The paper demonstrates the construction, training, and uncertainty quantification analysis of an artificial neural network model for single phase liquid pipeline leak detection through pressure drop and flow rate monitoring. The demonstrated methods reached acceptable error levels efficiently using theoretical data produced by a theoretical physics model. The results demonstrate that the randomly simulated leakage can be efficiently detected using the trained ANN (Artificial Neural Network) model based on theoretical data derived from physical equations. However, complexity appears when simulated leakage with modeled uncertainty are used in the training of the AI models. The propagation and influence of the uncertainty in the input data on the ANN method are discussed. Randomness following certain probability distributions is introduced into the data to measure the influence on the efficiency and reliability of the training and results of the ANN model. The paper also discusses the influence of the range of input data on the predictability of the ANN model in leak detection.
Data driven Machine Learning technologies such as ANN (Artificial Neural Networks) provide great innovation opportunities towards the design and operation of oil and gas pipeline systems. ANN based models have proved to be efficient in predicting the pressure drop within a pipeline (Brkic, D., and Cojbasic, Z., 2016; Shayya, W.H., and Sablani, S.S, 1998; Salmasi, F., etc. 2012; Fadare, D.A., and Ofidhe, U.I., 2009) as well as solving the inverse problems such as detecting pipeline roughness progression, inner diameter change, and leakage (Cheng, D., Zeosky, D., 2019).
The increasing interest towards machine learning models causes the need to highlight the concerns of uncertainty propagation into the engineering applications of the methods. Machine learning models are developed through analysis of data derived either from physics models or measurements.