Distributed cloud-edge analytics and machine learning for transportation emissions estimation
(English) In recent years IoT and Smart Cities have become a popular paradigm of computing that is based on network-enabled devices connected providing different functionalities, from sensor measures to domotic actions. With this paradigm, it is possible to provide to the stakeholders near-realtime...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Fecha de publicación: | 2022 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/380817 |
| Acceso en línea: | https://hdl.handle.net/2117/380817 https://dx.doi.org/10.5821/dissertation-2117-380817 |
| Access Level: | acceso abierto |
| Palabra clave: | Àrees temàtiques de la UPC::Informàtica |
| Sumario: | (English) In recent years IoT and Smart Cities have become a popular paradigm of computing that is based on network-enabled devices connected providing different functionalities, from sensor measures to domotic actions. With this paradigm, it is possible to provide to the stakeholders near-realtime information of the field, e.g. the current pollution of the city. Along with the mentioned paradigms, Fog Computing enables computation near the sensors where the data is produced, i.e. Edge nodes. This paradigm provides low latency and fault tolerance given the possible independence of the sensor devices. Moreover, pushing this computation enables derived results in a near-realtime fashion. This ability to push the computation to where the data is produced can be beneficial in many situations, however it also requires to include in the Edge the data preparation processes that ensure the fitness for use of the data as the incoming data can be erroneous. Given this situation, Machine Learning can be useful to correct data and also to produce predictions of the future values. Even though there have been studies regarding on the uses of data at the Edge, to our knowledge there is no evaluation of the different modeling situations and the viability of the approach. Therefore, this thesis aims to evaluate the possibility of building a distributed system that ensures the fitness for use of the incoming data through Machine Learning enabled Data Preparation, estimates the emissions and predicts the future status of the city in a near-realtime fashion. We evaluate the viability through three contributions. The first contribution focuses on forecasting in a distributed scenario with road traffic dataset for evaluation. It provides a robust solution to build a central model. This approach is based on Federated Learning, which allows training models at the Edge nodes and then merging them centrally. This way the models in the Edge can be independent but also can be synchronized. The results show the trade-off between accuracy versions training time and a comparison between low-powered devices versus server-class machines. These analyses show that it is viable to use Machine Learning with this paradigm. The second contribution focuses on a particular use case of ship emission estimation. To estimate exhaust emissions data must be correct, which is not always the case. This contribution explores the different techniques available to correct ship registry data and proposes the usage of simple Machine Learning techniques to do imputation of missing or erroneous values. This contribution analyzes the different variables and their relationship to provide the practitioners with guidelines for correction and data treatment. The results show that with classical Machine Learning it is possible to improve the state-of-the-art results. Moreover, as these algorithms are simple enough, they can be used in an Edge device if required. The third contribution focuses on generating new variables from the ones available with a ship trace dataset obtained from the Automatic Identification System (AIS). We use a pipeline of two different methods, a Neural Networks and a clustering algorithm, to group movements into movement patterns or \emph{behaviors}. We test the predicting power of these behaviors to predict ship type, main engine power, and navigational status. The prediction of the main engine power is compared against the standard technique used in ship emission estimation when the ship registry is missing. Our approach was able to detect 45\% of the otherwise undetected emissions if the baseline method was to be used. As ship navigational status is prone to error, the behaviors found are proposed as an alternative variable based in robust data. These contributions build a framework that can distribute the learning processes and that resists network failures in low-powered devices. |
|---|