Hierarchical Multi-Label Classification of Building Management System Time-Series Data

The Brick by Brick (BBB) challenge focuses on automating the classification of building management system (BMS) time-series data to enhance energy efficiency and operational insights. This dataset presents a multi-label hierarchical classification problem, requiring models to capture dependencies among labels. Our solution involves extracting statistical and spectral features from raw time-series data and expanding the training set using a splitting-based augmentation technique, increasing its size 28-fold. Labels are organised into five hierarchical tiers, and we train an ensemble of Random Forest (RF) and XGBoost (XGB) models using 10-fold cross-validation and sequential tiered predictions. Our submission achieves a macro F1-score of 0.528 on the public + private data leaderboard, highlighting the effectiveness of hierarchical ensemble learning and data augmentation for multi-label classification in smart building applications.