Event Detection and Localization Using Machine Learning on a Staircase

Abstract

Recent years have seen a push towards smart buildings that are energy efficient and proactive in decision making by detecting building events. Instrumentation of structures with sensors such as accelerometers or thermocouples is an essential element for providing the building with the necessary capabilities to enhance the occupant’s comfort, safety and overall quality-of-life. As a result, vast amounts of data are collected that if correctly parsed can produce meaningful information for this purpose. One of the much-needed information about a building’s activities is event localization. In many situations, event localization, based on traditional wave propagation techniques associated with vibrations, is a challenging task in an active environment as there is little control over the noise concurrent with the event. Determining ways to process sensor data efficiently and effectively will make user interaction with the building more intuitive and enhance user experience.

The present work pursues and evaluates an in-situ machine learning based approach for detecting and localizing footsteps on an instrumented staircase. The first part of the algorithm takes in live data from three accelerometers on a staircase and identifies footsteps based on a spike in the signal-to-noise ratio based on power spectral densities. The second part of the algorithm is the localization of the footstep once it is detected. Additionally, the performance of various features extracted from the time data (collected through controlled experiments) to generate an accurate machine learning model is also part of the current work. A nested tree algorithm is developed which yields 87% accuracy, showing potential for future stand-alone applications.