Research On In-pipeline Anomaly Detection Method Based On Reinforcement Learning And Hierarchical Reward Exploration Mechanism

As an important part of China’s infrastructure,pipelines play a crucial role in the country’s economic and social development.Therefore,ensuring their safety and reliability is of paramount importance,and regular detection and maintenance of pipelines is necessary.However,the development of pipeline detection technology in China is relatively lagging behind,and the current difficulties mainly include:(a)the complex internal environment of pipelines in the field,and the related detection data containing a large amount of irregular noise,which can completely cover the effective components of the original data when various noises are superimposed,resulting in difficulties in designing anomaly detection methods;(b)the complex state of detection data inside pipelines makes the anomaly detection process highly dependent on manual experience and there is no fixed reference standard for evaluation,which poses challenges to the design of automated detection methods.To address the above difficulties,our thesis proposes a pipeline anomaly detection framework based on a hierarchical reward exploration mechanism,referring to the basic ideas of reinforcement learning,and achieves accurate automation of complex on-site pipeline internal detection signals.The main research work is as follows:(1)Hierarchical reward mechanism: Constructed the entire reinforcement learningbased pipeline anomaly detection intelligent diagnostic framework based on the hierarchical reward mechanism.By adaptively dividing the original pipeline data into different sized window sets and performing attribute and type judgments,the process of expert defect recognition based on personal experience is deeply simulated,achieving accurate identification of defects and pipeline components under conditions of low noise interference.Ultimately,accurate automated detection of internal defects in tensioned pipelines is achieved,with specific performance as follows:(a)The actual detection rate for weld types and defect types in tensioned pipelines is around 95%;(b)Compared with traditional multi-classification algorithms,the overall detection performance of unused pipelines is improved by more than 15.5%.(2)Hierarchical exploration mechanism: Based on a hierarchical exploration mechanism,the entire reinforcement learning intelligent diagnostic framework is structurally optimized.Through deep search and feature learning of the complex state signals of on-site pipelines in the temporal and spatial dimensions,a comprehensive measurement of the global and local feature relationships between different signals is achieved,solving the problem of difficult defect signal recognition under high noise interference and improving the identification efficiency of defects and pipeline components in complex environments.Ultimately,relatively accurate automated detection of internal defects in on-site pipelines is achieved,with specific performance as follows:(a)The recall rate of defect types in on-site pipelines under experimental conditions increased from 51.9% to 97.4%,and the precision rate increased from 69% to94.9%;the recall rate of weld types remained unchanged at 100%,but the precision rate increased from 56% to 100%.(b)Compared with traditional multi-classification algorithms,the overall detection performance of in-use pipelines under experimental conditions is improved by more than 15%.