Research On Intelligent Diagnosis Method Of Working Conditions For Beam-pumping Unit Based On Torque-angle Indicator Diagram

The common diagnosis method of working conditions for beam-pumping unit(the method of indicator diagram)has limitations in data collection,high cost and low reliability,which seriously restricts the construction of "smart oilfield".The intelligent diagnosis method based on the torque-angle indicator diagram is based on the motor parameters and combined with the deep learning model,which can realize the automatic diagnosis of beam-pumping unit,improve the accuracy and efficiency of diagnosis,and reduce the costs of oil production.With the beam type pumping engine as the research object,by analyzing the combinations of parameters that can form the indicator diagram,the torque-angle is compared to form the indicator diagram,and the ground system model is constructed based on this derivation.The transformation experiment is carried out through the measured field data,and,the Cartesian coordinate system is used to form the indicator diagram with the characteristics of the parameters.Finally,a library of 15 typical working conditions is established and the corresponding characteristics are analyzed.The eight-category dataset of the torque-angle indicator diagram was constructed based on the above model,and combined with five typical convolutional neural network models(AlexNet,VGG16,ResNetV2,MobileNetV3 and DenseNet201)for recognition training,and various categories and overall performance analysis were completed through various metrics.Finally,considering the small sample size and data imbalance characteristics of the dataset,the ResNetV2,MobileNetV3 and DenseNet201 models based on transfer learning are trained for recognition to explore the performance impact of transfer learning on typical deep convolutional neural networks.The research has shown: the angle can directly characterize the operating position of the beam-pumping unit,the electrical parameters or the data of indicator diagram are more accurate for transforming torque,and the composed torque-angle indicator diagram can comprehensively characterize the operating state of the system.In addition,the feature analysis of the torque-angle indicator diagram under 15 typical working conditions reveals a high recognition degree.Through multi-indicator analysis,the recognition accuracy of the five convolutional neural network models,except MobileNetV3,meet the actual production requirements,but the recognition performance of the small sample size categories is not fully consistent with actual requirements.The transfer learning effectively alleviates the problems of small dataset size and unbalanced category samples,and significantly improves the recognition performance and generalization ability of the network models.The change rate of accuracy and the convergence speed of the loss function of the network model based on transfer learning are generally faster than the basic model during the training process,the training loss is relatively less,and the oscillation problems of ACC curve and Loss curve have been greatly alleviated.Its overall performance and recognition performance of each category have reached an excellent level,more in line with actual production requirements.