Validation of Subway Indoor Air Quality (IAQ) Data Using Memory-Augmented Autoencoders with Learned Normal Prototypes

Indoor air quality (IAQ) monitoring in subway stations depends on sensors prone to failures due to confined spaces, cyberattacks, and prolonged use. Soft sensor validation frameworks using statistical or machine learning models can detect, diagnose, and reconstruct faulty data but struggle with complex fault patterns. This study introduces a memory-augmented autoencoder-based framework for reliable IAQ sensor validation, leveraging memorized normal prototypes. To the best of our knowledge, this is the first validation method that utilizes normal prototypes for reconciling corrupted measurements. Tested on real IAQ data from Seoul Metro's C-station, the method achieved a 97.03% detection rate, a 4.33% false alarm rate, and demonstrated potential for 10.25% energy savings while maintaining healthy IAQ.