
Depositing User:	Mr. Tondo Indra Nyata
Date Deposited:	19 Jan 2023 02:34
Last Modified:	19 Jan 2023 02:34
URI:	http://repository.its.ac.id/id/eprint/95478

Actions (login required)

View Item