ARX modeling approach to leak detection and diagnosis

摘要

This work presents a time series strategy for detection, location and quantification of leaks in large pipeline systems. The technology has two active components, which operate sequentially: the Detector and the Localizer. The Detector continuously screens real-time data, searching for any anomalies such as leaks, which are detected (or not) depending on their size and position. The Detector is based on autoregressive multi-input/multi-output (MIMO) ARX predictors with one input filter. Subsequent to successful leak detection, the Localizer is launched to diagnose the leak via estimation of its parameters - diameter and location - using recorded data on a Search Time Window that includes information in the neighborhood of the instant of detection. The Localizer is also an ARX predictor, but with two input processors, the first is a filter for dynamic plant inputs and the second filter processes "parameter signals" of active leaks. The Localizer is developed beforehand via model identification with plant data under the action of known, artificially simulated, leaks. It is, therefore, able to recognize an active pattern of leak parameters, by maximizing the adherence of its predictions to data in the Search Time Window. The proposed detection and location methods were successfully tested in simulated leak scenarios for an industrial naphtha pipeline.