A Collective Learning Workflow for Remaining Useful Life Estimation of Building Assets Under Sparse Degradation Data

Abstract

Building maintenance in the commercial sector is still dominated by reactive and schedule-based strategies, yet the shift to predictive maintenance is held back by a chronic shortage of degradation data: individual assets such as fan coil units, air handling units and pumps rarely accumulate enough failure or degradation history to train standalone prognostic models, and the records that do exist are short and heavily skewed towards normal operation. Existing approaches do not resolve this. Deep sequence models require continuous multi-year recordings that newly instrumented buildings cannot provide and overfit on small, imbalanced datasets; semantic ontologies organise building data but carry no prognostic capability; and baseline forecasting methods generate operating features without estimating degradation. This paper presents a collective-learning workflow that groups functionally identical assets by their Brick semantic class, pools their operational records into a shared dataset, and trains a single Random Forest model to estimate each asset's daily degradation increment, which is then accumulated into a remaining-useful-life projection and updated as maintenance is recorded. Applied to operational data from a pilot building, the workflow produced per-unit daily degradation estimates and remaining-useful-life projections with quantified uncertainty for fan coil units that individually lacked sufficient history to be modelled in isolation. Under leave-one-out cross-validation, operational features alone did not recover the per-unit degradation level (cross-validated R² below zero), whereas fusing a static condition indicator with the operational dynamics was required to do so (R² = 0.67, CVRMSE 20 percent), reported as a preliminary finding rather than a validated prognostic capability. The results show that pooling sparse degradation labels across semantically aligned assets makes data-driven prognostics feasible for building portfolios under data scarcity, providing a transferable route to predictive maintenance that does not depend on multi-year failure records.