Study On The Analytical Method Of Oil Pipeline Signal Detection Based On Deep Learning

As an important part of oil and gas transportation,Its safety is of great significance to the entire oil and gas process and the task of loading and unloading.Aiming at the problems of complex noise,insufficient defect signal data set,and unclear defect types encountered in the process of processing oil pipeline inspection signals,this paper is based on the magnetic stress signal data collected on site,the pulsed magnetic eddy current signal data and five types of laboratory detection defect signals.Data,analyze and compare the characteristics of different pipeline defect signals and signal abnormal points in detail,and realize the identification of oil pipeline defect types based on continuous wavelet transform,LSTM,CNN and other methods.The main research contents are as follows:(1)The wavelet threshold noise reduction method is optimized based on particle swarm optimization algorithm.This method has the advantages of adaptively selecting the threshold and avoiding the result falling into the local optimum.For the characteristics of multi-channel detection equipment,the correlation analysis of the signal data was carried out,and the results showed that there was a strong correlation between the detection signals of different channels.(2)Four detection models for abnormal signals are established based on the change-point detection algorithm,which realizes the automatic capture of abnormal signals in oil pipelines.The experimental results reveal that the sampling points of most defect signal segments are concentrated between 500 and 800,and the voltage at the abnormal point is proportional to the depth of the defect and inversely proportional to the size of the defect.(3)In order to expand the oil pipeline defect signal data set,a signal dynamic prediction model based on LSTM neural network is established.The performance of the model is tested with the field pipeline signal dataset,and the average prediction accuracy of qualified samples in two signal datasets is 93.8 % and 94.6 %,respectively.The results show that the model can capture the variation trend of oil pipeline signals and the correlation of data,which provides theoretical support for subsequent data augmentation work.(4)The oil pipeline signal is converted into a time-frequency map based on the continuous wavelet transform method,which enhances the patterns that are not visible in the original signal and strengthens the defect signal features.For the labeled oil pipeline defect signal data,a defect signal classification model based on convolutional neural network is established.At the same time,the LSTM prediction model is used to expand the oil pipeline defect signal sample set,which improves the defect signal classification accuracy from 74.8% to 81.0%.