Research On Fault Diagnosis Of Pumping Unit Based On Auxiliary Classification Generative Adversarial Network

Pumping unit plays a very important role in oil production system and is widely used in oil field industry.Generally,the working environment of the pumping unit is very bad and the operation cycle is long,so it is easy to cause the failure.Therefore,the real-time monitoring of the operation status of the pumping unit and timely repair and maintenance when the failure occurs are important means to ensure the safe and reliable operation of the pumping unit.Deep learning can directly learn the salient features of samples from the data set and automatically identify them.It is a more effective method in pumping unit fault diagnosis.However,due to the lack of pumping unit fault data set and high fault similarity,it will lead to the problem of low diagnosis accuracy.How to solve these problems is the key to realize the accurate fault diagnosis of pumping unit.Therefore,this paper takes the pumping unit as the research object,uses the auxiliary classification to generate the expand sample data of the countermeasure network,and uses the classification network to distinguish the running state of the pumping unit to diagnose the fault of the pumping unit.The main research contents of this paper are as follows:(1)To solve the problems of low sample quality and lack of diversity in the traditional generative adversarial model,an approach is proposed for the Squeeze-and-Excitation Networks(SEnet)module and the focus loss function(Floss)to improve the indicator diagram data.Firstly,SEnet module is added into generator and discriminator network to extract more effective features and reduce invalid or ineffective features,so as to achieve better generation effect;Secondly,the focus loss function is added to the discriminator,so that the network can be trained on samples with high similarity,and the training degree of samples is balanced to prevent over-fitting.Finally,experiments show that the quality and diversity of fault indicator images generated by this method are better than those generated by traditional networks.(2)In order to improve the deep learning method’s ability to extract fault features of pumping units,based on convolution neural network,this paper introduces self-attention network module into the network and improves the learning rate,activation function,loss function and optimizer.Firstly,the learning rate is optimized by cosine annealing method,and the learning rate is adjusted adaptive in the training process,so that the model can be optimized more easily;Secondly,the Ranger optimizer is used instead of the traditional optimizer to stabilize the initial training,improve the convergence of the network and improve the generalization performance;Again,the mesh activation function is selected to directly cut off the negative network,and the gradient descent is smoother.;Finally,self-attention module and focusing loss function are introduced to make the network pay more attention to the relationship between channels and extract more sample features with high similarity.Experimental results show that the proposed method is superior to the traditional network in extracting fault features of pumping units.(3)In order to make the deep learning algorithm applied to the old equipment with weak computing capability,this paper makes a lightweight research on fault diagnosis of pumping unit.Shuffle Net network was used to construct the fault diagnosis model,and grouping convolution and channel-by-channel convolution were introduced to reduce the complexity of feature extraction while reducing the number of parameters and computation.In addition,channel rearrangement operation was used to fuse information of different dimensions to improve detection accuracy.Finally,the experimental results show that compared with the traditional network,the lightweight network model can greatly reduce the number of network parameters,the amount of calculation and the size of model weight on the premise of ensuring accuracy.