Research On Fault Diagnosis Method For Sucker Rod Pumping Systems Based On Dimensionality Reduction Algorithms

Sucker rod pumping system is one of the most widely used artificial lift system.In the actual production process,all kinds of faults in the pumping system will cause the reduction of oil production and even damage to equipment.The dynamometer cards describe the relationship between the displacement and the load of suspension point in one stroke.By analyzing the shape of the dynamometer cards,abnormal conditions in the well can be monitored and key parameters in production can be adjusted timely.Therefore,it is very important to realize the automatic diagnosis of sucker rod pumping system.Based on the background of fault diagnosis of suck rod pumping systems,this paper aims to develop advanced methods to improve the accuracy and the applicability of dynamometer cards diagnosis.By introducing dimensionality reduction algorithms,the fault diagnosis methods of rod pump pumping systems have been deeply studied.The main contributions are summarized as follows:(1)In order to utilize the unlabeled dynamometer cards effectively,a nonlinear dimension reduction method called Data-dependent kernel semi-supervised sparsity preserving projection(DKSSPP)is proposed.This method combines the data-dependent kernel and the sparse preservation projection algorithm.Firstly,the data-dependent kernel is used instead of the standard kernel function,and Fisher criterion is introduced to map all labeled samples into a high-dimensional feature space,in which all labeled samples have the maximum separability.Then the sparse reconstruction relationship of all samples is calculated in this high-dimensional feature space.Finally,all data is mapped into a low-dimensional space by a transform matrix,and the reconstruction relationship is maintained.Experiments show that DKSSPP can effectively utilize unlabeled samples and extract features of dynamometer cards for diagnosis.(2)Existing dimensionality reduction algorithms based on data-dependent kernel generally have the problem that they can't simultaneously consider the combining coefficients optimizing and the manifold structure calculating.To solve this problem,Data-dependent sparse discriminant analysis(DKSDA)and Supervised data-dependent sparsity preserving projection(SDKSPP)are proposed from two perspectives.DKSDA uses label information to optimize the coefficients in data-dependent kernels by constructing the inter-class and the intra-class graphs,while SDKSPP introduces the smooth assumption that the feature space and the label space share the same manifold structure to solve this problem.By comparing the two algorithms,it can be found that the SDKSPP algorithm based on smooth assumption can extract the feature of dynamometer cards more effectively for fault diagnosis.(3)To solve the problem that dimensionality reduction algorithm based on sparse representation can only extract local structure of data,the method named Data-dependent kernel based low-rank discriminant analysis(DKLRDA)is proposed.This method extends the low-rank preserving embedding into a supervised non-linear dimensionality reduction method by using data-dependent kernels.By imposing a sparse constraint on the low-rank reconstruction coefficient matrix,DKLRDA can extract both local and global features of data.Based on the conclusion of the previous chapter,the smoothness assumption is used as the optimization criterion,so DKLRDA can simultaneously consider the combining coefficients optimizing and the low-rank structure calculating.The Inexact Augmented Lagrange Multiplier algorithm(IALM)is used to solve the non-convex optimization problems.Experiments show that the proposed DKLRDA can achieve better performance for fault diagnosis of suck rod pumping systems.(4)When the categories of historical data from one well are incomplete,the cards of the missing categories can not be effectively identified by the model that trained from only the historical data of this well.To overcome this problem,a method named Supervised dictionary-based subspace transfer learning(SDSTL)is proposed based on dimensionality reduction algorithms and transfer learning.This method can effectively utilize the complete historical data from another pumping well.After dimensionality reduction,the source data that contain all kinds of fault samples and the target data that lack some kinds of fault samples can be represented by a shared dictionary matrix.By introducing two idea regularization terms,the structure information of source data and target data are included into the dictionary learning process.Thus,the obtained dictionary has the discriminative ability.Extensive experiments are conducted to evaluate the effectiveness of the proposed method.(5)In view of the categories of historical data from multiple well are all incomplete,a method named Subspace transfer learning based on discriminant dictionary for multi sources(STLDDM)is proposed.This method assumes that the categories of historical data from each well are incomplete,but the categories of all data from multiple wells are complete.Data from different wells can be mapped into a common low-dimensional space by different transform matrices,and the mapped data can be linearly reconstructed by the same dictionary matrix.The intra-class graph is introduced into the objective function to make dictionary matrix more discriminative.Experiments show that the proposed method can use the incomplete data from multiple pumping wells effectively to identify cards of missing categories from the corresponding well.