Study On Fault Diagnosis Method Of Indicator Diagram Based On Convolutional Neural Network

Petroleum is a vital strategic resource,which is related to national economy and people’s livelihood.Our country has to pay a huge bill for oil imports every year.Therefore,the realization of automatic,rapid and accurate fault diagnosis of the pumping system is very important to reduce the production cost and ensure the output of the oilfield.In this paper,the fault indicator diagram is taken as the research object.The deep learning theory is combined with the convolutional neural network theory method and the integrated learning thought as the theoretical basis,then,with mature design thinking of convolution model as the guiding ideology,to optimize the network.Then,combined with the Bagging algorithm with improved voting mechanism,the modeling of the fault dynamometer recognition model is completed.The specific research content is as follows:1st)A convolutional neural network based on classical network structure is proposed for image classification,and the effects of convolution kernel size,initial learning rate,network depth and weight initialization method on network model recognition rate and convergence performance are analyzed respectively.The adjustment process was demonstrated and a base model called Baseline with a test accuracy of 94.6%was obtained.2st)Complex adjustment process often requires a wealth of experience and business knowledge.That is one of the difficulties of deep learning.Aiming at this problem,this paper proposes a dynamometer card recognition method that integrates integrated learning and convolutional neural networks.The method uses the preliminary convolutional neural networks as weak classifiers,and then integrates tlhem with Bagging method.This method not only reduces the dependence on the parameter tuning technology,but also improves the accuracy of 4.1%based on Baseline.3st))The voting strategy adopted by the traditional Bagging method does not take into account the problem that the classification performance of the weak classifiers is different for each data category,and in view of the higher requirements on the recall rate in the actual.