Bayesian based design of real-time sensor systems for high-risk indoor contaminants
- Author(s): Sreedharan, Priya
- et al.
The sudden release of toxic contaminants that reach indoor spaces can be hazardous to building occupants. To respond effectively, the contaminant release must be quickly detected and characterized to determine unobserved parameters, such as release location and strength. Characterizing the release requires solving an inverse problem. Designing a robust real-time sensor system that solves the inverse problem is challenging because the fate and transport of contaminants is complex, sensor information is limited and imperfect, and real-time estimation is computationally constrained. This dissertation uses a system-level approach, based on a Bayes Monte Carlo framework, to develop sensor-system design concepts and methods. I describe three investigations that explore complex relationships among sensors, network architecture, interpretation algorithms, and system performance. The investigations use data obtained from tracer gas experiments conducted in a real building. The influence of individual sensor characteristics on the sensor-system performance for binary-type contaminant sensors is analyzed. Performance tradeoffs among sensor accuracy, threshold level and response time are identified; these attributes could not be inferred without a system-level analysis. For example, more accurate but slower sensors are found to outperform less accurate but faster sensors. Secondly, I investigate how the sensor-system performance can be understood in terms of contaminant transport processes and the model representation that is used to solve the inverse problem. The determination of release location and mass are shown to be related to and constrained by transport and mixing time scales. These time scales explain performance differences among different sensor networks. For example, the effect of longer sensor response times is comparably less for releases with longer mixing time scales. The third investigation explores how information fusion from heterogeneous sensors may improve the sensor-system performance and offset the need for more contaminant sensors. Physics- and algorithm-based frameworks are presented for selecting and fusing information from noncontaminant sensors. The frameworks are demonstrated with door-position sensors, which are found to be more useful in natural airflow conditions, but which cannot compensate for poor placement of contaminant sensors. The concepts and empirical findings have the potential to help in the design of sensor systems for more complex building systems. The research has broader relevance to additional environmental monitoring problems, fault detection and diagnostics, and system design.