The Study Of Pump Well Fault Recognition On Convolutional Neural Network

In the actual production process of an oil field,the production guarantee of oil is very important.The rod pump oil extraction technology is the most common way of oil production in China,accounting for more than 90%.Therefore,in order to improve the production efficiency of an oil field,fault diagnosis research is very important.The working environment of pumping wells is complex.After a certain production cycle,physical fittings such as sucker rod,pumps and pumps will fall off and age,and will also be affected by underground environment.Faults are easy to occur.By real-time monitoring of production parameters,the fault type can be determined by the indicator diagram analysis method,which has always been the mainstream method of fault diagnosis.However,the traditional classification method of dynamometer mainly depends on manual experience,which is not universal and difficult to popularize,and can not meet the requirements of intellectualized oilfield put forward by modern production.By relying on the artificial prior knowledge to model the features,the fault diagnosis based on the artificial dynamometer characteristics is heavy workload and inefficient.At the same time,the complex working condition in underground increases the difficulty of manual diagnosis,reduces the accuracy of diagnosis,and makes it difficult to meet the requirements of field construction.Therefore,the improved convolution neural network proposed in this paper has important significance for the fault diagnosis of pumping wells.Deep learning is the mainstream method at present.Convolutional Neural Network has good classification effect.However,among the types of fault of indicator diagrams of pumping wells,there are some faults with small amount of data,small datasets and different types of faults,but the reasons for the similarity of the images are so similar that the image classification ability of network models is poor and the classification accuracy is low.In view of the above problems,this paper mainly studies the content and specific work as follows:(1)Build up Generative Adversarial Networks to enhance the data of the types with few failures in the dataset,to ensure the balance of the number of failures in the dataset.The loss function in the model is improved.By combining the focus loss function with the original network loss function,the contribution degree of the network to the difficult-to-distinguish feature information is improved.The weight of the contribution of the difficult-to-distinguish sample to the loss function is increased,the influence of the easily distinguishable feature is reduced,and the resulting sample has better quality.(2)A fault diagnosis model based on convolution neural network is established,and attention mechanism is introduced to improve the effective recognition ability and accuracy of convolution neural network.For low-differentiated implicit features,the rescaled channel weights generated by the attention mechanism are suppressed,while for High-differentiated implicit features,the attention mechanism specifically enhances their impact weights to enhance recognition accuracy.The network classifier is improved,and a progressive classifier is designed and built to make the feature information extracted in the backbone network progressively overpass the number of specific classification targets in the progressive classifier,to make the transmission of feature information more stable,to screen out more effective hidden features for image classification,and to improve the accuracy of network classification.(3)For the network structure of convolution neural network,a lightweight structure is proposed,which can significantly reduce the size of parameters and computational load,and ensure the good integrity of the image hidden feature information during feature extraction.Finally,through experimental validation,the improved model proposed in this paper reduces the structure complexity of the model and reduces the size of parameters and the amount of calculation compared with the ordinary convolution network on the premise of ensuring the accuracy of model fault diagnosis.