ML4IoT: A Framework to Orchestrate Machine Learning Workflows on Internet of Things Data

Self Cite

It is well known that power plants worldwide present access to difficult and hazardous environments, which may cause harm to on-site employees. The remote and autonomous operations in such places are currently increasing with the aid of technology improvements in communications and processing hardware. Virtual and augmented reality provide applications for crew training and remote monitoring, which also rely on 3D environment reconstruction techniques with near real-time requirements for environment inspection. Nowadays, most techniques rely on offline data processing, heavy computation algorithms, or mobile robots, which can be dangerous in confined environments. Other solutions rely on robots, edge computing, and post-processing algorithms, constraining scalability, and near real-time requirements. This work uses an edge-fog computing architecture for data and processing offload applied to a 3D reconstruction problem, where the robots are at the edge and computer nodes at the fog. The sequential processes are parallelized and layered, leading to a highly scalable approach. The architecture is analyzed against a traditional edge computing approach. Both are implemented in our scanning robots mounted in a real power plant. The 5G network application is presented along with a brief discussion on how this technology can benefit and allow the overall distributed processing. Unlike other works, we present real data for more than one proposed robot working in parallel on site, exploring hardware processing capabilities and the local Wi-Fi network characteristics. We also conclude with the required scenario for the remote monitoring to take place with a private 5G network.

Section: Proposed Architecturementioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

An Edge-Fog Architecture for Distributed 3D Reconstruction and Remote Monitoring of a Power Plant Site in the Context of 5G

Vidal

Honório

Pinto

et al. 2022

Self Cite

“…In this sense, parallel execution in a distributed environment is evaluated. This is an innovative aspect of this work, where part of the code is executed onboard the robot and part of it in the fog close to the robot in an edge-fog scheme [ 39 ].…”

Section: Introductionmentioning

confidence: 99%

Dynamic Optimization and Heuristics Based Online Coverage Path Planning in 3D Environment for UAVs

Melo

Pinto

Marcato

et al. 2021

Self Cite

Path planning is one of the most important issues in the robotics field, being applied in many domains ranging from aerospace technology and military tasks to manufacturing and agriculture. Path planning is a branch of autonomous navigation. In autonomous navigation, dynamic decisions about the path have to be taken while the robot moves towards its goal. Among the navigation area, an important class of problems is Coverage Path Planning (CPP). The CPP technique is associated with determining a collision-free path that passes through all viewpoints in a specific area. This paper presents a method to perform CPP in 3D environment for Unmanned Aerial Vehicles (UAVs) applications, namely 3D dynamic for CPP applications (3DD-CPP). The proposed method can be deployed in an unknown environment through a combination of linear optimization and heuristics. A model to estimate cost matrices accounting for UAV power usage is proposed and evaluated for a few different flight speeds. As linear optimization methods can be computationally demanding to be used on-board a UAV, this work also proposes a distributed execution of the algorithm through fog-edge computing. Results showed that 3DD-CPP had a good performance in both local execution and fog-edge for different simulated scenarios. The proposed heuristic is capable of re-optimization, enabling execution in environments with local knowledge of the environments.

“…The number of devices required in these reconstruction processes could be exceptionally large and with different types of architectures. Therefore, an orchestration is necessary to gather [ 17 ] information from these devices and to produce a synchronized differentiated result from these nodes. This issue is common in IoT architectures, where many devices with limited computational power are distributed along with the network [ 18 , 19 , 20 ].…”

Section: Introductionmentioning

confidence: 99%

A Heterogeneous Edge-Fog Environment Supporting Digital Twins for Remote Inspections

Silva

Vidal

Honório

et al. 2020

Self Cite

The increase in the development of digital twins brings several advantages to inspection and maintenance, but also new challenges. Digital models capable of representing real equipment for full remote inspection demand the synchronization, integration, and fusion of several sensors and methodologies such as stereo vision, monocular Simultaneous Localization and Mapping (SLAM), laser and RGB-D camera readings, texture analysis, filters, thermal, and multi-spectral images. This multidimensional information makes it possible to have a full understanding of given equipment, enabling remote diagnosis. To solve this problem, the present work uses an edge-fog-cloud architecture running over a publisher-subscriber communication framework to optimize the computational costs and throughput. In this approach, each process is embedded in an edge node responsible for prepossessing a given amount of data that optimizes the trade-off of processing capabilities and throughput delays. All information is integrated with different levels of fog nodes and a cloud server to maximize performance. To demonstrate this proposal, a real-time 3D reconstruction problem using moving cameras is shown. In this scenario, a stereo and RDB-D cameras run over edge nodes, filtering, and prepossessing the initial data. Furthermore, the point cloud and image registration, odometry, and filtering run over fog clusters. A cloud server is responsible for texturing and processing the final results. This approach enables us to optimize the time lag between data acquisition and operator visualization, and it is easily scalable if new sensors and algorithms must be added. The experimental results will demonstrate precision by comparing the results with ground-truth data, scalability by adding further readings and performance.