Winning Brick by Brick with Daily Slices: A 94-Task Unified XGBoost Solution for Brick Schema Classification

This paper introduces a unified solution for multilabel time-series classification in the Brick by Brick Challenge 2024, addressing the automation of building data standardization using IoT sensor streams. We propose a method combining feature engineering, label hierarchy compliance, and novel probability calibration to classify 94 Brick sub-classes from irregularly sampled data. The raw sensor data is initially segmented into daily intervals, followed by multi-domain feature engineering across four categories: statistical, temporal, spectral, and peak characteristics, yielding a comprehensive set of 78 engineered features. The multilabel task is transformed into a 91-class problem via label concatenation (-1/0/1). A key innovation is hierarchical probability calibration using a 64th-root transformation and aggregated averaging, suppressing extreme probabilities while preserving label dependencies. Our XGBoost model demonstrates consistent performance across evaluation sets, achieving macro F1 scores of 0.7366-0.7391 through stratified five-fold cross-validation on the validation set. The model attained a macro F1 score of 0.6127 on the public leaderboard, with an overall score of 0.544 when evaluated across both public and private test sets. Temporal difference features prove most discriminative. The calibration strategy enhances robustness against segmentation artifacts, aligning with competition metrics through prediction remapping (-1→0.1, 1→0.9). The solution's computational efficiency (14-hour GPU training and inference) and interpretable features enable scalable, real-world building energy management, advancing fine-grained equipment status classification for sustainable operations.