Transfer learning-informed sensor validation for detecting and diagnosing unseen air quality faults in underground building environment

In modern underground building environments, data-driven integrated systems are necessary for accurate early warning and health risk assessment. However, sensor validation frameworks in both existing and newly deployed monitoring networks often face challenges due to data insufficiency and unseen fault scenarios, leading to increased energy consumption resulting from inaccurate ventilation control. To address these issues, this study proposes a sensor validation framework that integrates a gated residual network (GRN) with an autoencoder and network adapted transfer learning (TL) to ensure reliable performance under faulty conditions. Fault detection, diagnosis, and identification were initially performed on the source station using the squared prediction error and sensor validity index. Subsequently, a TL-based scheme was applied to adapt the model to the target station, mitigating the impact of data scarcity and unseen fault types on validation accuracy. Finally, the influence of faulty sensor measurements and AE-based TL-GRN reconstructions on ventilation control performance was assessed. The proposed TL-GRN achieved a 94.79% fault detection rate for unseen scenarios, significantly outperforming the GRN-AE (78.65%). Moreover, the proposed framework reduces overall resource usage and lowers carbon emissions by decreasing energy consumption by 13.8% and 19.9% compared to GRN-AE and the faulty condition, respectively. Overall, the proposed framework makes a significant contribution to the development of resilient and self-regulating ventilation systems for next-generation smart and sustainable buildings.