Research On Deep Learning Based Pipeline Leak Detection Methoed

In recent years,the acoustic method has been widely used in pipeline leak detection systems due to high sensitivity,high positioning accuracy,low cost and other advantages.However,due to its high sensitivity characteristic,the leak detection and location system based on the method is very easy to produce false alarms.In addition,the existing pipeline leak detection methods used a complete signal as detection object to classify whether the abnormal sub-signal exits,but the classification results cannot give the local information of the local abnormal sub-signal in a signal,that is,cannot give the number of abnormal sub-signal and its location in a complete signal.However,in practical applications,obtaining the number and location information of abnormal sub-signals plays an important role in accurate localization and elimination of false alarms and missed alarms.To address the above-mentioned problems,this topic treats both the interference signal and the leakage signal as local abnormal sub-signals,and researches the deep learning-based local abnormal sub-signal extraction method to realize the detection of local abnormal sub-signals in the pipeline transmission process from a new perspective,and combines the existing signal matching-based mutual correlation localization method to realize the accurate detection and localization of pipeline leak.Since it is difficult to obtain a large number of leak signal samples for actual pipelines,and the deep learning-based training of the abnormal signal detection model requires the support of a large number of samples,this paper presents the signal’s time-domain local joint features from the perspective of the signal’s time-domain local features,combining multiple time-domain local features such as the amplitude and width of the local sub-signal,and analyzes the feasibility of the local abnormal sub-signal extraction model to achieve small sample modeling by combining the characteristics of the network itself.To address the phenomenon of false alarms caused by non-homologous local abnormal sub-signals upstream and downstream in pipeline leak detection and localization,this paper proposes a signal homology detection method based on the combination of Siamese network and convolutional classification network to achieve upstream and downstream identification of non-homologous abnormal sub-signals,which solves the disadvantage that the Siamese network needs to set the threshold to judge whether the samples are similar or not after getting the similarity value and reduces the false alarms caused by non-homologous signals.The proposed method was tested using the actual acoustic signal history data collected in the field,and the test results show that:the deep learning-based local abnormal sub-signal extraction model has good experimental results and strong generalization ability on multiple pipelines;the non-homologous signal identification method based on Siamese network and convolutional classification network can effectively reduce false alarms;the pipeline leak detection method proposed in the topic provides an effective and reliable implementation method for industrial applications.