Accurate, real-time gas detection is crucial for applications from environmental monitoring to industrial processes, yet sensors are limited by low accuracy, slow response, and drift — the same gas reads differently months later. This project develops scalable machine learning for both problems. Encoder–decoder architectures and a decision-fusion model consolidate the strengths of multiple carbon dioxide sensors into a single virtual sensor, achieving a mean absolute percentage error (MAPE) of 2.97% while reducing response and recovery times from approximately 9 minutes to 2 minutes. A second, label-free framework compensates long-term drift on a public 36-month, 16-sensor array benchmark: it identifies the gas and reports its concentration using no labeled data from the drifted period, raising mean identification accuracy from 81% to 92% while also improving concentration accuracy. The architecture extends to other gases and other time-series sensing applications, pointing toward instruments that stay accurate without periodic recalibration.
Project is currently funded by: Member Fees