Release of chemicals or biological agents in the subsurface often results in plumes migrating in the medium, posing risk to human and ecological environments. Temporal and spatial monitoring of the plume concentrations are needed to assess risk, make decisions and take remedial action. Current underground contaminant plume monitoring technologies are inefficient, expensive and ineffective. Wireless sensor technologies have the potential to dramatically improve this process. A closed-loop system integrating wireless sensor network based monitoring with numerical models for plume tracking is being developed, in which sensor data continuously calibrates and validates the system identification and prediction models, while the output from these models direct the sensor network operation to optimize constraints such as accuracy and power consumption. The system is based on a novel virtual sensor network architecture with broader applicability beyond plume tracking. Algorithms and protocols being developed support the formation, usage, adaptation and maintenance of dynamic subsets of collaborating sensors, named Virtual Sensor Networks (VSNs). VSN protocols for collaboration among groups of sensors will greatly ease the task of deploying sensor networks, especially in environments where multiple geographically overlapping applications are deployed. A proof-of-concept laboratory test bed that captures the complex subsurface processes is used for integration and evaluation of VSN protocols. This interdisciplinary project will significantly advance the state-of-the art in subsurface plume tracking and sensor networking technologies. It will stimulate a unique partnership of electrical engineers, computer scientists and environmental researchers, and demonstrate closed-loop operation of computer models and sensor networks to solve complex environmental problems.
The purpose of this project is to develop a closed-loop system integrating a wireless sensor network based monitoring with numerical subsurface contaminant transport models in order to dramatically improve the effectiveness, accuracy and utility of underground contaminant plume monitoring. Three major contributions have been made in this project: (1) sensor data quality issues; (2) data assimilation into subsurface contaminant transport models; and (3) experimental test bed construction. We have investigated sensor data quality issues. One of the most notable outcomes is that we designed and developed REDFLAG, a fault detection service that addresses the two most worrisome issues in data-driven sensor applications: abnormal data and missing data. REDFLAG exposes faults as they occur by using distributed algorithms in order to conserve energy. Simulation results show that REDFLAG is lightweight both in terms of footprint and required power resources while ensuring satisfactory detection and diagnosis accuracy. REDFLAG has been applied into a subsurface contaminant transport model to improve the model performance in the presence of erroneous sensor data. A model's predictive performance is dependent on being able to match the model's outputs to the actual observations and is quantified by the objective function which contains the residuals between the data ``received'' from the network and the data computed by the predictive computational model. REDFLAG has reduced the objective function by almost three orders of magnitude due to the removal of a large majority of faulty data. We have addressed the challenges rising from sensor data assimilation. We have focused on the periodic inclusion of concentration data into a computational advection-dispersion transport model. The data are synthetically generated and a WSN simulator is used to inject authentic anomalies into the data set. To provide practical insight into the directions of future modeling efforts in a new data context, we have considered the fate of accepted numerical subsurface contaminant models (i.e., MT3DMS, MODFLOW-2000, and PEST) in light of a new data context. A complex 3D synthetic model was developed using field site measurements and, from this, four datasets were constructed to mimic different scenarios: (1) infrequently collected data with minimal noise; (2) daily WSN data containing only theoretical noise levels according to sensor specifications; (3) data containing WSN faults (including noise); and (4) data with fewer WSN faults due to the use of a WSN fault detection application. A simple data assimilation methodology was employed to calibrate a 2D numerical transport model and subsequent model predictions were compared to the reference dataset. We have found that even small amounts of erroneous data may significantly affect the outcome of the calibrated model. Satisfactory parameter estimation results do not guarantee reasonable model predictions if faults persist. Transport model forecasts are highly sensitive to parameter estimation input variables, particularly regularization parameters. Model agreement generally improves with each iteration of data assimilation even when most previous data has been eliminated through data reduction. This iterative approach to parameter estimation may be appropriate whenever transient data is available. Predictions of transport fate only remain valid for short time periods after model calibration. In order to validate the closed-loop system, we have developed a large three dimensional, laboratory-scale synthetic aquifer, designed experiments which create complex plume configurations from an ionic tracer, employed a network of electrical conductivity sensors to track plume movement, and developed a methodology which uses the sensor data to calibrate numerical models designed to simulate plume transport through the synthetic aquifer. The key lessons learned during the course of constructing the tank and collecting data from sensors are as follows. (1) Sensor calibrations are critical to obtaining accurate data; however it is often difficult to obtain absolute value precision from sensors. Often they perform better at measuring changes and change rates, therefore applying relative measurement scales is frequently more advantageous than using absolute scales. For WSN applications an absolute sensor measurement scale is preferred because it provides a baseline for comparison. It is recommended that WSN employ an absolute value reference adjustment to all sensors. This could be accomplished electronically if the sensors prove to be sufficiently robust or it could be accomplished by placing each sensor in a control solution for a reference reading prior to site installation. (2) WSN data analysis procedures must account for drifting and faulty data. (3) Fluid density differences in open hole wells have a considerable influence on the depth-wise aqueous concentration distribution within the well. Long term emplacement of sensors will have to account for possible density driven distributions. Sensors located at multiple depths within a single well may be required to accurately assess the mass of a contaminant.
