Research On Diagnostic Method For Working Conditions Of Pumping Wells Based On Deep Learning

The downhole working environment of the pumping wells is very complicated,and the working condition of the pumping wells is analyzed by establishing the motion equation of the pumping wells in the early stage.For pumping systems,too complex motion equations often contain unmeasured state variables,which are difficult to predict and control;and too simple motion equations cannot fully reflect certain key features.An appropriate mathematical model can be established and combined with a data-driven approach to solve the fault diagnosis problem of the pumping unit well.In this paper,the ground dynamometer card obtained by direct measurement as the object.Because the downhole pump dynamometer card can directly reflect the working state of the pump in a reciprocating process,the paper established the wave equation of the sucker rod firstly.Taking the viscous damping wave equation proposed by Gibbs as the main body,constraining the load and displacement of the polished rod suspension point as the boundary conditions,the analytical expressions of load and displacement at any cross-section of the sucker rod string are solved by Fourier transform,and the pump dynamometer card of each cross-section of the sucker rod string is obtained.Then the data of the pump dynamometer card is preprocessed.In this stage,the signal of the pump dynamometer card is first converted into a pump dynamometer card image,and then the image is binarized to reduce the complexity of the calculation.Subsequently,the paper proposes a deep learning-based offline training and online diagnosis of pump dynamometer card.For deep learning models,such as LeNet,AlexNet,and VggNet networks,are used to model and train the pump dynamometer card data in the experiment to explore suitable models and methods.The simulation results show that most of the performances of the above three networks under the pump dynamometer card data provided by the paper are relatively close,but the overall performance of the LeNet network is better,so it is selected for subsequent research.Then,by optimizing the structure and parameters of the LeNet network,a local optimal LeNet model based on the greedy strategy is designed.Finally,in order to further improve the model,a hybrid model based on LeNet network and SVM is proposed from the point of view of the algorithm.Through experimentation and comparison with the previous method,the overall performance of the hybrid model under the pump dynamometer card data provided by the text has been improved,verifing the effectiveness and feasibility of the method.