The AquaSCALE Project, a Global Cities Challenge Project,  is a community government/academic/industry partnership effort that aims to design robust water systems by capturing the structure, components and operation of community water systems. AquaSCALE is a cyber-physical-human middleware for gathering, analyzing and adapting operations of increasingly failure-prone community water distribution services. Today, detection of anomalous events in civil infrastructures (e.g. pipe breaks, contamination) is time consuming and often takes hours or days. AquaSCALE leverages dynamic data from multiple information sources including IoT sensing data, geophysical data, human input, and simulation and modeling engines to accurately and quickly identify vulnerable spots in water networks. Such sensor-simulation-data integration platforms can assist in design time tasks (e.g. optimize IoT device placement), enable fault detection (e.g isolation of leaky pipes), trigger runtime adaptations (e.g. control of valves) and predict/reduce cascading impacts (e.g. flood). AquaSCALE bridges the infrastructure/application gap by transforming input sensor data streams collected at lower infrastructure layers to semantic streams that capture application-level entities at a higher service layer. 

During this project, we developed a range of design-time and operational methodologies to create smart water infrastructures. At design-time, we developed techniques for sensor/IoT placement leveraging the intuition that failure events can have varied impacts on the surrounding community depending on size and demographics of affected populations, economic impact and intensity of events.  We developed impact-driven sensor deployment strategies that ensure minimum impact to the community under different failure events of varying intensities.  Since instrumenting underground water networks with in-situ sensors will incur significant costs for trenching and maintenance, we designed methods to augment in-situ deployments with mobile water sensing units that can be inserted into and extracted from pipes through existing infrastructure. E.g.manholes, fire hydrants. We proposed a hybrid deployment that combined advantages of mobile and in-situ instrumentation to provide a cost-efficient, low impact approach that was tested on several real-world water networks.

Operational techniques we developed include new methods for fault detection in water infrastructure for pipe failures and contamination events, leveraging simulators and machine learning algorithms. Our methodology uses a two phase approach - in Phase1, an extremely large number of offline fault profiles are generated using commercial grade simulators (EPANET, WNTR) resulting in a rich  dataset to train ML classifiers.  To address pipeline leaks, we designed an ensemble-based approach to accurately classify the likelihood of multiple failure events. In Phase2, external data sources like weather information and human input, were incorporated to improve the accuracy and speed of event detection in real-time settings.  To address water contamination events, we explored human-in-the-loop techniques that can guide human participants (e.g. utility providers, field staff) in progressively locating contamination sources. Here, apriori profiles capturing the physical nature of contaminant transport are used in an iterative event processing strategy with two steps: location inference and location refinement. The inference step uses HMMs to generate an approximate set of contaminant sources that are then refined using human-driven grab samples.. The refinement step implements reinforcement learning to determine optimal sampling locations to localize the source quickly. 

We also addressed the problem of  enabling resilience in large catastrophic events such as earthquakes by estimating the current operating states of the network quickly and accurately, often with very limited information. Our state estimation methods utilize AI-based methods from graphical models, a structured probabilistic framework, along with inference methods based on belief propagation in order to derive optimal estimates of the infrastructure state in near real-time.  Working with ImageCat Inc, we modeled failures using fragility models under large earthquake scenarios and conducted a series of validation studies to compare pipe break scenarios to actual patterns documented during several California earthquakes, including the 1994 Northridge earthquake. Finally, we proposed a framework to incorporate edge computing to collect and analyze data from water infrastructure to provide lower latency of analysis, reduce network bandwidth consumed and leverage the heterogeneity of available sensors to ensure a more resource-efficient monitoring approach towards resilience of water infrastructures. 

The project has trained several graduate students (Ph.D.,M.S) on techniques to improve  resilience of community scale infrastructures; it has established the viability and importance of data science techniques in creating resilient cyberphysical infrastructures. K-12 outreach was realized through programs such as the  IoTSITY Site REU and via undergraduate student projects. The AquaSCALE design and datasets are available to researchers; research results have been disseminated via several publications and talks in top tier venues and won best paper awards. Community outreach efforts include partnerships with local governments at the city/county level with whom we have established a good working relationship. The work on AquaSCALE has also instantiated new efforts that address  challenges across the broader set of integrated water systems (potable water, wastewater, stormwater) in communities today, 

Last Modified: 12/18/2020
Modified by: Nalini Venkatasubramanian