Master's Thesis

Throughout the project we had master's students writing their thesis on topics linked to the AI4IoT pilot.

2019 / 2020

Daniel Svendsen: A User-Based Look at Visualisation Tools for Air Quality Data harvested by IoT (microsensor units)

Abstract: The aim with this thesis is to perform a user-based look at visualization tools for air quality data harvested by IoT units in Trondheim. Several visualisations, including time-series graphs, a 2D heat-map and a noble 3D simulated environment have been developed and tested on 9 selected users. The scope of this project have been users in the three categories: citizens, researchers and policy makers. In addition three low-cost micro sensors, developed by Exploratory Engineering, was used to gather air quality data over a period of two months. Two of the sensors where mounted on stationary locations, at Lerkendal and Voll, and the third was placed on top of a moving bus. All of the sensors was monitoring in real-time and automatically transmitting the results to cloud server via Narrow Band IoT. In Norway, particulate matter (PM$_{2.5}$ and PM$_{10}$) and nitrogen dioxide (NO$_2$) are the most important components of local air pollution. Other pollutants such as carbon monoxide (CO), sulphur dioxide (SO$_2$) and ground level ozone (O$_3$) can also contribute to poor air quality, and can cause serious health problems for humans, animals and vegetation. Our aim is to visualise the complex "invisible" air pollution data, such that citizens, researches and policy makers can take a decision, to improve a routine or to change a method towards reducing the emissions of harmful gases and particles. We have reviewed related work of air quality visualisation, and projects that include low-cost air quality sensors. Further we developed three types of visualisation platforms, including a line-graph dashboard, a 2D heat-map and a 3D heat-map. Finally our visualisations was validated by a group of users through multiple video interviews. Our findings from the experiment performed shows that mobile air quality sensors are prone to power and connectivity failure. We also discovered that a 2D heat-map visualisation is the preferred way to present the data among all three user groups in a decision making context. All code developed for our visualisations is published online on GitHub:

Link to the thesis: TBA

Erling Ljundgren: Deep Learning for Blind Calibration of Wireless Sensor Networks - A comparative study of convolutional and recurrent neural networks

Abstract: Temporal drift of low-cost sensors is a crucial problem when considering the applicability of wireless sensor networks (WSN). Since they provide highly local measurements, which is key to combat the ever increasing problem of air pollution, calibrating such networks effectively becomes a high priority. The emergence of wireless sensor networks in locations without available reference data makes calibrating such networks without the aid of true values a key area of research. While deep learning (DL) has proved successful on numerous other tasks, it is sorely under-researched in the context of WSN calibration. To further this research, this thesis will explore the applicability of DL for blind WSN calibration by improving upon the only previously existing DL model and explore other possible models. Promising architectures are found by a structured literature search on DL methods in other related fields. To test architectures, a synthetic dataset has been implemented after analysing real sensor data. The new models presented in this thesis obtains a smaller calibration error with an order of magnitude compared to the previous model, with temporal convolutions in 2 dimensions proving most promising. All code used in this thesis is available at: TBA

Link to the thesis: TBA

Mykhaylo Marfeychuk: Learning Control Policies in Smart Cities from Physical Data

Abstract: Motor vehicle emissions are the primary contributors to the increase in ambient pollution levels. Rapid urbanization and lack of a good solution to manage the traffic are forcing cities to take drastic measures against the automotive industry. In this thesis, we build a case study of the Norwegian city of Trondheim’s traffic, and create a realist simulation based on real world data, that simulates the traffic and the emissions. We then propose a Reinforcement Learning based solution that controls the access to the different regions of the city to optimize the traffic given a desired metric. We also take a look at different improvements, like using a multi-agent system and using pre-generated data for the training phase. We compare the obtained results with the baseline and against a reactive agent. At the end we assess the solution’s strengths and weaknesses, and propose possible future improvements.

All code developed for our visualisations is published online on GitHub:

Links: thesis and presentation

2018 / 2019

Andreas Jacobsen Lepperød: Air Quality Prediction with Machine Learning

Abstract: In recent years, air quality has become a significant environmental health issue due to rapid urbanization and industrialization. Because of the impact air quality has on people’s everyday life, how to predict air quality precisely, has become an urgent and essential problem. Air quality prediction is a challenging problem with several complicated factors with additional dependencies among them. We target our air prediction study to the city of Trondheim, Norway. The air quality in Trondheim is on average at a healthy level, but has periods of high variations of severe pollution, especially in the winter months. The study demonstrates the benefits of machine learning for predicting air pollutants general pattern, and to foresee sudden spikes of a high pollution level. This paper explores a multivariate time series approach to modeling and forecasting the pollution of PM2.5, PM10, and NO2 at three air quality stations. This study is concerned with combining data of pollutants, meteorological, and traffic data with statistical temporal-spatial feature engineering, to provide multi-step-ahead air quality forecasts for 24 and 48-hours. Extensive experiments of real-time air pollution illustrate the effectiveness of machine learning to forecast air pollutions in terms of general pattern and sudden changes. Results express that ensemble techniques could significantly improve the stability and accuracy of predicting the general trend of air quality. Among the ensemble techniques, using gradient boosting with dropouts results in prediction errors with the lowest deviation. In the case of predicting sudden changes in air pollution, using a recurrent neural network with a memory unit results in the highest accuracy of classified spikes. Lastly, the machine learning results were compared with the national air quality service, a knowledge-driven model, to evaluate real-world practice. The predictions of general pattern and anomalies of this thesis are shown to be superior for 24-hour, and more comparable results for the 48-hour forecast. The data-driven approach is thus believed to be an excellent complement for the knowledge-driven model.

Link to the thesis: