Multiplex Visibility Graphs-Based Hybrid Deep Learning Method for Recognizing Pipeline Operation Conditions Using Operating Data

Abstract Accurate recognition of system operating conditions is the basis for ensuring the safe and stable operation of oil and gas pipeline network systems. However, the existing operating condition recognition mainly relies on manual experience, which cannot dynamically track the changes of pipeline operating status, and the operating data stored in SCADA system lacks operating condition labels, which makes it difficult to explore the potential value of the data. In this work, a hybrid neural network model based on multiplex visibility graphs (MVG) is proposed for operating conditions classification. Firstly, the generative adversarial network algorithms (GANs) are adopted to increase the sample size of abnormal conditions and unsteady conditions to deal with sample imbalance in oil and gas pipeline systems, and the reliability of the generated samples is verified by comparing the similarity of probability distributions of the generated data with the real data. Then, the operating data are mapped into the MVG. Next, an attention mechanism-based graph convolutional neural network (GCN) model and a long short-term memory network (LSTM) model are successively employed in the MVG for capturing spatial and temporal features in the operating data. By introducing the attention mechanism into the GCN model, the model can adaptively discover signal nodes with higher correlation to the target operating conditions, thus improving the classification accuracy. Finally, the proposed model is validated with historical operating data from a real multi-product pipeline system. The results show that the proposed hybrid method can be effectively applied to binary classification of scheduled operations/sling pump (accuracy = 0.95) and multiple classification of pipeline startup and shutdown/sling pump/switching pump/pigging (accuracy = 0.72). The proposed model has the highest recognition performance compared to both the single GCN model and the hybrid model without the attention mechanism. The proposed method is universal and can be promoted and applied to oil pipeline systems and natural gas pipeline systems, which is of great significance for realizing online monitoring of pipeline operating conditions and guaranteeing the safe transportation of oil and gas pipelines. It also provides a new idea for future research on the recognition of small sample operating conditions in oil and gas pipeline systems.