Use of Artificial Neural Networks and fluid transient waves for active and passive inspection of water pipelines

Abstract

Water is a vital resource to society and complex interactions between nature and human infrastructure are constantly required. Water transmission and distribution pipelines are critical for modern cities; however, their sheer size and the fact that most of them are buried underground, makes the health monitoring of pipelines challenging. In addition, some water transmission pipelines cover long distances through remote areas that are not easily inspected regularly. Fluid transients have been used over the last 25 years as part of techniques to assess and monitor the condition of pipelines by interpreting the pressure response to the presence of anomalies (e.g. leaks, blockages) and the occurrence of abnormal events (e.g. bursts). Nonetheless, existing techniques require detailed information with regard to the properties of the pipe system (model-based techniques), and require manual interpretation of the measured pressure signal or involve optimization methods that result in large amount of computer processing time to obtain results. Consequently, the research in this thesis presents new noninvasive techniques for the active and passive inspection of the condition of a pipeline combining custom-designed machine learning algorithms based on Artificial Neural Networks (ANNs) and fluid transients. An active inspection technique has been developed that is based on the interpretation of high-frequency pressure measurements using ANNs after the generation of a small and controlled transient event in a pipeline. The initial transient pressure wave produced by the rapid closure of a side discharge valve propagates through the pipeline and interacts with any anomalies that are present in the pipeline. The measured transient pressure trace is then processed using an ANN trained with transient pressure traces obtained from a physical model of the pipeline to locate and characterize the anomalies. The other method that has been developed is a passive inspection technique that is also based on high-frequency pressure measurements; however, in this case, there is no artificial generation of a transient event. This methodology provides for the continuous transient pressure monitoring of the pipeline by analyzing changes in pressure due to the occurrence of abnormal events such as bursts. This method uses ANNs at different stages of the process to determine if the pressure condition of the pipeline is normal or whether a potential abnormal event, such as a pipe burst, might have occurred. The major research contributions of this thesis are presented in three journal publications. These publications describe i) a novel framework for the use of ANNs for the active inspection of pipelines based on measured transient pressure traces applied to the detection of junctions and leaks in a numerically modeled pipeline, ii) a complete methodology for the active inspection of pipelines for the detection of leaks in a laboratory pipeline using an array of ANNs trained with datasets using different noise intensities to obtain robust, accurate and fast predictions when background pressure fluctuations are present in the pipeline, and iii) a complete methodology passive inspection of pipelines for the detection, location and characterization of bursts in numerical and laboratory pipelines. The overall contribution of this research is the development of new non-invasive techniques for the active and passive condition assessment of pressurized pipelines using machine learning algorithms. These techniques have the advantage of being data-driven, meaning once the ANNs have been trained using a physical model of the pipeline, no model of the analyzed pipeline is required when new measured pressure traces are interpreted by the ANNs. In addition, results can be obtained fast (near real time) and are accurate in locating and characterizing leaks and bursts in pipelines.