Intelligent Time Series Prediction and Anomaly Detection Method and Engineering Validation for Industrial Scenarios of Mobile Oil Pipelines
Year: 2026
Type: conference-paper
Abstract: Mobile oil pipelines are the core equipment of the emergency oil support and field oil transportation system, and their transportation safety and stability determine support mission execution efficiency. Aiming at problems in existing anomaly detection methods under unsteady conditions, strong interference, and small abnormal sample scenarios, this paper proposes an AI time series prediction and anomaly detection method integrating hydraulic mechanism prior knowledge. The method constructs a time - frequency hybrid attention time series prediction model with dual constraints to achieve high - precision short - term prediction of pipeline key parameters under unsteady conditions. It designs a prediction - reconstruction dual - path collaborative anomaly detection framework driven by a task - aware gating network to accurately identify both sudden and gradual anomalies. A mechanism - constrained time series generative adversarial network is proposed to augment abnormal data in small sample scenarios and improve model generalization. Ablation experiments verify the necessity of the dual - path framework and the effectiveness of the gating network, and supplementary sensor fault simulation experiments prove the method's engineering robustness. Engineering validation on a 5 km pipeline shows that the proposed method has a pressure prediction MAPE of 2.15% under sudden condition changes, an overall anomaly detection accuracy of 98.2 %, and a false alarm rate of 0.32 %, significantly outperforming existing mainstream methods. The research results can support the intelligent operation and maintenance of mobile oil pipelines and provide a reference for intelligent anomaly detection in similar unsteady industrial scenarios.
Authors: Yang Zhang, Rulan Zhao, Qi Yang
Institution: PLA Army Service Academy
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7165754263