Bayesian Data Fusion for Pipeline Leak Detection

In this paper we introduce a probabilistic model for data fusion for leak detection in oil and gas pipelines. We propose a fusion algorithm for both detecting and localizing leaks. Our algorithm optimally combines two heterogeneous systems, fiber optic Distributed Acoustic Sensing (DAS) and Internal Leak Detection (ILD) technology. The output of these two systems, which throughout the paper will be denoted interchangeable as measurements or data points or more frequently as test statistics are not necessarily related to physical quantities directly measured by physical sensors. For instance, ILD virtual sensing systems, based on computer simulation of pipeline conditions using advanced fluid mechanics and hydraulic modeling, can detect leaks by comparing the measured sensor data (for example flow, pressure or fluid temperature sensors) for a segment of pipeline with the predicted modeled conditions. On the contrary, DAS systems map physical fields acting on the fiber by exploiting coherent optical time domain reflectometry and probing the fiber with proper interrogation systems. With ILD we typically have low sensitivity, and poor localization. With DAS we have relatively high sensitivity, but also high false/nuisance alarm rates. By fusing the two we are able to exploit the entirely different physical principle they are based upon, achieving high sensitivity with low false alarms. The fusion process is based on building a Dynamic Bayesian Network (DBN) using the test statistics (which are indicative of a leak) provided by the DAS and the ILD systems. The hidden nodes in the network may indicate the leak/no leak hypothesis the system wants to test and the leak location, respectively. The observable nodes may denote the test statistics from both systems in each bin or zone the pipe is partitioned into. The probabilistic and causal relationships among the nodes are represented and executed as graphs and can thus be easily visualized and extended. The Bayesian perspective of the new analytic allows to easily and naturally incorporate a priori information (e.g., wall thickness) on the zone or bin where the leak is most likely going to occur into the leak location node and propagating this new data point through the inference network. Finally we validate our method using computer simulations and real world experimentation conducted by the GE Global Research Center located in Niskayuna, NY. The results demonstrate the benefit of fusing two heterogeneous orthogonal technologies resulting in reduced false alarms, increased response time and improved sensitivity.