Machine learning techniques for sensor-based household activity recognition and forecasting - PhD thesis abstract

Aug 4, 2026

Upload by Anda

Recent progress in inexpensive, unobtrusive smart-home sensors enables ambient assisted living systems that can support everyday routines. Effective operation requires continuous monitoring and accurate forecasting of activities of daily living. This thesis develops machine learning and deep learning techniques for household activity signal data to support activity recognition and energy-related decisions. The work targets action recognition and energy consumption optimization, leveraging sensor signals to build intelligent, future-facing systems for efficient use of user and community resources.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages113

Reading Time4 h 45 min

AccessFree to download

Contents

Introduction

State of the art

Exploiting feature selection in HAR

Ph. D . DEGREE IN MATHEMATICS AND COMPUTER SCIENCE Cycle XXXV TITLE OF THE Ph. D . THESIS Machine learning techniques for sensor-based household activity recognition and forecasting Scientific Disciplinary Sector(s) INF/01 Ph. D. Student: Supervisor Co-Supervisor Marco Manolo Manca Prof. Daniele Riboni Prof. Ernesto Bonomi Final exam. Academic Year 2021/2022 Thesis defence: April 2023 Session Abstract Thanks to the recent development of cheap and unobtrusive smart-home sensors, ambient assisted living tools promise to offer innovative solutions to support the users in carrying out their everyday activities in a smoother and more sustainable way. To be effective, these solutions need to constantly monitor and forecast the activities of daily living carried out by the inhabitants. The Machine Learning field has seen significant advancements in the development of new techniques, especially regarding deep learning algorithms. Such techniques can be successfully applied to household activity signal data to benefit the user in several applications. This thesis therefore aims to produce a contribution that artificial intelligence can make in the field of activity recognition and energy consumption. The effective recognition of common actions or the use of high-consumption appliances would lead to user profiling, thus enabling the optimisation of energy consumption in favour of the user himself or the energy community in general. Avoiding wasting electricity and optimising its consumption is one of the main objectives of the community. This work is therefore intended as a forerunner for future studies that will allow, through the results in this thesis, the creation of increasingly intelligent systems capable of making the best use of the user’s resources for everyday life actions. Namely, this thesis focuses on signals from sensors installed in a house: data from position sensors, door sensors, smartphones or smart meters, and investigates the use of advanced machine learning algorithms to recognize and forecast inhabitant activities, including the use of appliances and the power consumption. The thesis is structured into four main chapters, each of which represents a contribution regarding Machine Learning or Deep Learning techniques for addressing challenges related to the aforementioned data from different sources. The first contribution highlights the importance of exploiting dimensionality reduction techniques that can simplify a Machine Learning model and increase its efficiency by identifying and retaining only the most informative and predictive features for activity recognition. In more detail, it is presented an extensive experimental study involving several feature selection algorithms and multiple Human Activity Recognition benchmarks containing mobile sensor data. In the second contribution, we propose a machine learning approach to forecast future energy consumption considering not only past consumption data, but also context data such as inhabitants’ actions and activities, use of household appliances, interaction with furniture and doors, and environmental data. We performed an experimental evaluation with real-world data acquired in an instrumented environment from a large user group. Finally, the last two contributions address the Non-Intrusive-Load-Monitoring problem. In one case, the aim is to identify the operating state (on/off) and the precise energy consumption of individual electrical loads, considering only the aggregate consumption of these loads as input. We use a Deep Learning method todisaggregate the low-frequency energy signal generated directly by the new generation smart meters being deployed in Italy, without the need for additional specific hardware. In the other case, driven by the need to build intelligent non-intrusive algorithms for disaggregating electrical signals, the work aims to recognize which appliance is activated by analyzing energy measurement

FAQs

What problem does the thesis address in smart homes?

It addresses continuous monitoring and forecasting of daily living activities, along with improving energy consumption decisions for users and energy communities.

Which sensor signals does the thesis focus on?

It focuses on data from position sensors, door sensors, smartphones, and smart meters, using advanced machine learning to recognize and forecast inhabitant activities and related power usage.

How does the thesis approach non-intrusive load monitoring (NILM)?

It investigates two NILM settings: disaggregating aggregate consumption to infer individual load on/off and energy consumption, and classifying which appliance is activated using combined single-label and multi-label learning with an event detector.

Aug 4, 2026

Upload by Anda