A Convolutional Transformer for Enhanced NILM in Human Activity Recognition

Monitoring human activities is crucial within Ambient Assisted Living (AAL) contexts, contributing significantly to diagnosis. Data collected from smart meters or home sensors allow the deduction of various Activities of Daily Living (ADL), including eating patterns, sleeping, and routine alterations, providing insights into overall health and well-being. Currently, activity recognition in AAL heavily relies on ambient sensors. Information regarding appliance usage offers valuable insights into health, including immobility, sleep disorders, and activity patterns. In this regard, Non-Intrusive Load Monitoring (NILM) systems represent an interesting alternative as they minimize installation invasiveness. This work proposes an innovative Convolutional Transformer model to enhance the performance of NILM systems based on the sequence-to-point approach, which can be applied to load identification and behavior pattern monitoring. The model has been validated and tested on a widely-used public dataset in the NILM context, and the metrics obtained have been compared with a state-of-the-art NILM algorithm.