Multi Resolution Spatial Attention-U-Net (MRSA-U-NET) : Model Segmentasi Retak Untuk Pemantauan Kesehatan Struktur Beton

PRASETYO, ANDREW (2026) Multi Resolution Spatial Attention-U-Net (MRSA-U-NET) : Model Segmentasi Retak Untuk Pemantauan Kesehatan Struktur Beton. Doctoral thesis, Institut Teknologi Sepuluh Nopember (ITS).

Text 7022231006-Doctoral.pdf Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Retakan pada struktur beton merupakan indikator kritis yang mempengaruhi kekuatan, durabilitas, dan keamanan infrastruktur. Standar internasional seperti SNI 2847:2019, ACI 224R-01, dan Eurocode 2 telah menetapkan batasan lebar retakan maksimum antara 0,1–0,4 mm tergantung kondisi lingkungan. Deteksi dini retakan sangat penting untuk mencegah kegagalan struktural dan mengelola biaya pemeliharaan secara efektif. Penelitian ini mengembangkan MRSA-U-NET, sebuah arsitektur deep learning yang dioptimasi untuk deteksi retakan kecil sesuai standar internasional. Model ini mengintegrasikan blok residual, Convolutional Block Attention Module (CBAM), dan Atrous Spatial Pyramid Pooling (ASPP) dalam kerangka kerja U-Net. Inovasi utama penelitian ini meliputi: (1) implementasi optimizer AdamW yang meningkatkan proses minimisasi fungsi loss melalui weighted averages dari gradien; (2) perbandingan komprehensif enam optimizer populer sebagai benchmark untuk studi masa depan; (3) pengembangan DBCE Loss yang mengintegrasikan Binary Cross-Entropy dengan Dice Loss untuk mencapai akurasi pixel-wise yang superior; dan (4) arsitektur MRSA-U-NET yang mampu mendeteksi retakan kecil dengan presisi tinggi. Evaluasi komprehensif dilakukan pada empat dataset benchmark (CFD, Crack500, DeepCrack, dan GAPS), yang menunjukkan peningkatan signifikan pada berbagai metrik kritis. Hasil penelitian menunjukkan bahwa optimizer Adam dan RMSProp mencapai akurasi tertinggi (97,25%), sementara AdamW menonjol sebagai pilihan yang paling robust dengan nilai loss terendah dan skor Tversky Index tertinggi. Fungsi kerugian DBCE Loss secara nyata meningkatkan kinerja segmentasi dengan akurasi, koefisien Dice, dan skor F1 yang lebih baik pada berbagai dataset. Arsitektur MRSA-U-NET terbukti sangat efektif dalam mendeteksi dan mensegmentasi retakan kecil dengan nilai F1 Score dan IoU yang superior di seluruh dataset, sehingga memfasilitasi deteksi dini masalah struktural potensial dan memungkinkan intervensi pemeliharaan yang tepat waktu. Kontribusi penelitian ini memberikan dampak signifikan terhadap keselamatan publik melalui deteksi dini untuk mencegah kegagalan struktural, efisiensi ekonomi melalui pemeliharaan preventif yang lebih cost-effective, keberlanjutan infrastruktur melalui perpanjangan umur layanan, optimalisasi manajemen aset berbasis data, serta pengembangan standar internasional untuk pemantauan infrastruktur. Penelitian ini berpotensi mentransformasi praktik manajemen infrastruktur modern dan meningkatkan keandalan sistem infrastruktur secara keseluruhan.
================================================================================================================================
Cracks in concrete structures are critical indicators affecting the strength, durability, and safety of infrastructure. International standards such as SNI 2847:2019, ACI 224R-01, and Eurocode 2 have established maximum crack width limits ranging from 0.1–0.4 mm depending on environmental conditions. Early crack detection is essential to prevent structural failures and effectively manage maintenance costs. This research develops MRSA-U-NET, an optimized deep learning architecture for detecting fine cracks in accordance with international standards. The model integrates residual blocks, Convolutional Block Attention Module (CBAM), and Atrous Spatial Pyramid Pooling (ASPP) within the U-Net framework. Key innovations include: (1) implementation of the AdamW optimizer that enhances the loss function minimization process through weighted averages of gradients; (2) comprehensive comparison of six popular optimizers as a benchmark for future studies; (3) development of DBCE Loss that integrates Binary Cross-Entropy with Dice Loss to achieve superior pixel-wise accuracy; and (4) the MRSA-U-NET architecture capable of detecting fine cracks with high precision. Comprehensive evaluation was conducted on four benchmark datasets (CFD, Crack500, DeepCrack, and GAPS), demonstrating significant improvements in critical metrics. Results show that Adam and RMSProp optimizers achieved the highest accuracy (97.25%), while AdamW stood out as the most robust choice with the lowest loss and highest Tversky Index score. The DBCE Loss function significantly improved segmentation performance with superior accuracy, Dice coefficient, and F1 scores across various datasets. The MRSA-U-NET architecture proved highly effective in detecting and segmenting fine cracks with superior F1 Score and IoU values across all datasets, facilitating early detection of potential structural problems and enabling timely maintenance interventions. Research contributions provide significant impacts on public safety through early detection to prevent structural failures, economic efficiency through more cost-effective preventive maintenance, infrastructure sustainability through extended service life, data-driven asset management optimization, and the development of international standards for infrastructure monitoring. This research has the potential to transform modern infrastructure management practices and improve overall infrastructure system reliability.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	Deteksi Retakan, Deep Learning, U-Net, Segmentasi, CBAM, ASPP, Pemantauan Kesehatan Struktural, Crack Detection, Deep Learning, U-Net, Segmentation, CBAM, ASPP, Structural Health Monitoring
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) > Electrical Engineering > 20001-(S3) PhD Thesis
Depositing User:	Andrew Prasetyo
Date Deposited:	10 Jun 2026 01:05
Last Modified:	10 Jun 2026 01:05
URI:	http://repository.its.ac.id/id/eprint/133560

Actions (login required)

View Item