SBIR Phase I: Early Detection of Anomalies in Large-Scale Gas Networks

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) project is to dramatically reduce the incidence of natural gas pipeline failures across the country, within the next 3 to 5 years. Every year there are a few hundred "significant" pipeline accidents (fatalities or significant property damage) causing massive damage to life and property, dispersing hazardous materials and disrupting gas distribution services. These events result in many hundreds of millions of dollars in repair and recovery, and large fines that can be up to a billion dollars or more. This scalable and economical capability will significantly reduce the likelihood of such failures, without requiring additional infrastructure. Performance has been validated at a large utility company, and the prototype has demonstrated the ability to capture a substantial fraction of previously undetected events with significant advance warning (90 minutes or more). This outcome represents a clear performance improvement over existing systems and is enabled by advanced models customized for the gas-utility domain. The methods developed in this project can be directly applied to improve detection accuracy in other contexts such as power-grid networks, computer cluster management and financial fraud detection. This SBIR Phase I project proposes to detect anomalies in large-scale gas-utility networks through statistical inference from continuously observed time-series data on pressure, prevailing temperature, and other characteristics of the network. Anomalies within gas-utility networks occur due to a variety of reasons, e.g., sulphur or ice buildup in the pipelines, and corrosion/aging of hardware, and are often preceded by detectable signatures in the time-series of gas-pressure data. A premise of the project is that the early detection of such signatures, leading to advance warning of 90 minutes or more, allows corrective action within the utility network to avoid significant property damage, loss of life, and service disruption. The project proposes new methods for the rapid estimation of short and medium timescale models of gas pressure behavior from voluminous streaming data, along with methods for constructing prediction bands through Monte Carlo and stochastic optimization techniques. Such methods are non-generic and their success relies crucially on exploiting specific structural properties that are unique to network-level gas-pressure time series, along with modern trends in statistical machine learning. The proposed stochastic optimization techniques will probabilistically classify identified anomalies into ``failure type," allowing the prioritizing of network level emergency operations. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.