Sistem Inspeksi Visual Isi Kurang Folding Box Dengan Metode Real Time - Detection Transformer (RT-DETR) Dalam Menangani Kelengkapan Isi Kurang Shipper Box

Anggono, Prayogi Galih (2026) Sistem Inspeksi Visual Isi Kurang Folding Box Dengan Metode Real Time - Detection Transformer (RT-DETR) Dalam Menangani Kelengkapan Isi Kurang Shipper Box. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 2040221080-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (6MB) | Request a copy

Abstract

Pengendalian kualitas pada lini produksi PT SC Johnson Manufacturing Surabaya, khususnya pada proses pengemasan folding box ke dalam shipper box, saat ini masih mengandalkan inspeksi visual manual. Hal ini rentan terhadap risiko ketidaklengkapan isi produk (underfill). Tingkat kegagalan akibat ketidaksesuaian jumlah produk di dalam kardus yang mencapai 16,6% masih menjadi kendala operasional dalam menjaga kelancaran jalur produksi. Oleh karena itu, diperlukan sistem inspeksi otomatis yang mampu bekerja secara konsisten dan real-time guna meminimalkan human error. Penelitian ini mengembangkan sistem inspeksi visual berbasis deep learning menggunakan metode Real-Time Detection Transformer (RT-DETR) untuk mendeteksi dan menghitung jumlah folding box secara otomatis. Model ini dipilih karena kemampuannya melakukan deteksi objek secara end-to-end tanpa proses Non-Maximum Suppression (NMS), serta unggul dalam menangani objek yang padat dan tumpang tindih sehingga mencegah hilangnya deteksi (missed detection). Sistem dirancang untuk mengklasifikasikan kardus secara spesifik ke dalam empat kondisi operasional: kosong, isi kurang, isi penuh, dan tertutup. Kinerja model RT-DETR ini kemudian dibandingkan secara komparatif dengan algoritma YOLOv8 yang digunakan sebagai baseline evaluasi. Sistem ini diimplementasikan menggunakan kamera sebagai sensor visual dan Industrial PC sebagai unit pemrosesan, yang terintegrasi pada jalur konveyor roller manual menuju mesin sealer. Hasil pengujian menunjukkan bahwa sistem secara fungsional berhasil mengidentifikasi keempat kondisi kardus berdasarkan evaluasi kinerja, RT-DETR beroperasi sangat presisi dengan nilai metrik mAP50-95 mencapai 0,969, akurasi klasifikasi kondisi lapangan sebesar 99,44%, dan kecepatan stabil rata-rata 37 FPS. Sebagai bentuk pengamanan (safety interlock), apabila kamera mendeteksi pelanggaran seperti kardus yang belum terisi penuh namun sudah tertutup (sekuens melompat dari "Isi Kurang" ke "Tertutup"), sistem akan memicu buzzer peringatan dan secara otomatis mematikan mesin sealer. Mekanisme pencegahan ini terbukti efektif menghindari masuknya produk cacat meskipun operator secara tidak sengaja tetap mendorong kardus. Seluruh aktivitas inspeksi direkap secara otomatis ke dalam laporan teks CSV dengan meniadakan fitur tangkapan layar, yang terbukti efisien menekan utilisasi RAM di angka 33% dan beban SSD 0%. Hal ini sangat mempermudah divisi Quality Control (QC) dalam mengevaluasi kinerja pada lini pengemasan. Sistem ini telah memenuhi tujuan perancangan dan siap mendukung peningkatan keandalan proses quality control di industri.
===================================================================================================================================
Quality control on the production line at PT SC Johnson Manufacturing Surabaya specifically during the process of packing folding box into shipper box currently still relies on manual visual inspection. This process is prone to the risk of underfill. The failure rate due to discrepancies in the number of products inside the cartons which stands at 16.6% remains an operational challenge in maintaining the smooth flow of the production line. Therefore, an automated inspection system capable of operating consistently and in real time is needed to minimize human error. This study develops a deep learning-based visual inspection system using the Real-Time Detection Transformer (RT-DETR) method to automatically detect and count folding boxes. This model was chosen for its ability to perform end-to-end object detection without the Non-Maximum Suppression (NMS) process, as well as its superior performance in handling dense and overlapping objects, thereby preventing missed detections. The system is designed to classify cardboard boxes specifically into four operational conditions: empty, underfilled, full, and closed. The performance of the RT-DETR model is then compared with the YOLOv8 algorithm, which is used as the evaluation baseline. This system was implemented using a camera as a visual sensor and an industrial PC as the processing unit, integrated into the manual roller conveyor line leading to the sealer machine. Test results show that the system successfully identified all four carton conditions based on performance evaluation; RT-DETR operates with high precision, achieving an mAP50-95 metric value of 0.969, a field condition classification accuracy of 99.44%, and an average stable speed of 37 FPS. As a safety interlock measure, if the camera detects a violation such as a carton that is not fully filled but has already been sealed (the sequence skips from “Underfilled” to “Sealed”) the system triggers a warning buzzer and automatically shuts down the sealer. This preventive mechanism has proven effective in preventing defective products from entering the system, even if an operator accidentally continues to feed the carton. All inspection activities are automatically recorded in a CSV text report without the need for screenshots, which has proven efficient in reducing RAM usage to 33% and SSD load to 0%. This greatly simplifies the Quality Control (QC) department’s evaluation of performance on the packaging line. The system has met its design objectives and is ready to support improved reliability of quality control processes across the industry.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Kata kunci: Deep Learning, Inspeksi Visual, Pengemasan, RT-DETR, Safety Interlock. Keywords: Visual Inspection, YOLO, Occlusion, Sealer Machine, Folding Box Packaging
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7882.P3 Pattern recognition systems
Divisions:	Faculty of Vocational > 36304-Automation Electronic Engineering
Depositing User:	Prayogi Galih Anggono
Date Deposited:	11 Aug 2026 02:49
Last Modified:	11 Aug 2026 02:49
URI:	http://repository.its.ac.id/id/eprint/144295

Actions (login required)

View Item