IoT Sensor Data Collection and Visualization for Homecare Support of Older Adults

Sammanfattning

The growing ageing population has increased the need for scalable and non-intrusive approaches to homecare monitoring that preserve privacy and autonomy. Smart-home environments equipped with Internet of Things (IoT) sensors provide continuous behavioral data; however, real-world sensor streams are typically event-based, irregularly sampled, unlabeled, and subject to strict privacy and data governance requirements. These characteristics limit the applicability of supervised activity recognition methods and motivate the use of unsupervised, privacy-preserving approaches. This thesis proposes an end-to-end, label-free framework for collecting, processing, modeling, and visualizing daily behavioral patterns from raw ambient IoT sensor data collected in real residential environments. Data were obtained from five individuals living independently over a four-week period using motion and contact sensors. Raw sensor streams were temporally aligned onto a uniform two-second grid, encoded using feature-engineering-light preprocessing, and segmented into fixed-length, non-overlapping windows. Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer-based autoencoder models were trained independently for each user to learn personalized baseline routines. Behavioral deviations were quantified using reconstruction-error–based daily deviation ratios derived from training-period thresholds and visualized to support interpretable inspection of routine changes without clinical or diagnostic assumptions. Results show that unsupervised sequence reconstruction captures individual routine structure and highlights short-term deviations from learned baselines. Transformer models achieved superior reconstruction accuracy and deviation separation for most users, while recurrent architectures provided clearer deviation contrast for specific individuals, emphasizing the need for per-user model selection. Overall, the study demonstrates that raw, unlabeled IoT sensor data can be ethically structured and analyzed using unsupervised methods to support interpretable, privacy-aware monitoring of routine behavior in home environments.