Chapter 2 is a background chapter that discusses the application and methodology. Chapter 3 describes the experimental data that comprise the case study. The dissertation “core” consists of three distinct investigations (Chapters 4-6), followed by a concluding
chapter (Chapter 7). In the following paragraphs, I broadly outline Chapters 4-7, summarizing the objective, methods, and achievements of each.
The investigations in Chapters 4-6 are executed using experimental data obtained from one building. These data are used to generate several sets of sensor signals from a hypothetical release event. The Bayes Monte Carlo algorithm is the main analytical tool. Details of the algorithm and the case study are presented in Chapters 2 and 3.
Chapter 4 investigates how the individual sensor characteristics can be selected in such a way to improve the overall system performance. A class of chemical sensors that produce binary or threshold-type signals is investigated. The performance trade-offs among different sensor characteristics — including threshold levels, response time, and sensor error — are evaluated.
Several sets of threshold sensor signals were generated by exercising the tracer data from one release over different values of threshold levels, response time and error. The success of each set of sensor characteristics is evaluated by the ability of the system to characterize the release through reducing the uncertainty of each model parameter. The effects of each sensor characteristic are explained through physical insight.
This chapter demonstrates that a system-level approach may be beneficial in selecting an effective mix of sensor characteristics for sensor system application. The Bayes Monte Carlo algorithm is evaluated using signals that are generated from real experimental data, rather than from simulated data.
While Chapter 4 considers the limitations imposed by the sensors, Chapter 5 considers the limitations imposed by transport physics and the associated model. Chapter 5 explores the concept that the information a sensor system may learn about a
release is ultimately constrained by the physics that govern contaminant transport and therefore also by the physical model used by the algorithm. Understanding these relationships may lead to better design decisions.
The connections among the governing physics, the transport model, and system performance are investigated from the perspective of the contaminant transport and mixing time scales. A discussion describes how the time scales may influence the ability of the sensor system to characterize the release location and strength. The time scales are then investigated in two ways: first, by blinding portions of the data to the algorithm in order to assess the value of the blinded information; and second, by evaluating the performance of several sensor networks.
The relationship between the time scales and the transport model suggests that the statistical algorthms must carefully treat model uncertainty. The resulting likelihood function, though developed for the particular case study and BMC algorithm, may be useful for representing model uncertainty for other algorithms using a similar modeling approach.
Chapter 5 provides further insight into how the transport physics may help explain the relationship between sensor characteristics and sensor-system performance. While the investigation is necessarily limited to the case study, the time-scale concepts and illustrative likelihood function may help guide design of sensor systems for more complicated systems, and may also help identify and frame the relevant questions.
The concepts introduced in Chapter 5 emphasize the importance of the transport physics and model in selecting sensors. While Chapters 4 and 5 considered contaminant sensors, Chapter 6 considers the fusion of information from other kinds of sensors.
Networks employing non-chemical sensors can reduce the overall information costs, or may improve performance relative to a network that exclusively uses chemical sensors.
Chapter 6 describes a Bayesian framework for fusing information from heterogeneous types of sensors. While the framework is general, it is evaluated using the case study to investigate the benefit of door position switches.
The results from using these sensors are interpreted physically. I discuss how fusing information from heterogeneous sensors may help address the challenge of system complexity. I identify where the BMC and modeling framework may impose limitations to extracting otherwise potentially useful information from other sensor types, and offer suggestions for future research directions.
Chapter 7 concludes the dissertation with a summary and a discussion of ideas for future research on the themes explored here.
1.7 Contributions
The achievements in this research are empirical, conceptual, and methodological. On an empirical level, this research uses experimental data rather than synthetic data. It also explores the use of binary sensors, concentration sensors, and non-chemical sensors. In terms of conceptual development, a framework was developed for understanding how sensor system performance may be related to the underlying transport physics and its model representation. I discuss concepts for fusing heterogeneous sensor types. While both of these frameworks are explored within the context of the BMC algorithm, they are likely to have relevance to other sensor interpretation algorithms, as they are connected to the underlying transport physics.
Methodologically, a likelihood function was developed to characterize model uncertainty for the illustrative example. The functional form may be applicable for other buildings. While the likelihood function is not optimal, opportunities are identified to further develop a optimal likelihood function, which is more robust to model and sensor uncertainties.
Many of the investigations reported here were conducted using data from a case study. A benefit from using a case study approach is that the research is grounded in real truth, rather than highly idealized conditions. However, a drawback from using a case study approach is that the generalizability from the case study approach is not always assured. Generalizability exists in degrees, and the specific extent of generalizability of truths revealed from a case-study approach must be assessed carefully.
Thus, the generalizability of this dissertation’s findings to different release conditions and different building types is not certain. It is likely the findings will have high relevance to single, instantaneous releases for buildings that are served by central, overhead air distribution. The extent to which the findings may be applicable to other releases and ventilation systems requires further investigation. Because the insight and frameworks presented in each chapter are supported by physical explanations, it is possible to hypothesize how the findings may apply to different systems. At a minimum, this dissertation is likely to provide insight into the questions that may be relevant when considering more complex buildings and airflow conditions. Even in the event that the direct methodologies fail when applied to different systems, such failure would provide valuable insight into improvement opportunities, and is therefore relevant.
This dissertation considers the development of sensor systems for high-risk contaminants that are released in or near a building. The methods explored also may be applicable to other types of environmental monitoring, such as energy monitoring, or fugitive greenhouse gas emission characterization. Additional design problems may also benefit from a systems-centered, Bayesian approach. While the specific findings may have limited direct relevance, the concepts, challenges, and questions that inspired the investigations are likely to be highly pertinent in many and diverse applications.
Chapter 2
2 Background
This dissertation develops real-time sensor systems for characterizing indoor high-risk contaminant releases. The task of the sensor system is to transform concentration measurements into information that describes an unknown release conditions. A Bayesian statistical approach is used to achieve this task. Several subjects are relevant to this topic that include the physics of airborne environmental contaminant fate and transport, building systems and operation, contaminant transport modeling, and statistical methods. It is important, also, to understand the research accomplishments in indoor sensing systems and airborne contaminant sensor systems as the application considered in this dissertation sits within a broader domain. This chapter provides background material for these relevant topics. Sections 2.3.1 and 2.4.4 describe the specific modeling and statistical tools used in the dissertation (i.e., multizone model and two-stage Bayes Monte Carlo algorithm).