Pengembangan Multi-Task Learning Model Berbasis Vision Transformer Untuk Klasifikasi Dan Verifikasi Hasil Ukur Optical Power Meter Pada Dokumen Pekerjaan Fiber-to-the-Home PT. Telkom Indonesia

Santosa, Hafiz Akmaldi (2026) Pengembangan Multi-Task Learning Model Berbasis Vision Transformer Untuk Klasifikasi Dan Verifikasi Hasil Ukur Optical Power Meter Pada Dokumen Pekerjaan Fiber-to-the-Home PT. Telkom Indonesia. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5027221061_Undergraduate_Thesis.pdf Restricted to Repository staff only Download (10MB) | Request a copy

Abstract

Perluasan infrastruktur jaringan Fiber-to-the-Home (FTTH) yang masif oleh PT. Telkom Indonesia menuntut proses administrasi yang efisien dan akurat, khususnya dalam validasi dokumen Berita Acara Uji Terima (BAUT) dan Commissioning Test (BACT). Saat ini, proses verifikasi kesesuaian antara data tabel rekapitulasi hasil ukur dan lampiran foto fisik Optical Power Meter (OPM) masih dilakukan secara manual. Metode ini rentan terhadap kelelahan visual (visual fatigue) dan inefisiensi waktu, diperburuk oleh noise visual seperti silau cahaya (glare) dan latar belakang kompleks yang sulit ditangani oleh sistem Optical Character Recognition (OCR) tradisional. Penelitian ini mengusulkan solusi otomatisasi berbasis pemahaman dokumen visual (Visual Document Understanding) menggunakan implementasi two-stage pipeline. Pendekatan ini menggabungkan YOLOv11 sebagai pra-pemrosesan lokalisasi objek berkecepatan tinggi dengan model Vision Transformer (Florence-2) dalam kerangka kerja prompt-based multitask inference. Pemisahan fungsionalitas ini terbukti secara arsitektural mampu menekan kelemahan efek jungkat-jungkit (seesaw effect) yang umum terjadi pada optimasi Multi-Task Learning. Hasil pengujian menunjukkan bahwa pipeline yang dikembangkan unggul dibandingkan arsitektur OCR-free konvensional (Donut), di mana Florence-2 sukses mencapai akurasi klasifikasi dokumen sebesar 100% pada ratusan sampel, serta mendemonstrasikan ketangguhan ekstraksi di kondisi lapangan riil dengan akurasi 90,00% untuk label ODP dan 77,78% untuk layar OPM meskipun dihadapkan pada gangguan ekstrem (seperti pantulan cahaya dan coretan tulisan tangan). Selain itu, pengujian User Acceptance Test (UAT) mendemonstrasikan bahwa sistem otomasi ini mampu mengeksekusi logika verifikasi redaman dengan konsisten pada kecepatan rata-rata inferensi gabungan 1,5 detik per halaman, menjadikannya instrumen pendukung keputusan (Decision Support System) yang andal bagi PT Telkom Indonesia.
====================================================================================================================================
The massive expansion of Fiber-to-the-Home (FTTH) network infrastructure by PT. Telkom Indonesia demands efficient and accurate administrative processes, particularly in validating Acceptance Test Report (BAUT) and Commissioning Test (BACT) documents. Currently, the verification process for matching data between tabulated measurement results and physical evidence photos of Optical Power Meter (OPM) is still performed manually. This conventional method is prone to visual fatigue and time inefficiency, exacerbated by visual noises such as glare and complex backgrounds that are difficult to handle by traditional Optical Character Recognition (OCR) systems. This research proposes an automation solution based on Visual Document Understanding utilizing a two-stage pipeline implementation. The approach combines YOLOv11 as a high-speed object localization pre-processor with a Vision Transformer model (Florence-2) within a prompt-based multitask inference framework. This functional decoupling architecture is proven to suppress the seesaw effect commonly found in Multi-Task Learning optimization. Evaluation results demonstrate that the developed pipeline achieves absolute superiority over conventional OCR-free architectures (Donut), where Florence-2 successfully achieved 100% page classification accuracy and recorded near-zero extraction error rates (Character Error Rate and Mean Absolute Error) on both micro-scale table structures and OPM screens. Furthermore, field User Acceptance Test (UAT) demonstrated that this automation system is capable of consistently executing attenuation verification logic at an average inference speed of 1.5 seconds per page, thereby successfully serving as a reliable Decision Support System for PT Telkom Indonesia.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Document Understanding, Donut Model, Fiber To The Home (FTTH), Multi-Task Learning, Optical Power Meter, Vision Transformer
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. Q Science > Q Science (General) > Q337.5 Pattern recognition systems
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information Technology > 59201-(S1) Undergraduate Thesis
Depositing User:	Hafiz Akmaldi Santosa
Date Deposited:	28 Jul 2026 06:32
Last Modified:	28 Jul 2026 06:32
URI:	http://repository.its.ac.id/id/eprint/138643

Actions (login required)

View Item