One of the main concerns regarding the facilitation of the elderly well-being monitoring system is to preserve the participants’ privacy, enable the older adults to live longer independently, and support caregivers. Human Activity Recognition (HAR) in smart homes allows us to foresee the residents’ needs by identifying changes in behaviour that might link to possible health conditions. We propose a Federated learning (FL) model within the health monitoring application to generalize for diverse participant populations and to achieve comparable performance without disclosing the raw data to the traditional centralized approach, which raises privacy issues. In this study, we evaluate an unsupervised variational autoencoder (VAE) in centralized, individualized, compared to federated learning settings to learn the features of normal patterns of daily activities and build an anomaly detector based on reconstructed error resulting as outcomes by the trained model. Further, we validated our proposed approach on real-world datasets collected over three years from six single-resident elderly households. The individual and centralized-based learning models were used as a baseline to compare with FL. Our results show that the personalized FedAvg models achieve RMSE of about 1%, while the Global FL models achieve RMSE of approximately. 4%. The centralized model achieves RMSE of about 0.5%, and the RMSE of individual models based on local training ranges between 1% to 6%. The FL models are relatively comparable to the centralized baseline model.

In Human Activity Recognition (HAR) systems, activities of daily living/human behaviour are recognized using sensor data by applying data mining techniques and machine learning algorithms to the collected data. This allows for customised and automated services to support humans’ daily living. Along with recognizing human behaviour, HAR can provide insights into the abnormal behaviour of individuals that might link to possible health conditions. Health monitoring applications are widely used for patients with chronic diseases, especially to give feedback to the users and promote self-awareness, for example, persons with dementia who need constant monitoring to minimize the risk of undesirable events. In developing HAR applications and services along with anomaly detection, obstacles and challenges remain that need to be addressed and motivate our research focusing on older adults living in real-world conditions. The need for care has increased, especially with the demographic developments and growing population of 65 years and over. Some of the challenges identified are installing IoT devices and sensors, collecting data and maintaining the system in an uncontrolled environment, and identifying valuable features to extract from sensors while ensuring the development of a non-intrusive and privacy-preserving system. Our research uses smart home sensors to infer ADLs (e.g., eating, sleeping, and bathroom visits) for older adults in single-resident homes. The research aims to develop and validate an anomaly detection based IoT system within HAR to support older adults in living longer independently and for the relatives (caregivers) to provide the necessary support. In addition, the system will help address the challenges of the ageing population and the increased burden on healthcare resources. This thesis identifies technologies and datasets suitable for IoT-enabled smart healthcare applications to recognize near real-time and short- and long-term behavioural changes.

We propose to develop a life conditions model for each individual by understanding the routines and activities of daily living based on interviews with older adults and their caregivers and collecting datasets with a focus on motion sensors. The developed life condition model for each individual is used to recognize human behaviour (based on context information such as time and location) and to analyze overall behavioural change. Hence, we develop and build an anomaly detection system that preserves privacy for near real-time behavioural changes and supports large-scale deployment. Our research methodology follows a quantitative research methodology as well as a qualitative approach based on interviews. We defined activity models based on contextual information such as time and location to extract features suitable for inferring daily living activities, model behavioural patterns, and, after that, detect abnormal activities in each daily routine by utilizing motion sensors as suitable sensing devices for non-intrusive IoT enabled smart healthcare applications in single-resident homes. For example, we applied a statistical method to build a routine or habit model for each older adult. We utilized unsupervised clustering methods K-means and LOF and reinforcement learning sleeping to recognise sleep activity patterns and detect anomalies. Further, to recognise all variations of the person’s behaviour and detect short and long-term behavioural changes in older people’s daily behaviour, we applied LSTM and VAE algorithms. In addition, we developed an anomaly detection system that preserves privacy to support large-scale deployments by utilizing federated learning to build a generalizable model that learns from different persons models.

Sleep is an essential activity that affects an individual’s health and ability to perform Activities of Daily Living (ADL). Inadequate sleep reduces cognitive capacity and leads to health-related issues such as cardiovascular diseases. Sleep disorders are more prevalent in older adults. Therefore, it is essential to recognize sleep patterns and support older adults and their caregivers. In our study, we collect data in real-world unconstrained and non-intrusive environments. This paper presents a novel sleep activity recognition method using motion sensors for recognizing nighttime and daytime sleep, which can further enable the development of insightful healthcare applications. The research objectives are to evaluate the application of using Multi-Armed Bandit methods to (i) learn normal sleep patterns, (ii) evaluate sleep quality, and (iii) detect anomalies in sleep activity for 11 elderly participants living in single-resident smart homes. We evaluate the performance of Thompson Sampling, Random Selection, and Upper Confidence Bound MAB methods. Thompson Sampling outperformed the other two methods. Our findings show most elderly participants slept between 6 and 8 hours with 85% sleep efficiency and up to 3 awakenings per night.

Real-time anomaly detection in real-time energy consumption helps identify the technical and infrastructure issues and events that result in significant energy waste and provides the end users feedback to address the issues, which reduces drift operation costs and saves energy in the building. This study proposes deep learning reconstruction models to detect anomalies in daily energy consumption data for nine school buildings. We evaluated the performance of three proposed models, stacked RNN-LSTM autoencoder, CNN-LSTM autoencoder, and LSTM Variational Autoencoder (VAE), to learn the features of normal consumption in an unsupervised manner and detect anomalies based on reconstruction error. We used Exponential Moving Average (EMA) and static threshold to detect local and global anomalies. The experimental results demonstrate that the local CNN-LSTM autoencoder performs better than the local Stacked Autoencoder(AE), with RMSE values ranging between 8-13% for electricity and 11-19% for district heating compared to 12-17% and 15-34% resulting from AE model, respectively. Local LSTM-Variational Autoencoder (VAE) outperformed both methods, with RMSE 4-6% for electricity and 5-7% for district heating. LSTM-VAE trained model on grouped training datasets of schools with similar energy consumption and building profiles has improved the local model by lowering RMSE values to 2-3% for electricity and 3-4% in district heating.

As the ageing population increases, predictive health applications for the elderly can provide opportunities for more independent living, increase cost efficiency and improve the quality of health services for senior citizens. Human activity recognition within IoT-based smart homes can enable the detection of early health risks related to mild cognitive impairment by providing proactive measurements and interventions to both the elderly and supporting healthcare givers. In this paper, we developed and evaluated a Multivariate long short-term memory (LSTM) to learn to forecast activities of daily living and detect anomalous behaviour using motion sensor data in 6 single-resident smart homes of the elderly. We use Mahalanobis distance to identify anomalies based on distance scores to build thresholds. The model's performance in terms of NMAE error values ranges between 2\% and 6\%. The experimental results show that the performance of LSTM for predicting the direct next activity versus the multiple forecasts is close. The method could identify participants' changing health conditions through the used predictive model and unsupervised anomaly detection method.

The growing population and demand for new public buildings contribute to increased energy consumption and greenhouse emissions. In Sweden, the largest amount of energy is consumed in school buildings, i.e., where schools form the highest number of public properties (30 million m2). In total, schools consumed 4 222 GWh of district heating and about 3 GWh of electricity for heating and other purposes in 2020. These figures lead to the realization of the need to apply effective measures to meet the European Green Deal target for 2030. Accurately forecasting energy usage is important for all stakeholders to conduct economic analysis and optimize decision-making. It is equally important in maintenance operations to allocate resources and enable the staff and students to adjust their behaviours and address the issues in buildings where peak forecasts occur. This paper develops and evaluates a power and district heating consumption for a single day and multiple days forecasting using Multivariate Recurrent Neural Network (RNN)-Long-Short term memory (LSTM) and convolutional neural networks (CNNs) and Autoencoders (AE), using daily real consumption data of six public schools provided by Skelefteå municipality in Sweden. The experimental results demonstrate that the hybrid model CNN-LSTM has achieved good accuracy compared to others, with RMSE and nRMSE error between 18%-25% and 5%-6% for electricity, respectively, and between 20%-30% RMSE and 5% nRMSE for district heating.

IoT-enabled innovative elderly healthcare facilitated by machine learning (ML) can address the challenges pertaining to the global aging population. For instance, it can enable the early detection of debilitating conditions such as Alzheimer's and dementia. This paper addresses this challenge by developing IoT and ML-based methods to recognize changes in long-term activities of daily living (ADLs) that may lead to the conditions mentioned above. In particular, we gather real-world long-term (approx. three years) data from 6 real-life single-resident elderly smart homes in Sweden, equipped with motion sensors in each room; and use unsupervised ML methods incorporating K-means clustering and local outlier factor to recognize changes in long-term behaviour efficiently. Our results have shown that K-means show similar performance in identifying outliers over all datasets while local outlier factorization fluctuates more but is more sensitive to identify small changes in living conditions. We foresee that our methods to detect long-term behaviour changes can support caregivers in carrying out their assessment for discovering the early onset of health conditions, thereby preventing further progression and providing timely treatment.