Design And Realization Of Indicator Diagram Classification Algorithm Of Pumping Unit’s Based On CNN

The use of computer technology to automatically,accurately,and efficiently classify the indicator diagram of pumping units is an important part of oilfield intelligence and specialization.In response to the problems of limited data volume and low accuracy in the current classification methods for indicator diagram,combined with oilfield production data,we propose a classification algorithm for indicator diagrams based on convolutional neural network is proposed.The main work is as follows:(1)A deformable convolution based indicator diagram classification algorithm is studied and implemented.The curve part of the indicator diagram is effective features.The established classification model of the indicator diagram should not only maintain a high classification accuracy,but also better meet the practical requirements of automatic classification recognition.The best performing Dense Net among several classical convolutional neural network was selected as the basic experimental model.From the perspective of improving the performance of the model,the loss function,convolutional layer,and optimization algorithm of the network are improved to build the Dense Net-DC network.The convolution of 3×3 in the Dense layer is replaced by deformable convolution to improve the learning ability of the network for geometrically variants and enhance the spatial modeling ability of the network;Cross-entropy loss function is replaced by the Focal-Loss function as a way to solve the problem of unbalanced samples;SGD optimization algorithm is replaced by Adam optimization algorithm to solve the problem of slow model training speed and excessive loss oscillation.The experimental results shown that the Dense Net-DC model can effectively improve the accuracy of the model to a certain extent,which is 1% higher than the accuracy of the Dense Net network classification.(2)A classification algorithm based on attention mechanism is studied and implemented.The dense connectivity unique to the Dense Net network can solve the gradient disappearance during deep network training,but it also leads to a large amount of redundant information reuse and does not consider learning the weight correlation between feature channels from the feature dimension level,which reduces the efficiency of the feature extraction process.To improve the efficiency of feature utilization in the model,an SE-Dense Net-DC network is constructed.In each Dense Block module of the Dense Net-DC network,add an "SE module" for feature adaptive recalibration,allowing the network to adaptively adjust the weights of each channel and pay more attention to the effective feature channels in the feature layer.And add separate SENet blocks after each transition layer,readjust the feature map,and maintain the dense connectivity characteristics of the network.Through a series of comparative experiments,it has been shown that the SE-Dense Net-DC network has the highest classification accuracy of 98.4%.The comprehensive experimental results indicate that,compared to other methods,the algorithm proposed in this paper considers the correlation between specific indicator diagram and actual production processes,so as to reflect the real situation.It can solve the complex process of manually extracting geometric features of indicator diagrams and the problem of improper feature selection in traditional networks,and automatically and accurately classify indicator diagram images.On the other hand,based on convolutional neural networks,the classification algorithm can improve the comprehensive management of oil fields and provide a new perspective for future intelligent operations.