Mahmudah, Haniah (2026) Optimisasi Performansi Sistem Deteksi Kerusakan Jalan Menggunakan Model Convolution Neural Network (CNN) Pada Edge Device. Doctoral thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 7009231001-Doctoral.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
7009231001-Doctoral.pdf - Accepted Version Restricted to Repository staff only Download (4MB) | Request a copy |
Abstract
Jalan adalah prasarana transportasi darat sebagai sarana mobilitas masyarakat mempunyai peranan sangat penting mendukung bidang ekonomi, sosial dan budaya. Kebutuhan pemeliharaan jalan secara berkala dan tepat serta mencegah kerusakan jalan berdasarkan pavement performance atau kinerja perkerasan. Perlu penilaian kondisi jalan digunakan sebagai acuan untuk menentukan jenis program penanganan kerusakan jalan. Salah satu metode untuk penilaian kondisi perkerasan jalan metode Surface Distress Index (SDI) yang melakukan pengukuran kondisi fungsional jalan oleh surveyor secara visual dilakukan secara konvensional. Hal ini menyebabkan ketidakakuratan dalam mengestimasi kondisi fungsional jalan sehingga berpengaruh pada ketidaktepatan pemilihan jenis penanganan jalan serta tidak efektif dari segi waktu, biaya, dan tenaga. Untuk itu membutuhkan pengembangan perangkat untuk survei kerusakan jalan yang digunakan pengambilan data bergerak secara real time. Beberapa penelitian kerusakan jalan menggunakan sensor akselerometer, sensor LiDAR, laser, sensor rekontruksi 2D dan 3D, sensor kamera dengan pengolahan image processing serta sensor kamera menggunakan Deep Learning yaitu model Convolutional Neural Network (CNN). Problem deteksi kerusakan jalan menggunakan sensor akselerometer, sensor LiDAR, laser, teknik pengolahan image processing, model machine learning mengekstrak gambar secara manual dan menyesuaikan parameter pemrosesan gambar. Perlu mengotomatisasi proses ekstraksi fitur dan klasifikasi secara bersamaan menggunakan model CNN. Solusi perangkat survei kerusakan jalan menggunakan model CNN pada edge device mempunyai storage yang terbatas dan pengambilan data bergerak dan real time. Hal ini membutuhkan akurasi tinggi dan efisiensi tinggi dengan waktu inferensi yang cepat.
Penelitian ini tentang deteksi kerusakan jalan menggunakan sensor kamera dan model CNN dengan mengoptimisasi performansi sistem mAP, AR dan F1-score dan efisiensi sistem berupa waktu inferensi yang cepat. Metode multilevel hyperparameter optimization yang menggabungkan optimasi Tree-Structured Parzen Estimator (TPE) dan Search Space (SS) menghasilkan performansi sistem model iYOLOV7 (improved YOLOV7) lebih baik dibandingkan model MobileNet V2, RetinaNet dan YOLOV7-tiny. Hasil penelitian menunjukkan bahwa iYOLOV7 memiliki performansi terbaik dengan nilai precision 0,986, recall 0,970, F1-score 0,978, mAP@0,50 sebesar 0,988, mAP@0,50:0,95 sebesar 0,806, serta loss terendah sebesar 0,031. Model iYOLOV7 juga memiliki skalabilitas ukuran yang mendekati YOLOV7-tiny sehingga layak diimplementasikan pada edge device. Hasil efisiensi sistem model iYOLOV7 mempunyai waktu inferensi yang cepat dibandingkan dengan model YOLOV7-tiny, MobileNet V2 dan RetinaNet152 sehingga sistem deteksi kerusakan jalan yang diusulkan dinilai akurat, efisien, dan aplikatif untuk pemantauan kondisi jalan secara real-time.
======================================================================================================================================
Roads are land transportation infrastructure that serve as a method of community movement and play a critical role in supporting the economic, social, and cultural areas, as well as many other facets of community life. There is a need for frequent and timely road maintenance to guarantee that the road is suitable for its age and to prevent further road damage based on pavement performance. Surveyors traditionally use the SDI method to visually measure the functional condition of roads. This causes mistakes in calculating the road's functional status, which leads to incorrect road treatment selection. Furthermore, it necessitates a significant amount of field surveyor labor and funding for road damage restoration, rendering it inefficient in terms of time, money, and energy. This necessitates the creation of a gadget for road damage surveys that leverages real-time mobile data capture. The difficulty of detecting road defects with accelerometer sensors, LiDAR sensors, lasers, image processing techniques, and machine learning models involves manually extracting images and adjusting image processing parameters. It is required to use the CNN model to automate both extraction features and classification. The challenge with road damage survey devices that use CNN models on edge devices is that they have limited storage and mobile and real-time data collection capabilities. This necessitates great precision, efficiency, and rapid inference time.
The current research investigates road damage detection utilizing video sensors and CNN models, with the goal of enhancing system performance for accuracy and speed of inference. To improve accuracy, optimize the CNN network with pre-trained MobileNet V2, RetinaNet, and YOLO models with multilevel hyperparameter optimization (M-HPO). This study focuses on road damage identification with camera sensors and CNN models, optimizing the performance of mAP, AR, and F1-score systems as well as system efficiency. The system performance is improved over the RetinaNet and YOLOV7-tiny models. In addition, it optimizes the YOLOV7 network with multilevel hyperparameter optimization (M-HPO) that combines Tree-Structured Parzen Estimator (TPE) optimization and search space optimization, which is improved YOLOV7 (iYOLOV7). The findings revealed that iYOLOV7 performed the best, with a precision of 0.986, recall of 0.970, F1-score of 0.978, mAP@0.50 of 0.988, mAP@0.50:0.95 of 0.806, and the lowest loss of 0.031. The iYOLOV7 model also has a size scalability similar to YOLOV7-tiny, making it appropriate for implementation on edge devices. The results of the system efficiency of the iYOLOV7 model have a fast inference time compared to the YOLOV7-tiny, MobileNet V2, and RetinaNet152 models so that the proposed road damage detection system is considered accurate, efficient, and applicable for real-time monitoring of road conditions.
| Item Type: | Thesis (Doctoral) |
|---|---|
| Uncontrolled Keywords: | kerusakan jalan, Convolutional Neural Network (CNN), optimisasi hyperparameter, post training, edge device,,Road damage, edge devices. |
| Subjects: | T Technology > T Technology (General) > T57.5 Data Processing T Technology > T Technology (General) > T58.8 Productivity. Efficiency |
| Divisions: | Faculty of Industrial Technology and Systems Engineering (INDSYS) > Physics Engineering > 30001-(S3) PhD Thesis |
| Depositing User: | Haniah Mahmudah |
| Date Deposited: | 07 Mar 2026 13:21 |
| Last Modified: | 11 Mar 2026 05:34 |
| URI: | http://repository.its.ac.id/id/eprint/132699 |
Actions (login required)
 |
View Item |