Simulated Data-based Adversarial Learning for Zero-shot Anomaly Monitoring of Pipeline Leakages

Fluid pipeline systems are widely used in key areas such as industrial water, petrochemical, and ship pipelines. However, pipeline systems are prone to structural failures such as leaks due to various complex factors during long-term operation. Traditional detection methods suffer from issues such as low efficiency, while existing data-driven intelligent diagnosis methods are limited by the scarcity of fault data, making it difficult to deal with unknown leaks. To address this, this paper proposes a zero-shot pipeline anomaly monitoring method combining simulation and a dual adversarial decoder. Firstly, a simulation model is built using Flowmaster software to obtain abnormal data of pipeline leaks through simulation. Subsequently, real normal samples and simulated abnormal samples are jointly used for pre-training to construct a model framework comprising a masked patch autoencoder and a dual adversarial decoder. In the inference stage, zero-shot anomaly monitoring of unknown faults is achieved by calculating reconstruction errors and dynamic thresholds. Experimental results on public datasets show that this method outperforms traditional methods in terms of accuracy, recall, and F1 score, verifying its effectiveness.