MetaSense: Calibration of Personal Air Quality Sensors in the Field

Abstract:

Cyber-Physical Systems (CPS) promise to change how we study and interact with physical world. CPS use sensors and actuators connected to an autonomic cyberinfrastructure. A key requirement for having dependable CPS is the correct calibration of their sensors. Unfortunately, the current state of the art is to calibrate sensors in laboratories, often manually. Moreover, commodity sensor precision degrades over time.  Consequently, sensors must be periodically recalibrated. Sensors are often sent back for calibration, requiring the availability of additional spare sensors, further driving up costs. The fact that sensors can lose calibration without users noticing also impacts the dependability of the data collected.  For example, small commodity personal sensors hold the promise to detect health risks as they arise, giving patients and doctors an upper hand in preventing life-threatening medical events as well as the onset of serious. Our CitiSense personal air quality monitoring system could help prevent asthma attacks in asthmatics, as well as help prevent the development of asthma.  However, a showstopper for ubiquitous health monitoring is the high cost of sensor calibration and the limited confidence that physicians can have in the data collected by inexpensive sensors. The MetaSense project will remove these roadblocks to affordable person-centered healthcare.  We propose leveraging large networks of mobile sensors connected to the “cloud” to support self-calibration. The cloud enables using large data repositories and intelligent computational power to cross-reference data from different sensors to detect loss of calibration.  As a simple example, if two sensors are at the same location at the same time, they should report the same values, if properly calibrated.  If not, the least-recently calibrated sensor’s values are suspect.  Using machine learning, this idea can be broadly generalized to using the full population of sensors for self-calibration. To make our vision a reality, we must solve numerous problems across several disciplines.  First, to make gradual loss of calibration detectable, MetaSense must be able to discern signal from noise.  Humidity, temperature, barometric pressure, and cross-contaminants can alter sensor readings irrespective of the precision of calibration.  In addition, to make the readings of two sensors comparable, as proposed above, it is necessary to know that they are truly in the same context.  Both of these require a significant application of context-aware computing, even detecting whether a sensor is indoors or outdoors, since pollution levels vary significantly with such changes.  With this contextual data in hand, we posit that advanced machine learning techniques will be able to separate signal from noise, permitting large-scale cross-device calibration.  We propose that these techniques, using advanced context-aware techniques, can be applied to detecting cross-sensitivity as well.