Deeppipe: A deep-learning method for anomaly detection of multi-product pipelines

Abstract

The multi-product pipeline is the main way of refined oil transportation to ensure the safety of the energy supply. Considering that abnormal conditions of the pipeline will cause huge economic losses, personal injury, and environmental pollution, it is important to conduct abnormal detection of multi-product pipelines in time. This work proposes a deep-learning method for multi-product pipeline anomaly detection from the perspective of the temporal and spatial characteristics of the pipeline system. First, analyzing the topological structure of the real pipeline and collecting relevant equipment parameters to establish an SPS simulation pipeline system. By simulating common normal conditions and abnormal conditions in the pipeline, the pressure matrixes of each station are constructed. After being processed by the adaptive padding network and the pairing network, it is input into a customized ranking model to realize the anomaly detection of the multi-product pipeline. Finally, a simulated pipeline and a real pipeline are taken as examples to verify the effectiveness of the proposed method. The results show that the proposed hybrid model has high accuracy, precision, recall, and F1 score of 99.5%, 99.5%, 99.4%, and 99.4%, respectively, which are better than the traditional ranking model. This method can be applied to guide the safe operation and management of on-site multi-product pipelines.