Data-Driven Methods in Pipeline Leakage Detection

Abstract

Nowadays, leakage detection is of great importance as pipelines are the major means of transporting hydrocarbon fluids and gases. In this thesis, we propose two methods based on supervised learning and filtering to deal with the pipeline leakage detection problem. First, a novel two-stage detection method is introduced to differentiate normal, leakage and transient conditions of pipelines. In this method, feature vectors are constructed from the flow rate and pressure using leakage characteristics. An artificial neural network (ANN) is used in the first stage of the detection to differentiate normal and abnormal conditions with the feature vectors as the inputs. In the second detection stage, simple logic is used to distinguish leakage and transient for data under abnormal condition. The method has been shown to have higher detection performance and fewer false alarms in comparison with the line balance and Kantorovich distance methods. As the pipeline leak data is not always abundant to train supervised learning models, a filter-based method is proposed to detect pipeline leakage that does not require prior leak data for training. Based on studies in field data, we model pipeline leakage as an increase in the mean value of the flow rate difference between the inlet and the outlet sensors, where the increased value is unknown and subject to change. Then, an adaptive filter is proposed based on the estimated cumulative distribution function (CDF) of the data in steady-state condition using kernel density estimation. The proposed filter has better performance in small leaks in comparison with different benchmarks.