Predicting pipeline leakage in petrochemical system through GAN and LSTM

In petrochemical system, how to predict the leakage of pipeline is one of the most critical problems due to its huge side effects. Prior studies are mostly based on statistical analysis and they focus on the healthiest state of the system. However, these approaches ignore the trend of changes among data and lead to low accuracies. In this paper, we conduct research on characteristic of petrochemical data and get four core findings. We find that the data are (1) noisy, (2) high dimension, (3) imbalanced, (4) temporal correlated. According to these findings, we propose a novel neural network based classification model using GAN and LSTM to predict the leakage of pipeline. Specifically, we firstly apply a sliding window and a key feature selecting method in data preparation stage to address the noise and high dimension problems. We then propose a GAN based data enhancement approach to synthesize fault data to address the imbalance problem. After that, we propose an LSTM based classifier approach to learn the temporal correlation of data and classify the state of pipeline to predict leakage. We conduct extensive experiments and they show that our approach achieves 2x F1 score (0.8166) and 3x AUC (0.8940) compared to the existing methods. Our proposed neural network based approach is more suitable for fault prediction of pipeline leakage in petrochemical system. (C) 2019 Elsevier B.V. All rights reserved.