Sistem Deteksi Kelengkapan Baut Pada Rangka Sepeda Motor Berbasis Klasifikasi Identitas Baut Krusial Menggunakan Metode Convolutional Neural Network

Maulana, Selfrin Surya (2026) Sistem Deteksi Kelengkapan Baut Pada Rangka Sepeda Motor Berbasis Klasifikasi Identitas Baut Krusial Menggunakan Metode Convolutional Neural Network. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 2040221111-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (8MB) | Request a copy

Abstract

Proses inspeksi baut pada rangka motor pada lini perakitan akhir di suatu perusahaan manufaktur otomotif merupakan titik quality control (QC) yang krusial untuk menjamin keamanan produk. Sistem inspeksi berbasis computer vision dengan pemrosesan citra konvensional yang pernah diterapkan sebelumnya kerap gagal membaca baut dengan warna yang menyerupai warna latar di sekitarnya sehingga menyebabkan kegagalan deteksi hingga 2000 unit per hari. Akibatnya, sistem tersebut dinonaktifkan dan inspeksi kembali dilakukan secara visual manual dengan memeriksa 20 titik baut dalam waktu 22 detik per unit pada target produksi 3000 unit per hari. Penelitian ini merancang sistem deteksi baut krusial pada rangka sepeda motor berbasis convolutional neural network (CNN) menggunakan arsitektur YOLO11 yang dilatih menggunakan sumber dataset internal perusahaan dengan membandingkan performa model YOLO11n standar dan model YOLO11n-P2. Hasil pengujian menunjukkan bahwa model YOLO11n-P2 memberikan performa klasifikasi terbaik pada test set dengan nilai accuracy, precision, recall, dan F1-score masing-masing sebesar 100%, mengungguli model YOLO11n yang memperoleh accuracy 99,97%, precision dan recall 99,75%, serta F1-score 99,74%. Pada pengujian terhadap 100 unit rangka motor di luar dataset pelatihan, model YOLO11n-P2 berhasil mendeteksi baut lengkap pada 92 unit, sedangkan model YOLO11n mendeteksi 88 unit. Model YOLO11n-P2 membutuhkan waktu inferensi rata-rata 436,57 ms per siklus inspeksi, lebih tinggi dibandingkan model YOLO11n standar sebesar 279,17 ms, namun masih jauh di bawah batas cycle time produksi sebesar 22 detik per unit sehingga tetap memenuhi kebutuhan implementasi di lini perakitan. Pengujian komunikasi web socket menghasilkan rata-rata latensi sebesar 0,56 ms dengan rata-rata jitter sebesar 0,31 ms, sedangkan komunikasi Modbus TCP/IP mencatat rata-rata latensi sebesar 1,10 ms dengan rata-rata jitter sebesar 0,37 ms, yang menunjukkan kemampuan pertukaran data secara cepat dan stabil antara sistem deteksi, PLC, dan web dashboard. Berdasarkan keseluruhan hasil tersebut, model YOLO11n-P2 dipilih sebagai konfigurasi optimal untuk diintegrasikan dengan database dan web dashboard pada sistem deteksi kelengkapan baut otomatis.
=========================================================================================================================================
The bolt inspection process on a motorcycle frame on the final assembly line in an automotive manufacturing company is a crucial quality control (QC) point to ensure product safety. The computer vision-based inspection system with conventional image processing that was previously implemented often failed to read bolts with colors that resembled the surrounding background color, resulting in detection failures of up to 2,000 units per day. As a result, the system was deactivated and manual visual inspection was carried out again by checking 20 bolt points in 22 seconds per unit at a production target of 3,000 units per day. This study designs a crucial bolt detection system on a motorcycle frame based on a convolutional neural network (CNN) using the YOLO11 architecture trained using the company's internal dataset source by comparing the performance of the standard YOLO11n model and the YOLO11n-P2 model. The test results show that the YOLO11n-P2 model provides the best classification performance on the test set with accuracy, precision, recall, and F1-score values of 100% each, outperforming the YOLO11n model which obtained 99.97% accuracy, 99.75% precision and recall, and 99.74% F1-score. In testing on 100 motorcycle frames outside the training dataset, the YOLO11n-P2 model successfully detected complete bolts on 92 units, while the YOLO11n model detected 88 units. The YOLO11n-P2 model requires an average inference time of 436.57 ms per inspection cycle, higher than the standard YOLO11n model of 279.17 ms, but still far below the production cycle time limit of 22 seconds per unit, thus still meeting the implementation needs on the assembly line. Web socket communication testing produced an average latency of 0.56 ms with an average jitter of 0.31 ms, while Modbus TCP/IP communication recorded an average latency of 1.10 ms with an average jitter of 0.37 ms, indicating the ability to exchange data quickly and stably between the detection system, PLC, and web dashboard. Based on these overall results, the YOLO11n-P2 model was selected as the optimal configuration to be integrated with the database and web dashboard in the automatic bolt completeness detection system.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Convolution Neural Network, YOLO, Deteksi Objek, Baut, Rangka Motor, Object Detection, Bolt, Motorcycle Frame
Subjects:	H Social Sciences > HD Industries. Land use. Labor > HD30.213 Management information systems. Dashboards. Enterprise resource planning. H Social Sciences > HD Industries. Land use. Labor > HD9980.5 Service industries--Quality control. Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Vocational > 36304-Automation Electronic Engineering
Depositing User:	Selfrin Surya Maulana
Date Deposited:	05 Aug 2026 09:09
Last Modified:	05 Aug 2026 09:09
URI:	http://repository.its.ac.id/id/eprint/144077

Actions (login required)

View Item