Research On Fault Diagnosis Of Beam Pumping Unit Based On Deep Learning

Oil is the lifeblood of a country’s economic development.In the process of oil extraction,the working environment of the pumping unit is complex.Timely diagnosis of pumping unit faults is of crucial importance for controlling oil well safety production and improving oil field production efficiency.Analyzing the dynamometer card to quickly determine the working conditions of the pumping unit has always been the main method of diagnosing pumping unit faults.Currently,deep learning convolutional neural networks are widely used in this field.This paper first introduces the basic principles,research status,and development trends of deep learning dynamometer card fault diagnosis.The main problem is the contradiction between computational complexity and accuracy.The computational complexity of the network increases exponentially with the growth of sample data and sample data features.Therefore,the further improvement of network accuracy will be limited by computer conditions.First,this paper takes the widely used rod pumping unit dynamometer card as the research object,preprocesses the actual dynamometer card to improve data recognition accuracy,and establishes a dataset.Then,the principles,structures,and training processes of convolutional neural networks are analyzed.After that,five different deep learning network models(Goog Le Net,Res Net18,Mobile Net,VGG-16,and Shuffle Net V2)are introduced.Through experimental comparison,the lightweight network model Shuffle Net V2 with the fastest operating speed is selected.Its unique network structure greatly simplifies the computational complexity,but its accuracy is lower.To solve the problem of low accuracy,a residual link mechanism and an increase in the number of network layers are used to significantly improve its accuracy.The effectiveness of the method is verified by comparing it with other networks.Finally,based on these networks,hyperparameter optimization methods such as grid search,random search,Bayesian optimization,and particle swarm optimization are analyzed.A simple experiment is designed to verify the optimization effects of Bayesian optimization and particle swarm optimization.Based on the experimental results,Bayesian optimization is used to optimize the hyperparameters of the residual Shuffle Net V2 network,and an improved residual Shuffle Net V2 network is obtained,which further improves the model’s accuracy.This paper designs and uses the improved residual Shuffle Net V2 network for dynamometer card fault diagnosis.The results show that this algorithm,while maintaining the advantages of speed and low computational complexity,has an accuracy 4.84 percentage points higher than the original model and is not inferior to deep networks such as Goog Le Net and VGG-16.It effectively solves the problems of existing deep learning dynamometer card fault diagnosis and has important application prospects and value.