Pipeline safety is of paramount importance for socio-economic development, necessitating regular inspection and maintenance. The complexity of the internal pipeline environment and the presence of irregular noise in detection data, however, pose significant challenges to pipeline inspection technologies. Additionally, the reliance on manual expertise for inspection and the absence of standardized assessment criteria further complicate the development of automated inspection methods in this context. To address these challenges, this paper proposes a pipeline anomaly detection framework based on reinforcement learning with hierarchical reward exploration mechanism (RHiREM). It includes two main aspects: Firstly, the hierarchical reward mechanism. By deeply simulating the process of defect recognition based on expert personal experience, the original pipeline data is first adaptively divided into sets of windows with different sizes, and then attribute and type judgments are performed on them. In this way, the approach achieves accurate identification of defects and pipeline structures in scenarios with minimal noise interference. Secondly, the hierarchical exploration mechanism. By leveraging the temporal exploration and spatial exploration, the mechanism enables further deep search and feature learning on complex pipeline signals, and facilitates comprehensive assessment of the relationships between global features and local features across different signals, effectively resolving the difficulties associated with identifying defect signals in the presence of high noise interference. the proposed framework has been demonstrated to automate the detection of complex on-site pipeline internal signals and successfully detected the common anomalies with high F1-score over conventional techniques.
Read full abstract