Abnormal events recognition and classification for pipeline monitoring systems based on vibration analysis and artificial neural networks

Abstract

Pipelines become the principal means of oil and gas transportation. The leakage usually takes place due to some natural or artificial damages and causes loss of life and properties. Now a pre-warning system based on distributed optical fiber sensor has been proposed and deployed in China. Now, its following key problem is how to recognize and classify damage activities along with pipeline, such as ramming, rotor working, manual digging, well knocking, and mechanical execution. This paper involves in-depth study on recognition method for this system. Firstly, original vibration signal is pre-processed and segmented according to energy threshold and sliding window. Through statistical and short-time Fourier transform (STFT) analysis in time and frequency domain, energy ratios and frequency centroid are extracted as feature vectors, which can describe and distinguish distribution characteristics of each vibration event effectively. At classification, event set is divided firstly into discrete and continuous events with kurtosis, which can decrease classified event dimension and improve recognition accuracy. Then BP artificial neural network is applied to identify damage and non-threatening events. Experiment results show that proposed algorithm can differentiate discrete events with accuracy rate of 99%, while continuous events with 97.5%.