https://scholars.lib.ntu.edu.tw/handle/123456789/611422
標題: | Pruning deep neural networks for efficient edge computing in internet of things: A structural health monitoring case study | 作者: | Wu R.-T. Singla A. Jahanshahi M.R. Bertino E. RIH-TENG WU |
關鍵字: | Antennas;Corrosion;Damage detection;Deep neural networks;Edge computing;Iterative methods;Learning algorithms;Motion planning;Structural health monitoring;Surface defects;Analysis capabilities;Civil infrastructures;Computing platform;Condition assessments;Detection performance;Internet of Things (IOT);Resource efficiencies;Resource-efficient;Internet of things | 公開日期: | 2019 | 卷: | 2 | 起(迄)頁: | 3170-3177 | 來源出版物: | Structural Health Monitoring 2019: Enabling Intelligent Life-Cycle Health Management for Industry Internet of Things (IIOT) - Proceedings of the 12th International Workshop on Structural Health Monitoring | 摘要: | Condition assessment of civil infrastructures is a key application of Internet of Things (IoT) while edge computing is a crucial component of IoT. In this context, swarms of autonomous inspection robots (e.g., unmanned aerial vehicles (UAVs)), that can replace current labor-intensive manual inspections, are examples of edge devices. Such devices need to have on-board data analysis capabilities for efficient inspection path planning. The incorporation of deep learning algorithms into these edge devices for damage detection is a challenging problem since these devices are typically limited in computing and memory resources. For instance, deep transfer learning, which leverages the advantages of pre-trained models, is usually used to detect a new type of damage when there is no sufficient training data. However, the pre-trained models are typically large in size and therefore are not resource efficient for those edge devices. This study introduces a solution to achieve quick inference and low memory demands through transfer learning and network pruning. Network pruning, inspired by the biological brain behavior, eliminates the redundant neurons of the pre-trained network with an iterative pruning and fine-tuning process. Results from comprehensive experiments on two pre-trained networks (i.e., VGG16 and ResNet18) and two types of prevalent surface defects (i.e., crack and corrosion) are presented and discussed in details with respect to damage detection performance, memory demands, and the inference time for damage detection. The experiments show that with the NVIDIA Jetson TX2 GPU, simulating the on-board computing platform, the proposed approach achieves an inference time which is nine and four times faster than the original VGG16 and ResNet18 networks, respectively. Also, the network size is reduced by 80% and 95% for the VGG16 and ResNet18 networks, respectively. Results demonstrate that the proposed approach significantly enhances resource efficiency for field applications without decreasing damage detection performance. ? 2019 by DEStech Publications, Inc. All rights reserved. |
URI: | https://www.scopus.com/inward/record.uri?eid=2-s2.0-85074248467&doi=10.12783%2fshm2019%2f32475&partnerID=40&md5=fa590b0f9c8753c556660c39a5d99847 https://scholars.lib.ntu.edu.tw/handle/123456789/611422 |
DOI: | 10.12783/shm2019/32475 |
顯示於: | 土木工程學系 |
在 IR 系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。