Another Brick in the Wall: Leveraging Feature Extraction and Ensemble Learning for Building Data Classification

Abstract

Accurately classifying building time series data enables the identification of patterns, detecting anomalies, and the understanding of how people behave within a building and eventually leads to the automated control, predictive maintenance, and efficient energy management of smart buildings. The 2024 Brick by Brick competition aimed to create machine learning models that classify devices, which output building time series data according to the standardized Brick schema. Since manually labeling all these devices can require significant effort, the competition's goal was to automate this process. In this work, the competition's multi-label classification problem is reframed as a multiclass task in order to apply a classical multiclass classification approach. Our solution involves a structured pipeline where we extract statistical, temporal and frequency-domain features from various time intervals and apply feature selection techniques to enhance the model's efficiency. An extra-trees classifier, optimized using stratified k-fold cross-validation, forms the core of our model. Our method achieves a macro F1-score of 0.5767 on the Brick by Brick's public leaderboard test set, resulting in the 4th place and demonstrating its effectiveness in automating building device classification. By reducing manual effort and improving classification accuracy, our approach contributes to more scalable and intelligent building management systems.